You are on page 1of 87

How to Identify and Predict Bull and Bear Markets?


Erik Kole†
Dick J.C. van Dijk
Econometric Institute, Erasmus School of Economics, Erasmus University Rotterdam

February 1, 2012

Abstract
This paper compares fundamentally different methods to identify and predict
the state of the equity market. Because this state is a comprehensive economic
indicator, good identification and prediction are important for financial decisions
and economic analyses. We consider rules-based methods that purely reflect the
direction of the market, and regime-switching models that take both average returns
and volatility into account. Because the state of the equity market is latent, we
develop a novel framework for statistical and economic comparisons of the different
methods. Rules-based methods are preferable for identification, because ex post only
the direction of the market matters. Regime-switching models perform significantly
better in making predictions. Focusing on average returns and volatilities leads to
more prudent forecasts and a higher utility from the perspective of risk-averse investor
who engages in market timing. Both the statistical and economic distance measures
indicate that these differences are significant.


We thank Christophe Boucher, Thijs Markwat and seminar participants at the 8th International Paris
Finance Meeting, Inquire’s UK Autumn Seminar 2010 and Erasmus University for helpful comments and
discussions. We thank Anne Opschoor for skillful research assistance and Inquire UK for financial support.
Kole thanks the Netherlands Scientific Organisation (NWO) for financial support.

Corresponding author. Address: Burg. Oudlaan 50, Room H11-13, P.O. Box 1738, 3000DR Rot-
terdam, The Netherlands, Tel. +31 10 408 12 58. E-mail addresses kole@ese.eur.nl (Kole) and
djvandijk@ese.eur.nl (Van Dijk).

1

1 Introduction
The state of the equity market, often referred to as bullish or bearish, is an important
variable in finance and more generally in economic analysis. Despite its importance, the
literature does not offer a single preferred method to identify and predict the state of the
equity markets. In this paper, we compare several existing methods that can identify and
predict bull and bear markets. These methods differ in their use of the characteristics of
bull and bear markets. During bull markets prices gradually rise and volatility is low, while
during bear markets prices can fall dramatically and volatility is high. Our comparison
offer the main insight that for ex post identification, methods that focus explicitly on price
increases and decreases work best, whereas for ex ante predictions, methods that combine
the trend in prices with volatility are preferable.
Information on the state of the equity market is obviously relevant for agents who are
closely involved in financial markets. Investors may follow a market timing strategy, with a
long position in the equity market when it is bullish, and a neutral or short position position
when it is bearish. Investors that do not engage in market timing strategies can incorporate
the different behavior of stock returns dependent on market sentiment (see Perez-Quiros
and Timmermann, 2000) in their risk management. Firms prefer to issue new equity during
bull markets. For regulators, the state of the equity market is important, because it can
affect the credit supply with destabilizing effects on the real economy. When financial
assets are used as collateral, bull markets extend the credit supply, while bear markets
reduce it (see Rigobon and Sack, 2003; Bohl et al., 2007). Finally, bull and bear markets
impact asset pricing, as they are an important source for time variation in risk premia (see,
for example, Veronesi, 1999; Gordon and St-Amour, 2000; Ang et al., 2006).
The importance of bull and bear markets concerns economic analysis in general, and
not just investments. As argued by Stock and Watson (2003a), stock prices can predict
macroeconomic variables, as they are discounted future dividends. As far back as Mitchell
and Burns (1938), the state of the stock market is considered as an ingredient for leading
indicators of the business cycle (see Marcellino, 2006, for an overview). Harvey (1989); Es-
trella and Mishkin (1998); Chauvet (1999) and Stock and Watson (2003b) report evidence
that the state of stock market helps predicting the business cycle.

2

The literature offers two fundamentally different types of methods to identify and pre-
dict the state of equity markets: non-parametric ones based on rules and fully paramet-
ric ones based on models. Rules-based methods are more transparent than model-based
methods that need statistical inferences, and more robust to misspecification. On the other
hand, a completely specified model for the price process on the equity market offers more
insight and its quality can be evaluated by statistical techniques, whereas rules-based meth-
ods typically require some arbitrary, subjective settings. As a final difference, model-based
techniques are statistically more efficient, because they treat identification and prediction
in one step, whereas rules-based methods are two-step approaches.
We develop new techniques to compare the identifications and predictions of the dif-
ferent methods. Existing techniques, for example to test for predictive ability, do not
suffice, because the state of the equity market is latent. There is no generally accepted
chronology of bull and bear markets like the NBER chronology of recessions. To deter-
mine the difference between identifications and predictions of the methods, we propose a
general statistical distance measure and a more specific economic measure. As a statistical
measure we propose the Integrated Absolute Difference (IAD) between identifications or
predictions. It is closely related to the Integrated Square Difference of Pagan and Ullah
(1999) and Sarno and Valente (2004), though easier to interpret as a difference in probabil-
ity. The economic measure quantifies the preference for an method over another method.
We determine this measure as the fee that a risk-averse investor would maximally pay to
switch from one method to another to engage in marketing timing.
In the category of rules-based methods, we consider Pagan and Sossounov (2003) and
Lunde and Timmermann (2004). These non-parametric methods first determine local
peaks and troughs in a time series of asset prices, and then use rules to select those
peaks and troughs that constitute genuine turning points between bull and bear markets.
They are based on the algorithms used to date recessions and expansions in business cycle
research (see Bry and Boschan, 1971, among others), and have been adapted in different
ways for application in financial markets. The main rule in Pagan and Sossounov (2003)
is the requirement of a minimum length of bull and bear cycles and phases.1 By contrast,
1
See Edwards et al. (2003); Gómez Biscarri and Pérez de Gracia (2004); Candelon et al. (2008); Chen
(2009) and Kaminsky and Schmukler (2008) for applications.

3

Lunde and Timmermann (2004) impose a minimum on the price change since the last peak
or trough.2
In the second category, we analyze Markov regime-switching models pioneered by
Hamilton (1989, 1990). In these models returns behave differently, depending on a discrete
latent state process that follows a Markov chain. We consider models with two and with
three distinct states. Conditional on the state, returns follow a normal distribution. Em-
pirical applications typically distinguish two regimes with different means and variances
and normally distributed innovations.3 The bull (bear) market regime exhibits a high (low
or negative) average return and low (high) volatility. The number of regimes can easily be
increased to improve the fit of the model (see Guidolin and Timmermann, 2006a,b, 2007)
or to model specific features of financial markets such as crashes (see Kole et al., 2006) or
bull market rallies (see Maheu et al., 2009). We follow the literature and consider models
with two and with three regimes.
A comparison of the identification of the state of the US stock market, proxied by the
S&P 500, over the period 1955-2010 shows a preference for the rules-based methods. The
rules-based methods simply separate periods with price increases from periods with price
decreases. The regime-switching models identify periods where the risk-return trade-off
is attractive, a positive mean and low volatility, or unattractive, a negative mean and
high volatility. However, periods with high volatility are sometimes labeled bearish, even
if they exhibit positive average returns. Periods with negative average returns may be
labeled bullish as long as the volatility is low. The IADs indicate that the rules-based and
model based methods produce identifications that are significantly different. A risk-averse
investor would pay up to 20% per year to time the market based on a rules-based method.
This fee is significant, but should be interpreted with care as it corresponds with perfect
foresight.
Because the state of the equity market is a comprehensive economic indicator, predicting
it well may be more important than identifying it. We set up a forecasting experiment,
where predictions are made from July 1983 onwards. From a utility perspective, regime-
2
Chiang et al. (2009) adopt this method.
3
See for instance Hamilton and Lin (1996); Maheu and McCurdy (2000); Chauvet and Potter (2000);
Ang and Bekaert (2002); Guidolin and Timmermann (2008a) and Chen (2009) for applications.

4

switching models are preferable. However. In the dating of recessions and expansions. the method of Lunde and Timmermann (2004) shows the best performance. The sooner a switch is identified. The different results for identifications and predictions show that quickly picking up bull-bear changes is crucial.4%. but find that the effects are at best marginal. looking at Sharpe ratios. Our research relates directly to the debate between Harding and Pagan (2003a. switches are identified immediately. The best performing regime-switching model yields a Sharpe ratio of 0. The resulting Sharpe ratio of 0.b) and Hamilton (2003) on the best method to date business cycle regimes. and the performance is less. Regime switch- ing models are fastest in this respect.7% in excess of the risk-free rate and a volatility 32. Harding and Pagan advocate simple dating rules to classify months as a recession or expansion. When predictions have to be made. While the IAD indicate significant statistical differences. For the rules-based approaches their use consistently lowers performance. the larger the gains. both methods base their identification mainly on the sign of GDP growth and produce compa- rable results. which partly explains their good performance.21. We use a specific-to- general selection procedure to include those predictive variables that work best in-sample. with an average return of 8. the methods detects switches with some delay. 1994). but produces many costly false alarms. while Hamil- ton proposes regime switching models.. We investigate whether inclusion of macro-financial variables improves the predictions of the different methods. The methods of Pagan and Sossounov (2003) rapidly picks up switches. They make prudent forecasts and yield highest utility.15. A market timing strategy based on regime switching models produces lower volatility than based on rules. and by the regime switching models are all significantly different. whereas perfor- mance improves when predictive variables are included in the transition probabilities of the regime-switching models (see Diebold et al. the fees indicate that the value of these differences are small and insignificant. The IADs indicate that the predictions by the methods of Lunde and Timmermann (2004). The maximum fee to switch to a regime switching model from a rules-based methods is around 16% and significant. With perfect foresight.27 compares favorably to the Sharpe ratio of a long position in the market at 0. of Pagan and Sossounov (2003). For dating bull and bear periods in the stock market by regime switching 5 .

On the other hand. we consider the dynamic combination of more predictive variables. Harding and Pagan (2002a) discuss similar issues with respect to business cycle dating. their identifications and predictions differs substantially from the rules-based approaches. a parametric setting offers statistical techniques to assess the quality of the model. Taken together. Unfortunately. The parametric techniques are based on models that specify the distribution of stock re- turns conditional on the state of the stock market and the dynamics of the state. We extend the anal- ysis of Chen (2009) in several ways. Our results for the rules-based approaches show that the in-sample added value of the predictive variables is not met with out-of- sample quality. Second. The 6 . First. The non-parametric methods consist of sets of rules. He treats the smoothed inference probabilities as observed dependent variables in a linear regression. the volatility of recent returns seems at least as important (if not more) than their sign. we find quite some variation in the selected variables and their coefficients. They differ in their selection of peaks and troughs that constitute the actual switch points between bull and bear markets. we include predictive variables directly in the regime switch- ing models and do not need Chen (2009)’s two-step procedure. can typically choose between non- parametric and parametric methods. Specifying a model implies the risk of misspecification. Both methods identify bull and bear markets by peaks and troughs in a price series. As rules-based methods. An economic agent who wants to infer it. these results are in line with those documented by Welch and Goyal (2008) for direct predictions of stock returns. which does not take their probabilistic nature into account. For the regime switching models. we consider the algorithms proposed by Lunde and Timmer- mann (2004) and by Pagan and Sossounov (2003). Consequently. to which non-parametric techniques are more robust. Strategies with predictive variables perform worse than those without.models. We also add to the discussion on predictability in financial markets. this state and its process is latent. 2 Rules or regime switching: theory The state of the equity market is an important variable when making investments and for economic decision making in more general.

no update takes place. (b) If the index has increased with a fraction λ1 . 1990).model-based methods we consider are Markov regime-switching models as pioneered by Hamilton (1989. no update takes place. (b) If the index has dropped by a fraction λ2 . (a) If the index has dropped below the last minimum. In this approach. The agent considers the subsequent period. 7 . (a) If the index has exceeded the last maximum. A bull (bear) market occurred if the index has increased (decreased) by at least a fraction λ1 (λ2 ) since the last trough. (2006) or bear market rallies and bull market corrections as in Maheu et al. a trough has been found. 2. the maximum is updated. The identification rules can be summarized as follows: 1. the minimum is updated. we write Stm to denote the state of the equity market at time t for method m. (c) If neither of the conditions in satisfied. for example to capture sudden booms and crashes as in Guidolin and Timmermann (2006b. In this paper. The agent considers the subsequent period. The identification of peaks and troughs in a price series {Pt }Tt=1 uses an iterative search procedure that starts with a peak or trough. 2007) and Kole et al. the state of the stock market follows a first order Markov process with a specified number of regimes. 2. peaks and troughs have to meet minimum requirements on their magnitude to qualify as switch points between bull (between a trough and the subsequent peak) and bear markets (between a peak and the subsequent trough). The number of regimes can be set equal to two (a bullish and bearish regime) or larger.1 Identification and prediction based on rules In the algorithm of Lunde and Timmermann (2004. LT henceforward). The last observed extreme value was a trough with index value P min. The last observed extreme value was a peak with index value P max . (c) If neither of the conditions is satisfied. a peak has been found. (2009). More states can be introduced.

unless the absolute price change exceeds a fraction ζ. This implies that an increase of 20% over the last trough signifies a bull market. PS hence- forward). 2002b). 3. We mostly follow PS for the values of these parameters. but instead put restrictions on the minimum duration of phases and cycles. τcycle = 70. We count the number of times the maximum and minimum of the index have to be adjusted since the first observation.20 and λ2 = 0. A local maximum (minimum) is higher (lower) than all prices in the past and future τwindow periods. Their approach is based on the identification of business cycles in macroeconomic data (see also Harding and Pagan. If the maximum has to be adjusted three times first. We have τwindow = 32. τphase = 16 and ζ = 0. otherwise it starts bearish. opposite to the 26 weeks taken by PS. 4. Appendix B). Eliminate cycles of bull and bear markets that last less than τcycle periods. Censor peaks and troughs in the first and last τcensor periods. We follow LT by setting λ1 = 0.After these decision rules the agent considers the next period. the market starts bullish. We censor switches in the first and last 13 weeks. Censoring for 26 weeks would mean that only after half a year an investor can be sure whether a bear or a bull market prevails. The second approach has been put forward by Pagan and Sossounov (2003. Their algorithm consists of five steps 1. Construct an alternating sequence of peaks and troughs by selecting the highest maxima and lowest minima. 5. which we consider a very 8 . Eliminate bull market or bear markets that lasts less than τphase periods. Locate all local maxima and minima in a price series. and that a decrease of 15% since the last peak indicates a bear market. adjusted for the weekly fre- quency of our data. They also use peaks and troughs to mark the switches between bull and bear markets. However. PS do not impose requirements on the magnitude of the change of the index.20 (see also PS. 2. To commence the search procedure we determine whether the market is initially bullish of bearish.15. their identification is quite different from LT.

The PS- alogirthm suffers from this problem too. m = LT. as it combines a logit model with the m Markovian property that the probability distribution of the future state St+1 is (partly) determined by Stm . In that case. PS. To form the one-period ahead prediction for πTm+1 . zt−1 ] = Λ(βqm ′ zt−1 ).long time. In the LT-approach. We adjust the standard logit model such that the effect of an explanatory variable on the probability of a future state can depend on the current state. the prevailing state at time T is needed. Some macro-finance variables may have a different (or no) effect on the probability of a switch from a bull market than from a bear market. For the rules-based approaches. In both methods. We opt for a logit model. only if PT equals the last observed maximum (minimum). the next step is to relate the resulting series of bull and bear states to a set of explanatory variables. q = u. zt−1 . We call this model a Markovian logit model. the state of the market is known until the period of the last extreme value. the market state process is a standard stochastic process with the Markov property. Since the dependent variable is binary. but this will only become obvious later. 9 . So. and βqm is the coefficient vector on the zt−1 variables. If only a constant is used. For notational convenience. we use a shorter period of 13 weeks to establish the initial and the ultimate state of the market. a normal logit model results. the market may already have switched. this information may not be available. The probability for a bull state to occur at time t is modeled as m πqt ≡ Pr[Stm = u|St−1 m = q. We code bull markets as Stm = u and bear markets as Stm = d. which depends on the previous state of the market q. Since we will use this information in making predictions. we assume the first variable in zt−1 is a constant to capture the intercept term. d (1) where Λ(x) ≡ 1/(1 + e−x ) denotes the logistic function. a logit or probit model can be used. since only the state up to the last τcensor periods is known. If the coefficient βqm does not depend on q. and is a fraction λ1 above (λ2 below) the prior minimum is the market surely in a bull (bear) state. as this model can be easily extended to a multinomial logit model when more states are present.

We construct the one-period ahead prediction recursively m m Pr[St+1 = s|zt ] = Pr[St+1 = s|Stm = u. Second. they may be more robust. Based on the results of Guidolin and Timmermann (2006a. whereas a simple characterization of the rule-based results may not be straightforward. In particular (missed or misspecified) changes in the data generating process can have severe impact on the results. we consider k = {2. 2007. (2) Starting with the known state at T ∗ . We consider several Markov chain regime switching models for the stock market. First. we now first write down a model that can be the data generating process of a stock market index that allows for prolonged bullish and bearish periods. which we use for the predictions of T ∗ + 2 and so on. Instead of applying a set of rules to a given series. and use S m to denote the 10 . we can easily extend the number of states in the model. Estimating such a model produces probabilistic inferences on periods of bull and bear markets in a certain index. T ∗ < t ≤ T + 1. it offers more insight into the process under study. 3}. having k regimes (used as first suffix) and having either constant or time-varying probabilities (suffix C or L). The advantages come at the cost of misspecification risk.2 Identification and prediction by regime-switching models We also consider a method for identifying and predicting bull and bear markets that is fundamentally different from the algorithms considered in the previous section. A third advantage is the ease with which we can compare results for different markets and different time periods. Models can typically be summarized by their coefficients. We can derive theoretical properties of the model and see whether it yields desirable features. For example. 2008b). zt ] Pr[Stm = u|zt−1 ]+ m Pr[St+1 = s|Stm = d. zt ] Pr[Stm = d|zt−1 ]. the label RS3L means a Markov regime switching model with three states and time-varying transition probabilities. we construct predictions for T ∗ + 1. As rule-based approaches do not make strict assumptions on distributions or on the absence or presence of variation over time. This iteration stops at T + 1. 2. We can test whether such extensions imply significant improvements. Using such a model-based approach has several advantages.which we denote by T ∗ < T .

rt ∼ N(µm m s . (2009). Our approach differs from Maheu et al. Timmermann (2000) shows that the mixtures implied by regime switching models constituted by normal distributions can flexibly accommodate these features. s. ∀t : πsqt s∈S m = 1 applies. (5) ς∈S e t−1 with ∀q ∃s ∈ S : βqs = 0 to ensure identification. (4) P m Of course. this restriction leaves k (k the probabilities are time-varying. (6) 11 . s. the restrictions ∀q. If ties are kept constant. ωs ). This matrix contains parameters m πqst ≡ Pr[Stm = s|St−1 m = q. zt−1 ] = Λ(βqm ′ zt−1 ). we treat it as a latent variable that follows a first order Markov chain with transition matrix Ptm that can vary over time. m = RS2L. though the added value for prediction is less obvious. Since the actual state of the market is not directly observable.set of regimes. They enable identification by imposing that the expected return during bear markets including rallies is negative. while it is positive during bull markets including corrections. The difference in the predicted return distributions between a bull market and a bear market rally is likely to be small. If the number of regimes equals 2. q ∈ S m . zt−1 ]. When the transition probabili- ˙ − 1) free parameters to be estimated. These authors allow for bear markets that can exhibit short rallies and bull markets that can show brief corrections. this multinomial specification reduces to the standard logit specification m πqt ≡ Pr[Stm = u|St−1 m = q. we use a multinomial logit specification m′ m eβqs zt−1 πqst =P m ′z βqς . While asymmetric or fat-tailed distri- butions can be used for the regimes. q ∈ S m . s ∈ S m (3) We order the regimes based on the estimated means. We assume that the excess index return rt follows a normal distribution in each regime s with regime specific means and variances. This setup can improve identification.

This comparison is complicated by the latent nature of the true regime. White (2000). While this statistic indicates how different two methods are. Therefore. For a purely statistical comparison. we propose a new framework that builds on the probabilities for the different states of the market. We consider an investor who wants to time the market.. (1994) to estimate the parameters of the multinomial logit specification. We treat the remaining parameters as free. ξsm ≡ Pr[S1m = s]. In appendix A we extend the method of Diebold et al. which define a loss function over the realization and the prediction of a certain variable. and derive the maximal fee that would make her ex post indifferent towards two methods. standard ways to compare the different identifications do not avail. and estimate them.This specification is mathematically similar to the logit models for the rules based ap- proaches in Eq. The evaluation of predictions suffers from the same problem.4 Instead. 24). Hamilton. 3 Comparing two filters We want to establish the difference of the results of the various approaches and test for their significance. As a consequence. Ch. We estimate the parameters of the resulting regime switching model by means of the EM-algorithm of Dempster et al. Again the restriction s∈S m ξsm = 1 should be satisfied. which is used in the business cycle literature. West (1996). To determine the optimal parameters describing the distribution per state. (1977). though it is an integrated part of the regime switching model. 12 . contrary to for example the NBER list of recessions and expansions. it does not point out which one is better. We cannot apply existing tests for predictive ability. We finish by introducing parameters for the probability that the process starts in a P specific state. (1).g. Corradi and Swanson (2007). we propose a statistic that is based on the absolute difference between the probability vectors that result from different methods. Because the true chronology of bull and bear markets is latent a prediction can never be classified as true or false. we develop a second statistic that is based on economic decision-making. Both statistics 4 See for example Diebold and Mariano (1995). Neither does an accepted reference list of bull and bear markets exist. we follow the standard textbook treatments (e. 1994.

(7) to an information set Ωt2 to determine the likelihood ps of each state s at a point in time t1 . β̂ m . 1)′ for m = LT. If regime switching models are used. To determine turning points. this probability vector can be interpreted as a forecast probability. it does not matter whether a prices series or a return series is taken as input. In case of the rules based approaches. so F m : (t. a larger absolute value for g(pm. Comparing the results of two different filters F m (t1 . The statistical comparison is more general. θ n ) and F n (t1 . θ m ) → p. β̂ m . Ch. Ωt = {(rτ . When a statistical measure is used. 1994.can be used for the identification and the predictions that the methods produce. pn ) → R. the function g is closely related to pm and pn . Ωt2 . So. the likelihood corresponds with identification. Each algorithm m applies a filter F m : (t1 . Ωtm 2 . The information set contains a return series and a set of explanatory variables. Ωtn2 . denoted by ΩT . or exogenously specified parameters θ m . The difference between pm and pn can come from a different filter algorithms F or from information sets of different length Ωtm 2 or Ωtn2 . the function g is defined in the context of the economic decision-making that is considered. 13 . From a statistical point of view. a stationary information set is advantageous. the state at time t is identified as either bullish or bearish. subsequent rules based on returns are applied. Since the initial index value P̃0 does not influence the results of the rules-based approaches. The filter may use parameters β̂ m ≡ β̂ m (Ωt3 ) that are estimated using the information set Ωt3 with t3 ≤ t2 . For an economic comparison. As pointed out by Pesaran and Skouras (2002). pn ) should indicate a larger difference. β̂ n . θ m ) = (1. we assume that a set of returns is included in Ωt . whose outcome can be interpreted as a difference. statistically based and economically based comparisons both have their merits. If the information set comprises all available information. The basis for both statistics is the interpretation of the different approaches as filters. ΩT . zτ )}tτ =1 . 22).5 If t1 > t2 . and we call it an inference probability. θ m ) is equivalent to comparing the two resulting probability vectors pm and pn . for example the boundaries λ1 and λ2 in the LT-algorithm. Therefore. There should be a function g : (pm . and 5 The rules-based approaches actually work with prices series to define peaks and troughs. the identification comes from the smoothed inference probabilities (see Hamilton. PS. 0)′ or (0. The regime-switching method are completely based on returns.

when we consider identification in the rules-based approaches. we next integrate over the different k states. such a measure can be easily linked to economic theory. d}. pm ) = φu |pm n m n m m m n m n u −pu |+φd |pd −pd | = φu |pu −pu |+(1−φu )|1−pu −(1−pd )| = |pu −pu |.can be relevant in various circumstances. the above expression simplifies to d(pn . and can take into account that the consequences of decisions can be asymmetric. (9) s∈S where φs ≡ Pr[S = s] is the probability that state s prevails under the true probability measure. The expression can be interpreted as an integrated absolute difference. In contrast. Calculating the difference between two methods of identification or prediction would be easy. The weight of the difference between pm n s and ps increases when state s is more likely to occur. similar to the integrated square difference in Pagan and Ullah (1999) and Sarno and Valente (2004). For the binomial case S m = S n = {u. 3. 14 . Wrong decisions may be more harmful than good decisions are beneficial. If the sets do not coincide. Since S is latent. we can reduce the larger set by aggregating two or more states. an economic comparison relates the comparison to a specific decision problem. West et al. S m = S n . X d(pn . when the realization is known. corresponding with the absolute difference. whereas the existing literature relates them to evaluate forecasting accuracy. 2000. we would calculate d(pm .1 The statistical difference We base our statistical distance measure on the L1-norm. pm ) ≡ φs |pm n s − ps |. pn |S = s) ≡ |pm n s − ps |. We cannot measure the difference between pm n s and ps by the ratio of their logarithms as proposed by Kullback and Leibler (1951) since either probability can equal zero or one. (8) We can only compare two filters if their set of states coincide. 1993). By using cost functions or utility functions (as in Granger and Pesaran. A novelty in our use of these concepts is their application to compare identification.. Conditional on the realized state s.

2 The economic difference To construct an economic measure for the difference between two predictions. We take the perspective of an investor who uses the predictions to speculate on the occurrence of bull or bear markets. When pm n s = 1 and ps > 1/2. In that case we would like to have a zero distance. which we discuss in Section 3. R > 1 and the observations before R are used as in-sample period to estimate model parameters. we can use the full sample and typically have R = 1. we need a framework that links the state probabilities to a decision. We can.t − ps.t |. Together. the two approaches lead to the same rounded inference or prediction. and a method to evaluate this decision and compare it to another state probability. where I() is the indicator function. We propose a slightly adjusted distance measure when pm n s ∈ {0. which means replacing pm n s by ps . (10) T − R + 1 t=R s∈S where R is the first period for which we compare the probabilities. 1951). 3.n . we replace pm n m n s by 1 − ps when ps = 1 and ps ≤ 1/2. When we compare identifications. p |S = s) = (11)  |1 − 2pns | otherwise. This situation arises when pm s corresponds with full-sample identification by the rules-based methods and pns with predictions from the rules-based methods or identification as well as predictions from the regime-switching models.n ≡ φs. dbm.n t t will depend on the choice for φt in a similar way as the Kullback-Leibler divergence (see Kullback and Leibler. 1} and ps ∈ [0. φ is irrelevant. We propose a bootstrap to derive the variance of of dd m. it means we replace the conditional distance in Eq. For predictions. While 15 .t |pm n s.3. By a similar logic. She speculates by taking long or short position in one-period futures contracts. (8) by the function   0 if pm n s = I(ps > 1/2) ˜ m n d(p . 1]. for example. pm )] by its sample equivalent 1 XX T dbm. In the binomial case. We estimate the expected value E[d(pn . In the multinomial case we need to make an assumption on φt . assume that φt = pm or φt = pn .

see e. Nonetheless. we think that this approach is direct and easy to interpret. this approximation fits in nicely with the regime-switching models that reflect both the mean and variance in its identification and predictions. The rules-based approaches do not provide predictions on the complete distribution of r. Harvey and Siddique (2000). Dividing by U ′ (W0 )W0 produces a standardized utility function 1 U ′′ (W0 )W0 2 2 1 Ũ (W0 (1 + wr)) = wr + ′ w r = wr − γw 2 r 2 . (13) produces the optimal portfolio wtm = µm m t /(γψt ). (13) 2 U (W0 ) 2 where γ is the coefficient of relative risk aversion. We assume the investor maximizes the expected value of her utility function U(W ). other approaches. given a vector with state probabilities pm t .t µs (15) s∈S m X X ψtm ≡ pm m s. (12) 2 Since the current utility level does not influence the optimization. We approximate the utility function to the second order around her initial wealth W0 : 1 U(W0 (1 + wr)) ≈ U(W0 ) + U ′ (W0 )W0 wr + U ′′ (W0 )W02 w 2 r 2 . The quadratic approximation only requires an estimate for the mean and the variance of the distribution.t ((µs ) + ωs ). X µm t ≡ pm m s. we derive the optimal decision. a similar approach with higher order moments would be straightforward. However. (14) where µm m t is the predicted mean and ψt the predicted raw second moment of rt+1 . Max- imizing the expected value of Eq.g.other frameworks are possible. for example using bull.t ψs = pm m 2 m s. (16) s∈S m s∈S m 16 .and bear markets for macroeconomic predictions. we can ignore the first term. We express her position as a fraction w of her initial wealth W0 . Jondeau and Rockinger (2006) or Guidolin and Timmermann (2008a). are also possible. Second. but just predict its sign. First. Her next-period wealth equals W0 (1 + wr). Both depend on the forecast probability for each state. We choose a quadratic approximation instead of other fully specified utility functions for two reasons.

West et al. wealth at time t + 1 equals W0 (1 + wtm rt+1 − ηm. refer 17 .n .n > 0. the bootstrap is best used to construct the distribution of ηd m.n is negative. (1993) interpret the relative increase in initial utility that would be needed to equate to expected utility levels of two strategies also as a fee. Based on series for {wtm }Tt=R−1 −1 .n − γηm. As discussed in the previous section. Including the fee. we are the first to adopt it as a test statistic. ηn.n = 0 2 which can be solved analytically for ηm. equal the utility resulting from method n. and denote it ηm. it can be interpreted as a compensation that the investor wants to receive for adopting an apparent inferior method m instead of n.where µm m m s . The utility resulting from method m should. the investor is willing to pay a fee for adopting m instead of n.n ) = EGr wt rt+1 − γ (wt rt+1 ) .n .m 6= −ηm. raw second moment and variance for model m. {wtn }Tt=R−1 −1 and {rt }Tt=R . by calculating the unconditional expected utility with respect to the unconditional distribution Gr of rt+1 . after paying the fee.n ). unless m = n. If λm. The measure is not symmetric. ψs and ωs are the state-specific mean. (17) 2 We determine the economic difference between two methods m and n by calculating the fee that an investor would be willing to pay to use method m instead of n. while other papers. for example Das and Uppal (2003) and Ang and Bekaert (2002). While a fee and the closely related concept of a certainty equivalent return have been used before to determine the economic value of investment strategies. we can estimate ηd m. If ηm. so she prefers method m over n. The fee is expressed relative to the investor’s wealth. Consequently. Fleming et al. so exchanging methods m and n does not produce the negative of the original fee. but also whether method m is more attractive to risk- averse investors.n .n . we evaluate the optimal portfolio produced by method m.n .n − γ (wt rt+1 − ηm.   m m 1 m 2 V (w ) = EGr wt rt+1 − γ(wt rt+1 ) .n is not only a measure for how different method m is from n. (2001) and Marquering and Verbeek (2004) calculate fees to compare dynamic investment strategies with a fixed benchmark of a static buy-and-hold strategy. Next. 2 2 (18) m n m  1 2 ⇔V (w ) − V (w ) − 1 − γ EGr [wt rt+1 ] ηm.     m 1 m 2 n 1 n 2 EGr wt rt+1 − ηm. ηm.

Hence. we calculate bootstrap estimates dd m. Since the investment strategies we consider are zero-cost. we deduct the fee from the investment return. 3. in this case the first part of the information set ΩjR−1 = ΩR−1 for each bootstrap j. In both cases. First. so including returns and predicting variables. depending on whether we compare the identification or predictions of different methods. j With these series. In the case of comparing predictions.to it as a certainty equivalent return. we construct bootstrapped samples from the original information set ΩT . Their approaches imply that the fee is paid up- front and reduces the amount available for investment.3 Bootstrap Conventional asymptotic techniques may not avail to determine the distribution of dd m. These set of bootstrapped estimates converges to the true distributions and can be used for testing and the construction of confidence intervals. the distribution of dm.n and construct their distribution.n has a lower bound at 0. We resample from the second part of the information. We implement the bootstrap in two ways.n and ηd j m. starting from the first prediction R. we apply the stationary bootstrap of Politis and Romano (1994). we follow the approach of White (2000). So.n for two reasons.n .n and ηd j m. we use the bootstrapped sample ΩjT to calculate new estimates β̂(ΩjT ) and to construct new series of inference probabilities. In turn. we propose to use the bootstrap.j T t }t=R and {pt }t=R . but use the estimates from the original series. the limiting distribution will be non-normal.n or ηd d m. We create new series {pm. To account for autocorrelation in these series. When we compare the different methods for identification.j T n. series of pm n t and pt may exhibit a high degree of autocorrelation that is not properly addressed by considering a limited number of autocovariances as proposed by Diebold and Mariano (1995). Therefore. Second. 18 . these lead j to bootstrapped estimates dd m.

1 Stock market data The state of the stock market should be determined against the benchmark of a riskless investment. We use weekly observations because of their good trade-off between precision and data availability. the data sample that we analyze starts on January 7. whereas the FRED series starts on January 4. On the other hand. Weekly data does not cut back too much on the time span. (19) τ =0 From a stock market index Pt the relevant series to determine the state follows P̃t = Pt /Bt . Higher frequencies lead to more precise estimates of the switches between bull and bear markets. data of predicting variables at a lower frequency is available for a longer time-span. as they are cheap and easily available. A riskless bank account Bt earns the risk-free interest rate rτf over period τ . This series starts on January 8. Studying the excess market index P̃t thus corresponds directly with the return on an investment opportunity. Louis has made available on FRED. 1885 until July 2. Our analysis considers the US stock market. 6 We kindly thank Bill Schwert for sharing his data with us. Starting with B0 = 1. Because of the availability of the predicting variables. proxied by the S&P500 price index on a weekly frequency. (20) The return on this index is the market return in excess of the risk-free rate.6 Schwert’s data set runs from February 17. the value of this bank account obeys Y t−1  Bt ≡ 1 + rτf . For the risk-free rate we use the three-month T-Bill rate. 1954.4 Data and implementation 4. 1955. also from FRED. 1957 and is kept up-to-date. It also corresponds with the return on a long position in a one-period futures contract on the stock market index. 19 . Futures contracts are the natural asset to speculate on the direction of the stock market. We splice together a time-series for the S&P500 by combining the data of Schwert (1990) with the S&P500 series that the Federal Reserve Bank of St. and gives a satisfactory precision. 1962.

the crash of 1987. The graph exhibits the familiar financial landmarks of the last sixty years. The T-Bill rate and the D/P-ratio exhibit a unit root.. 1991. [Figure 1 about here.2 Predicting variables We consider macro-economic and financial variables to predict whether the next period will be bullish or bearish. As a consequence. Avramov and Chordia (2006) provide evidence favoring the term spread and the dividend yield. we transform some of the predictive variables. We observe a clear alternation of periods of rise and decline in the 1950s and 1960s. Our choice of variables is motivated by prior studies that have reported the success of several variables for predicting the direction of the stock market.g. and the dividend-to-price ratio. unemployment. Hamilton and Lin (1996). we standardize each series. We transform the industrial production series to yearly growth rates. and finally the fall during the recent credit crisis. In Appendix B we provide more information on the predictive variables. 2005). We do not transform the inflation. 20 . Figure 1 shows the excess stock price index for the US. coefficients all relate to a one-standard deviation change and the economic impact of the different variables can be compared directly. We construct a stationary series by subtracting the prior one-year average from each observation. We join this literature and gather data accordingly for inflation. Avramov and Wermers (2006) and Beltratti and Morana (2006) use business cycle variables like industrial production. the T-Bill rate. industrial production and change in unemployment being the most successful.. Campbell. To ensure stationarity. To ease the interpretation of coefficients on these variables. For the unemployment rate we construct yearly differences. the dramatic rise during the IT-bubble of the late 1990s and the subsequent bust in 2000-2002. the inflation rate. industrial production. Ang and Bekaert (2002) show the added value of the short term interest rate.] 4. the prolonged slump during the late 1970s and early 1980s. Chen (2009) considers a wide range of variables with the term spread. the term and credit spread. the term spread or the credit spread series. Rapach et al. used more often in forecasting (see e. The index has been set to 100 on 1/7/1955.

Since we consider the T-Bill rate as a difference to it’s yearly moving average. The D/P-ratio is available at a weekly frequency for all of our sample period. For the same reason. Weekly observations of the term and credit spreads become available from January. industrial production and unemployment is available at a monthly frequency. The procedure stops when no further significant improvement is found. We select the variable-transition combination with the largest improvement and test whether this is significant with a like- lihood ratio test. If the improvement is significant. which also determines the starting date for our analysis. this sets the starting date one year later. The data for inflation. dating back to 1950 or earlier. we would need to estimate a model with 3 · 2 · 7 = 42 transition coeffi- cients.3 Variable selection We consider in total seven variables that can help predicting the future state of the stock market. For the T-Bill rate. This approach differs from the general-to-specific approach. we combine the stationary bootstrap of Politis and Romano (1994) discussed in Section 3. we use monthly observations. 4. We lag this data by one month. we do not follow Pesaran and 21 . which would include all variables first and then consider removing the variables with insignificant coefficients. we propose a specific-to-general procedure for variable selection. In both the rules- based and the regime switching approaches we start with a model with only constants included. we calculate for each variable and transition combination the improve- ment its inclusion would yield in the likelihood function. Next. We apply the stationary bootstrap on a monthly frequency. and assume that the series are constant within a month. we draw all four or five corresponding weekly observations. For the RS3L-model. we add the variable to our specification for that specific transition and repeat the search procedure with the remaining variables- transition combinations. which is typically infeasible. There- fore. Not all these variables might be helpful in predicting specific transitions. We lag weekly observations by one week. 1962. Because the predicting variables have a mixed frequency. 1954. weekly observations are avai- lable from January 8. Before that date. If a certain month is drawn.3 with a block-bootstrap.

In the periods 1955–1970 and 1985–2000 long bull markets and short bear markets alternate.] [Table 1 about here. but also their durations and the return distributions. However. switches all take place at maxima and minima.Timmermann (1995). The PS-method shows more bull and bear markets (19 and 18). the LT-method identifies some bull-bear market cycles that last too short to be picked up by the PS-method. The IAD and switching fee allow us to summarize these differences in a single number. On the other hand. but is not identical. 22 . Figure 2 shows which periods are qualified as bullish (white area) or bearish (pink area) by the different methods. This comparison shows in detail for the largest available information set how and why the outcomes of the various methods differ. BIC and R2 . Bull markets last slightly over two years on average. 5 Full Sample Identification We first compare the identification of the different methods for the full sample. who compare all different variable combinations based on general model selection criteria such as AIC. After 2000. the variation in duration is large. so it identifies some bull and bear markets that do not pass the hurdles of the LT-model. The PS-method imposes minimal duration on bull and bear markets instead of minimal changes. In the period 1970–1985 bear markets dominate and prices decline over the years. The shortest bull (bear) market from the LT-method lasts 15 (7) weeks. bear markets slightly longer than one year.] The identification by the PS-method in Figure 2b closely resembles the identification by the LT-method. We consider the actual dating of bull and bear markets. Since the LT-algorithm is based on peaks and troughs. We report summary statistics on the duration of bull and bear markets in Table 1. we see bear market periods with pronounced declines in prices. compared to 27 (15) from the PS-method. The LT-method produces 16 cycles that are spread over our sample period. [Figure 2 about here.

1955–1975 and 1983–1995) and periods of bear markets that are interrupted by brief bull markets (e. The smoothed inference probability for a bull market at time t gives the probability that a bullish regime prevails. and use them to calculate smoothed inference probabilities at each point in time. This indicates that the two regimes are quite distinct.g. since the RS2C and RS2L-model produce a highly similar identification. and that the approach gives a clear indication which regime prevails. we categorize the observation as bullish. In the RS2-models. and rarely equal to values around 0. In Figures 2(c–d) we show the identification by the RS2-models. the volatility ratio is around 2. When using rules-based methods. whereas bear markets ex- hibit negative returns and high volatility. If a bull probability exceeds 0.g. but the ratio of bear to bull market volatility is around 1. The RS2-models differ substantially from the LT and PS models. The variation in duration remains large. We do not see a comparable alternation of bull and bear markets.25.5. while the difference in average returns is 23 . To see why the rules-based methods and the RS2-models produce such different identi- fications. and the longest bear market 78 weeks.. 1979–1983 and 1997–2003). The volatility of bear markets is higher.However. We see that the probability for a bull market is either close to one or close to zero. but instead periods of bull markets that are interrupted by brief bear markets (e.30. The impact of time-variation in the transition probabilities is small. based on the full sample. which is about a full 1% lower when a bear market occurs. since it identifies more cycles. we also plot the series of rounded probabilities. Bull markets have a positive average return and low volatility. To compare the identification with that of rules-based methods. the difference between bull and bear markets is concentrated in the average return. we report the means and volatilities of the bullish and bearish regimes in Table 2. Bull markets last on average 52 to 64 weeks. otherwise as bearish. The number of cycles is much larger at 34 (RS2C) and 43 (RS2L).. bear markets only 16 to 21 months. the longest bull market still lasts 336 or 337 weeks.5. average duration is lower for the PS-method. We estimate the model parameters. Still. Regime switching models pay more attention to volatility. and consequently the duration is considerably shorter. We plot the series of bull probabilities by a thin black line. Consequently we observe the largest difference there.

The RS3C (RS3L) model identifies 24 (25) mild bear markets and 7 (8) strong bear markets. we compared the identification of the different methods by figures or summary statistics. Therefore. which puts the RS3-models more in line with the rules-based approaches. we end up with 17 bearish periods that last on average 69-72 weeks. we see more and longer bear markets in the period 1955-1990. even if prices eventually increase. So far. Table 2 shows that the bullish regime of the RS3-models has the same characteristics as the bullish regime in the RS2-models. For the models with three regimes. the RS2-models identify high volatility periods as bearish. Instead of one bearish regime. which is again quite in line with the rules-based results. the duration shows quite some variation. the identification is between the rules- based methods and the RS2-models. though their maximum is still half a year. whereas strong bear markets last on average 9–11 weeks. They identify 17 bull markets.1. When we aggregate the identification of mildly and strongly bearish regimes. and concentrate on 24 . which compares the identification by two methods on a week-to-week basis. Such comparisons cannot be summarized in one number and may be misleading. we now see two: a mild one with an average return just below zero and volatility similar to the rules-based methods. we calculate the integrated absolute difference (IAD) of Section 3.44%. Again. that last on average 2 years. and low volatility periods as bullish. we aggregate the mildly and strongly bearish regimes. Compared to the RS2-models. and the big drops during the credit crisis in 2008. The presence of the strongly bearish regime is limited to a few periods. notably the big declines by the end of 1974. [Table 2 about here. We consider such a comparison when we calculate the switching fees. That’s why the periods with gradually falling prices in the 1950s are seen as bull markets. and a strongly bearish regime with a large negative average return and very high volatility. Of course.] When we allow three regimes in Figures 2(e–f). Table 1 also shows that the RS3-models are closer to the LT and PS-methods than the RS2-models. even if prices fall. the crash of October 1987. a risk-averse investor may indeed judge volatile price increases as unattractive. Mild bear markets last around 45 weeks. As a consequence. and why the volatile price increase from 1997 to 2000 is qualified as bearish.only around 0.

Since bull and bear markets are highly persistent (see Table C. Even though the summary statistics for the RS3-models were close to those for the rules-based methods. We can say that with a probability of only 6. only 0. Of course. They also indicate that the distribution of the IADs is skewed to the right. but still indicate a proba- bility of around 17% of a different identification.17.247 to 0.313. The confidence intervals are quite wide. which implies that all methods produce significantly different identifications. a difference of 1 simply means that two method produce completely the opposite identification. Within the class of regime-switching models IADs are simply not very precise. [Table 3 about here. for a given drawing the next drawing is random with probability p or the next observation with probability 1 − p.8% the identification by the LT- and PS-methods differs.] The probability that the rules-based methods produce a different identification than the regime switching models is much higher. There we also consider conventional techniques for model comparison.141. For the LT and PS-methods. The IADs between the regime switching models with two or three regimes are smaller. 25 .the bullish versus the two bearish regimes. Because the differences between constant and time-varying transition probabilities are so small. No confidence interval includes zero.032. their IAD can be as large as 0. The average difference between the LT and PS-method is small. we put this probability p at a low value of 0. with a 5% lower bound on the IADs of around 0. as the IAD’s range from 0. Time-varying transition probabilities lead only to minor changes in the identification. In this bootstrap. We use the stationary bootstrap of Politis and Romano (1994) to construct confidence intervals around the estimated IADs. The average IAD for all combinations of two methods in Table 3 can be interpreted as a probability. The 90%-confidence intervals also show the distinction between the rules-based and regime-switching models. so they are bounded between 0 and 1.1). which indicates that the methods can produce quite varying results. the IADs show that the differences between these models are comparable to the differences between the RS2-models and the rules-based models. we postpone a more detailed comparison to Appendix C. with IADs of 0.05.068.

We use the same bootstrap as for the IADs to construct confidence intervals. so the RS2-models lead to a significantly inferior identification. ranging from 7. the results of this investment setting give an upper bound to what could possibly be reached in real-life. The IAD is a statistical measure of the difference between two methods. For example. but also to considerably less volatility of around 11%. the Sharpe ratio is twice as high for the rules-based methods. Still. Therefore. while the various regimes switching models always attribute a non-zero probability to every regime. The regime-switching models lead to smaller returns. So. The difference in magnitude comes to some extent from the fact that the rules-based identification is binary (either a bull market or a bear market prevails). would be willing to pay a considerable fee to switch from the regime-switching models (columns) to the rules-based methods (rows). but it does not tell how important this difference is. We report the performance measures and fees in Table 4. with a coefficient of relative risk aversion equal to five.63% per year to switch from the RS2C model to the LT-method. 26 . The rules-based methods lead to a stunning average return of 48% per year with a volatility of 30%.3% to 9. In Section 3. she would pay up to 18. So. Since we consider identification here. Here we see the influence of qualifying volatile periods with price increases as bearish and tranquil periods with price decreases as bullish. Confidence intervals for the fees to switch from the RS3-models to the rules-based methods are wider. Utility is four to five times higher.2 we derive this fee in an investment setting. if an investor would be able to correctly predict bull and bear markets. this fee corresponds with a situation of perfect foresight. and 18. We find that the fees to switch from the RS2-models to the rules-based methods differ significantly from zero. we also look at the fee that an investor would be willing to pay to switch from one method to another.77% to switch to the PS-model.9% per year. The investor knows with certainty whether a method labels the next period as bullish or bearish. They indicate that the outperformance of the RS3-models by the rules-based methods is less precise. this is her expected performance. and may even include zero. instead of having to predict it. [Table 4 about here.] The fees in Table 4b indicate that a risk-averse investor.

rules-based methods are best to determine these periods. we see which method performs best. The investor would also pay a small fee to switch from models with constant transition probabilities to models where they are time-varying. While the rules-based approach tend to identify relatively long periods of bull and bear markets. it signals whether the economic outlook is good or bad. Here the added value of the predicting variables can be a bit larger. It is 27 . and whether risk premia are low or high. In case of the RS2-models this fee has small confidence intervals. More generally. include zero and indicate that this fee is also often negative.85%. From an investor’s perspective.37–2.29% to 11. the fees that we calcu- late correspond with perfect foresight. An investor would pay a small fee of 1. In the next section. 6 Predictions The state of the stock market being a comprehensive economic indicator. In case of the RS3-models. which is important for identification by regime switching but completely ignored by the rules-based methods. However. Good predictions can tell investors whether to expect a good or bad performance of investments. However. For identification. We conclude that the rules-based approaches produce substantially different bull and bear markets than the regime-switching models. the 90% confidence intervals are wide. we are more inter- ested in predicting it than just identifying it. because they excel in ex-post separating profitable from losing periods. we can just as easily find outperformance. the regime-switching models exhibit periods when bull (bear) markets dominate with short interruptions of bear (bull) markets. (2009) explicitly accommodate rallies during bear markets and corrections during bull markets in their regime switching models. This means that we should not only compare the different methods by their ex-post identification. The driving force behind these differences is volatility. the fee is less precise and can vary from -1. So though we actually observe an underperformance by the RS3-models.14% to switch from the RS3-models to an RS2-model. identification by rules is definitely preferable. which include zero. Also in other economic decisions and analyses where bull and bear market periods are needed. but mainly by the predictions that they yield. when the label of the next period has to be predicted. Maheu et al. rules-based methods are preferable.

For the regime-switching models. and calculate the maximum fees to exchange one method for another. When 52 weeks have passed. the investor combines the most recent parameters estimates with the current information set to predict the state of the market at t + 1. At each point in time t. For the regime-switching models. The PS-method excludes the last 13 weeks from identification. and uses these to make one-step-ahead predictions for the second part of the sample period. An investor uses the first part of the sample period until June 24. For example. she will make her first prediction for t∗ +1. using standard techniques for predictive accuracy and the Intergrated Absolute Differences. To assess the economic importance of these differences we compare the performance of investment strategies based on the different methods. Given the duration of bull and bear markets. 52 weeks offers an acceptable trade-off between estimation speed and accuracy. she updates her information set to construct a prediction for the next week. In the LT-method. the investor expands the estimation window and reestimates the parameters of the different models. (2) to arrive at the prediction for t + 1. and use the recursion in Eq. predictions start at t − 13. Regime switching models may respond quicker to switches. In the PS-method. she knows the sentiment of the market until the last extremum. she first identifies bull and bear markets in the in-sample period. If the last extreme price was at t∗ < t. the investor estimates the parameters over the in-sample period. We compare these predictions in a statistical way. The rules-based methods indicate only after some time what the sentiment of the stock market is. To see which method is best in predicting the future state of the stock market. This approach implies that the set of selected variables varies over time. 1983 for identification and the estimation of model parameters. At every point in time. in particular when a switch has just occurred. the specific-to-general approach is used to select the variables. she applies the filter 28 . When transition probabilities (either in the Markovian logit or the regime switching models) can be time-varying. these methods may fail to correctly predict the future state. If it is unknown which state currently pre- vails.not obvious whether this advantage carries over to forecasting. the LT-method needs a price increase of 20% to categorize a period as bullish. we set up a prediction experiment. and next estimates a Markovian logit model with or without predictive variables. as switches in the last 13 weeks are ignored. When she uses the LT or PS-method. The investor applies the methods in the same way as in the previous section.

whereas for monthly variables the values of the month ending before week t are taken. the hit rate is 71. the values of time t are used. It balances the percentage of hits with the percentage of false alarms. The IAD from predictions by LTC- method differ with a probability of 0.9% of the bull market weeks correctly. but only 43. We also calculate the IAD between the predictions and the realizations of a specific method.] Table 5 reports statistics on the quality of the predictions. The Kuipers score.] [Figure 3 (continued) about here. We present the predictions of the different methods with and without predictive vari- ables in Figure 3. 29 . which equals percentage of correctly predicted bull markets minus the wrongly predicted bear markets. We compare the predictions with the full-sample identification of the different methods. In total. we see that the predictions of the LT-method without predictive variables (LTC-method) lag the ex-post identification. [Table 5 about here. but not very large.to infer the state of the market at time t.1%.194 from the realizations of the LT-method. giving both an equal weight (see Granger and Pesaran. but at other points in time it takes longer before the LTL-method predicts the correct state of the 7 We present and discuss the evolution of the parameters in Appendix D. 2000.6% of the bear market weeks. switches are sooner predicted. until a new maximum (minimum) is reached and the probability jumps back to one (zero). which is lower than the hit rate that results from always predicting a bull market. [Figure 3 about here. is positive. It can take quite some weeks before the LT-method predicts the correct state of the market after a switch. The LTC-method predicts 79. Regarding the weekly predicting variables. Sometimes. Multiplying these inference probabilities with the transition matrix produces the forecast probabilities. In Figure 3a.] Figure 3b shows that the use of predictive variables in the LT-method leads to stronger swings in the forecast probabilities.7 The black lines shows the probability for a bull market at each point in time. for a discussion). We also see that the during a bull (bear) market the probability of a continuation gradually declines.

However. an improvement of 19. The downside of this method are the frequent false alarms that last a couple of weeks. that is revised one or two weeks later. The addition of predictive variables only leads to marginal differences. Compared to the LT-methods. Overall. pointing at a larger difference between prediction and realization.9%. bear markets are more difficult to predict. As indicated by Table 5. as the overall hit rates. This effect is stronger during bear markets than during bull markets. the hit rate is impressive with 85. the improvement over the default bull strategy and the Kuipers score are higher. with a hit rate of 71. without rounding them first to integers. To compare the forecast probabilities of the RS3-models with the other models. but worse of bear markets. compared to the LTC-method.7%. the accuracy is slightly less than with constant transition probabilities. the PSC-method predicts true switches sooner. Figure 3e corresponds with the two-state regime switching models with constant transi- tion probabilities. we also see some false alarms. we first aggregate the probabilities to a bull and a bear probability. The RS2C-model is quicker than the rules-based methods in picking up switches in the state of the market. where the RS2C-model indicate a switch. while the IAD is lower.158.4%. Using predictive variables leads to better predictions of bull markets. Here we compare the predictions with the rounded full-sample smoothed inference probabilities. Overall. switch points are determined by lower bounds on the duration of cycles and phases. also for the regime switching model. The PSC-method scores better at predicting bear markets with a hit rate of 62. where the hit rate decreases to 38. For bull- markets the performance is comparable. A low peak or shallow trough can be (mis)taken for a switch point as long as the restrictions on duration are met. The IAD for the LTL-method is higher than for the LTC-method. The hit rate for the bullish state is 92.6%. The predictions of the PS-method with constant transition probabilities (PSC) in Fig- ure 3c look quite different from the LT-predictions. The IAD can be calculated directly from the fore- cast and the smoothed inference probabilities. The overall hit rate and the Kuipers score are also lower.4%.4% over the “default bull”-strategy. So. this holds in particular for bear markets. In the PS-method.market after a switch. The overall performance compares positively with the LT-methods. The full-sample results in the previous section showed a single bullish regime (positive mean) and two bearish regimes 30 . It indicates an average difference between prediction and realization of 0.

the LT-method with predictive variables are substantially different than the predictions of the PS-method without predictive variables. So while using the LT- method or the PS-method produces largely the same identification. which exceed the hit rates of “default bullish” predictions. this does not necessarily imply that these are useful predictions of bull and bear markets. Therefore.068 according to Table 3. [Table 6 about here. This is a lot larger than the difference between the identification. The predictions of the RS3-models in Figure 3(g–h) are good for bull markets (hit rates are 99. It is difficult to draw conclusions from the accuracy statistics in Table 5. since regime 2 is only mildly bullish or bearish. We conclude that the predictions of the RS3 models are less accurate than the predictions of the RS2-models. Furthermore. but less accurate for bear markets (hit rates of 42. one being mildly and the other strong. balancing the profits of hits of bull and bear markets with the costs of false alarms. which was only 0. Overall. The IADs for the RS3-models are considerably higher than those for the RS2 models. because the accuracy for each method is determined with respect to the ex-post identification of that method. with hit rates around 70%. we aggregate the regimes depending on the sign of their means. while the forecast probabilities of regime switching models may be close to the full-sample smoothed inference probabilities. The average difference between the predictions of the LT-method and the PS-method is approximately 0.6% and 100%). So. 31 . This is partly caused by the aggregation that we choose. Figure D. These differences are due to different ways in which both methods identify the current regime.1(e–f) in Appendix D actually shows two bullish regimes (mild and strong) before 2009. these methods yield quite different predictions.27. Neither can we say whether predictions of. During bear markets. the confidence intervals do not overlap. because the evolution of the regimes is more stable for the RS2-models. predictions oscillate frequently between bullish and bearish.] We report the IAD between the different predictions in Table 6. say.2% and 36. Calculating IADs quantifies the difference between the various predictions.(negative means).8%). The fees indicate which predictions are more valuable to a risk-averse investor. predictions are reasonable.

231 to 0. we evaluate the predictions of the different methods in an investment setting. The differences between the predictions by rules-based methods and those from the regime-switching models vary from 0. The upper bounds on these probabilities confirm that the IADs here are typically small. the IAD is 0. Figure 3a shows that this difference comes from the lag in identifying switches. The volatility of the strategy remains high at 30. As a reference. Taking both means and volatilities into account yields not only different identifications but also different predictions. We plot the cumulative result of this portfolio for each method in Figure 3 (blue lines) and report the corresponding summary statistics in Table 7a. In this setting. these methods produce substantially different results. The confidence intervals are quite narrow. We conclude that predictive variables do not help much when predicting the state of the stock market. They have the same magnitude as the differences between their identifications. The state-dependent means and volatilities are updated together with all other parameters every 52 weeks.27 compares positively to the Sharpe ratio of an investment in 32 . Using two or three states in regime switching models produces largely similar predic- tions. The LTC-method yields the highest return of all methods. w = 1.082 is for the difference between the predictions of the PSC and the PSL-method. It ends up with a cumulative return of 237% over 27 years. The performance of the method is similar during bull and bear markets.095). Finally.331. this result is not surprising. The largest IAD with a value of 0.3% per year.73% per year.090 (0. Using predictive variables does not change the predictions by much. The resulting Sharpe-ratio of 0. Given that the state of the stock market is an important predictive variable of other economic variables itself. So. we add the performance of a strategy that always takes a long position in the stock market. The lower bounds of the confidence intervals also show that these differences are substantial. also in terms of predictions. When we compare that to the results in Table 3. which corresponds 8. (14). Money is lost at the beginning of bull and bear markets.1% that would result from 100% correct forecasts. When transition probabilities are constant (time-varying). the investor combines the forecast probabilities with the state-dependent means and volatilities to construct the optimal portfolio in Eq. we conclude that the two and three-state regime switching models differ more in their ex post identification than in their predictions. This yearly return is considerably less than the 48.

075 and lower values for utility. As a result. The RS2-models give the highest utility of all methods. a volatility of 16.48% to 2. However.15 exceed those of the PS-methods. The PS-methods perform considerably worse than the LT-methods. though they are still lower than the those of the market and the LT-methods. The average returns of these strategies are rather small. [Table 7 about here.053–0. The small magnitude is caused by the smaller magnitude of the expected returns in the RS2-models and the higher volatility of the bearish regime (see Table 2 and Figure D.the market.] Introducing predictive variables worsens the performance of the LT-method.1). but produce a loss of 2. The lagged response to actual switches and the many false alarms have a disastrous result on the performance. which results in a low Sharpe ratio of 0. the overall return decreases from 8. Including time-variation improves the performance during bull markets.43% per year. the many false alarms particularly hurt the performance during bull markets which is lowest among all methods. The false alarms are less of an issue when predictive variables are used. because of the higher volatility that it yields. but performance during bear markets deteriorates. but volatility (9. The cumulative returns are negative for a prolonged period. Volatility is comparable to the LT-methods. The stronger focus on volatility of the RS2-model leads to more prudent forecasts and investments. The utility is actually lower than that of the long position of the market.7% and a Sharpe ratio of 0.1%) is also much lower than for the rules-based methods. volatility increases as well with a dismal effect on utility.13% per year.5–10. The pre- dictive variables lead to a higher return during periods that are ex post identified as bullish.21. with the Sharpe ratio and utility decreasing accordingly. which is meagre at 1.26%. In the PSC-method. The Sharpe ratio still exceeds the ratio of the market.69% during bearish periods.26% per year. 33 . The reason for this good performance is the smaller magnitude of the portfolios that the investor takes when adopting an RS2-model. compared to the LT-methods and the market. as the average return increases from 1.48% to 2. It it the only strategy that performs better during bear markets than during bull markets.52%. As a consequence.73% to 7. the Sharpe ratios of 0. which has a return of 3.97 to 1. ranging from 0.10–0.

and higher utility. in both cases the confidence intervals contain zero. but reverts after 1997. the performance of the RS2L-model is even better. Since the RS2L-model yields the highest utility. with a maximum of 2. the investor would pay a small fee for these variables. indicating that more regimes are not that beneficial for forecasting. In terms of utility. Though its volatility is also low. the RS2L- model outperforms the RS2C-model with a higher Sharpe ratio. The magnitude of these fees is also low. While the LT- methods yield higher average returns and utilities than the PS-methods. the Sharpe ratio remains lowest for this strategy. the Sharpe ratio and utility for the RS3L-model exceed those for the RS2L-model. However. they lead to more accurate predictions.Allowing for time-variation in the transition probabilities improves the performance during bull markets at the expense of the performance during bear markets. The investor does not want to significantly pay for other switches. In particular. she wants to pay a considerable maximum fee of around 17% per year to change from the rules based methods to it. However. The RS3C-model yields the lowest average return of all strategies at 0.75% per year. The confidence intervals indicate that this amount is significant. regime-switching models perform best from a utility perspective. this strategy performs worse than the RS2- models. the investor is willing to pay a fee to adopt this method instead of any other. The fees to switch between the regime-switching models do not exceed 2% per year. The investor actually wants to pay a fee to discard predictive variables in the rules-based methods. A risk averse investor is willing to pay a significant fee to use a regime-switching model. since the average return. When working with regime-switching models. Since they take both means and volatilities into account. but better than the rules-based models and the market strategy. A higher utility of a strategy corresponds with a higher fee for that strategy. Also here. 34 . the confidence intervals for the fees all contain zero. The performance of the RS3-models if Figure3(g–h) is at first better than for the RS2- models. adding predictive variables produces better results. First. The same conclusion applies to all switches from a rules-based method to a regime-switching model. The fees for exchanging one method for another method in Table 7b are derived in a utility setting. less extreme positions and a higher utility. and are also not significantly different from zero. We can draw several conclusions from this analysis.34% per year. In total.

we find difference between the LT and PS-methods that are just as large. It produces a higher average return and a better Sharpe ratio than all other strategies. In the economic comparison of the different methods. whereas the rules-based method just look at the trend of the market. this difference is obvious for the rules-based methods on the one hand. looking just at average returns. As expected. Restrictions on price changes lead to quite different results compared to restrictions on duration. While this method is slow in picking up switches in the state of market. Second.This fee mainly represents the lower volatility that results from regime-switching models. Also for other economic decisions where both the predicted direction and volatility of the stock market are important. as indicated by the low hit ratios. the coefficient of relative risk aversion is a crucial coefficient. The regime-switching models classify volatile periods as bearish. including a long position in the market. the predictions from the rules-based methods and the regime switching models differ considerably. and take short positions. 35 . volatilities and Sharpe ratios. they end up with a loss. and most strategies actually lose money during bear markets. The rules-based methods require parameters to determine which peaks and troughs signal switches between bull and bear markets. Hit rates are higher during bull markets. In this section we analyse how our results change when we change these settings. and the regime switching models on the other hand. Third. Fourth. The rules-based methods are too late spotting the beginning of a bear market and suffer from false alarms. Regime switching models take both means and volatilities into account. 7 Robustness checks Our results and conclusions may be sensitive to some arbitrary choices that we have made in our analyses. If prices increase. Similarity in identification does not imply similarity in predictions. most strategies perform better during bull markets that during bear markets. its performance does not suffer much from false alarms. However. regime switching models are best used. the LT-methods perform best.

The improvements in predictive accuracy are not fully matched by economic improvements. 0. The utility of the (0.20. Even though the changes are statistically small. in particular when both thresholds are lowered.15. The values of 20% and 15% we have used so far have been argued by LT to be most conventional. the IADs between the different identifications indicate that they differ in less than 5% of the weeks. The IADs between the predictions of the LT-methods with different thresholds indicate a different prediction for 7–16% of the weeks.15.20. 0. We compare the performance of the LT-methods with different thresholds as in Sections 5 and 6. Lower thresholds make it easier to identify a switch.7. but utility is sometimes lower. 0.15. fees. and a decrease of more than λ2 for a bear market to begin.10). They also consider lower threshold combinations of (0. but also higher volatilities. We find that the predictive accuracy in- creases for lower thresholds.1.10)-thresholds are positive and significant. We consider these values as well. because our results show that the speed with which the current regime is identified is crucial for good predictions. The fees that an investor is willing to pay to switch to identification with (0.though insignificant. 0.15. For predictions the results are more mixed. Only a lower threshold for λ2 of 0. While larger than the differences for identification.1 Thresholds in the LT-method The LT-method requires an increase in exceedance of λ1 since the last trough for a bull market to start. but may also lead to more false alarms. We discuss the main results here. in particular for bear markets. 0. However. Lower thresholds may be particularly interesting. Sharpe ratios still increase. These results strengthen our conclusion that the LT-method works better for identification than the regime-switching methods. The full results and a detailed discussion are available in Appendix E. The full sample results on identification show that lower thresholds leads to more cycles. The hit rates and Kuipers score of the (0. Lower thresholds lead to investments that produce higher average returns.10 leads to an increase in utility and positive. 0. The choice of thresh- olds does not much affect the pattern of bull and bear markets in Figure 2a. (0.10).10)-thresholds is still substantially lower than 36 . the economic comparison indicates that they present improvements. they are smaller than the differences between the different methods in Table 6.10)-thresholds come close to the results for the two-state regime switching models.15) and (0.

Censoring 7 weeks leads to an extra bear market at the end of the sample period. it means that the conclusion that the PS-method works well for identification is robust to these parameter changes. do we sometimes see accuracy improvements for bull markets at the expense of bear markets. We relax the minimum price change to overrule the phase constraint to 15%. Censoring shows a larger impact. we refer to Appendix E. The IAD between the predictions for different settings and the predictions produced by the basic setting are not significantly different from zero. As PS do not consider robustness checks themselves. but do not improve the economic quality enough to beat them. Neither does a relaxation of the constraints on minimum length or minimum price change largely impact predictions. we consider some changes in the parameters based on our results so far. 7. For the minimum on phase duration we consider 12 and 20 weeks (standard at 16 weeks). We discuss the main results here.the utility for the RS2-models in Table 7. We consider a lower minimum on cycle duration of 52 weeks (instead of 70 weeks). Identification does not change at all when we change the constraints on phase duration or price change. Together. Lowering the cycle constraint to 52 weeks leads to one extra cycle. so we mainly investigate relaxations. but censoring 26 weeks has no effect. For the full results and a discussion. We conclude that lower thresholds bring the statistical quality of the predictions at the same level as the regime-switching models. For identification. The LT-method performs better with relaxed restrictions. The overall predictive accuracy stays the same. and censors a set of first and last observations. The fees for switching are economically small and insignificant. The settings for these constraints are based on the algorithm of Bry and Boschan (1971) for business cycle identification and common market lore.2 Parameters in the PS-method The PS-method use minimum constraints on the length of cycles and phases to select the peaks and troughs that indicates switches between bull and bear markets.2. in particular for the economic difference measures. 37 . changes are negligible. We also investigate the consequences of censoring more (26 weeks) and less (7 weeks).

1b) and typically become less extreme. lower volatilities. These effects can all be explained by the effect that censoring has on a prediction. When more observations at the end are censored. 7. the IADs between the predictions of the standard approach and approaches with more and less sensoring are quite substantial. we have set γ = 5. both effects of risk- aversion exactly cancel out in the Sharpe ratio. and to a better performance. To the contrary. (14) is proportional to the inverse of γ. When 26 weeks are censored. Overall. less extreme predictions lead to less extreme investments.2 depends on the coefficient of relative risk aversion γ. Throughout our analysis so far. This relation caries over to the expected return of an investment strategy. an investor essentially makes a forecast for a longer horizon. In this subsection we show how sensitive our results are to the choice for γ. which is proportional to 1/γ as well. (2). we find that our conclusions are robust to changes in the design of the PS- method. Shrinking predictions towards their long-term average also leads to less extreme investments and a better performance. When predictive accuracy is determined. The effects for censoring show that rules-based methods rely too much on the past direction of the market when making predictions.. and the variance which is proportional to 1/γ 2 . The expected utility being the sum of the 38 . which is conservative though still in the generally accepted range for this parameter. Consequently. utility is close to the utility of long position in the market and only slightly below the utility of the two-state regime switching models. The optimal portfolio w m in Eq. the prediction converges to the long-term average (see e. applying a longer recursion of Eq.3 Risk Aversion The economic measure for comparison that we propose in Section 3. Table C. The economic comparison shows that more censoring leads to less extreme portfolio weights.g. Paying more attention to volatility as regime switching models do leads to better performance. higher Sharpe ratios and higher utility. while censoring less has the opposite effect. If the prediction horizon rises. predictions are rounded.Predictive accuracy is largely unaffected by censoring more or less data. so shrinkage to a long-term average does not matter. However. higher means. Constraints on the duration of cycles or phases or on price changes only have a small effect.

fees that are significantly different from zero remain significant independent of the choice of γ. One way to address identification and prediction is by formulating rules to determine bullish and bearish periods. where we use V (µm /ψ m ) for the expected utility for γ = 1 corresponding with model m   m  n    m  1 µ µ µt 1 2 V − V − 1 − EG r t+1 ηm. we consider the approaches of Lunde and Timmermann 39 .expected return and the second moment weighted by −γ/2 is also inversely related to γ. However.   "  m 2 # m 1 µ 1 µ V (w m ) = EGr w m r − γ(w m r)2 = EGr r − γ r 2 γψ m 2 γψ m "  2 # (21) 1 µm 1 µm = EGr r− r .n produces   m  1 µt ηm. utilities and switching will all decrease (increase) by the same relative amount.n − γη =0 (22) γ ψm ψn r m ψt 2 m. as the proportionality of the switching fee applies similarly to simulated fees. the ordening of the strategies will not be affected and our conclusions regarding the preferred method are robust. In this category.n Solving for ηm. γ ψm 2 ψm Finally. (18) explicit. The confidence intervals around the fees are also proportional to 1/γ. volatilities. the switching fee is proportional to 1/γ. 8 Conclusion In this article we compare the identification and prediction of bull and bear markets by four different methods.n = − 1 − EGr rt+1 ± γ ψtm s   m 2   m  n  (23) µt µ µ  1 − E Gr r m t+1 +2 V − V ψt ψm ψn An increase (reduction) in risk-aversion will lead to less (more) extreme results in Tables 4a and 7a. To derive this result. So. we make the effect of γ in Eq. Means. and then as a second step use binary models for prediction.

Looking at Sharpe ratios. meaning positive average returns and low volatility. and they lead to more prudent investments. Regime-switching models are able to react quicker to bull-bear switches than the rules-based methods. rules-based methods are preferable ex post. and volatile price increases as bearish. We apply the different methods to the S&P 500. an investor can formulate a model that simultaneously handles identification and prediction. The PS-method performs less. Our results for predictions show that paying attention to expected returns and volatil- ity is preferable ex ante. A market that has exhibited price decreases since the last peak is bearish. The switching fee quantifies the preference for one method over another from the perspective of a risk averse investor who wants to time the market. From the identification we conclude that the rules based approaches produce more or less the same results. while negative average returns and high volatility are qualified as bearish. whereas Lunde and Timmermann (2004) impose restrictions on price changes. Be- cause they yield the highest utility. To the contrary. as it produces many false alarms. the regime switching models also pay attention to volatility.(2004) and Pagan and Sossounov (2003). Since periods with low volatility but price decreases can be identified as bull markets. To find switch points Pagan and Sossounov (2003) impose restrictions on the length of cycles and phases. we propose two new measures to compare the identifications and predictions of the various methods. In line with the somewhat depressing results on return predictability in Welch and 40 . The Integrated Absolute Distance is a generally applicable statistic quantifying the difference between the inferred or predicted probability of regimes. are bullish. price increases after the last trough qualify as bullish. Both base identification on peaks and troughs in price data. To deal with the latent nature of bull and bear markets. As an alternative. Their identification is significantly different from the identification by regime-switching models. and an extended version that includes three states. a risk-averse investor is willing to pay a significant fee of around 16% to switch from rules-based methods to regime-switching models. Periods with an attractive risk-return trade-off. the method by Lunde and Timmermann (2004) outperforms the other methods and a long position in the US stock markets. We consider a simple regime switching model with a bull and a bear state. and is worth a significant fee.

but the variables that are selected for predictions and their coefficients vary considerably over our sample period. For the regime switching models we observe improvements in performance. rules-based methods are preferable. Harding and Pagan (2003a. Overall. applied to dating business cycles. the effect is clearly detrimental. the IAD’s between models with and without predictive variables are small. differences are larger where the rules-based approaches purely reflect the tendency of the market. For financial time series. the main differences were the larger transparency for the rules-based approaches versus the deeper insight into the data gen- erating process for the regime switching models.Goyal (2008). regime-switching model are best used. and fees are negligible. For the rules-based approaches. If the state of the stock market is needed in a predictive setting. As the resulting identifications were largely similar. we also find that the inclusion of predictive variables is limited and subject to changes. For applications that require an ex post series of bull and bear markets.b) and Hamilton (2003) have already discussed the differ- ence between rules-based and model-based approaches. with per- formance uniformly worse. while the regime switching models reflect the risk-return trade-off. 41 .

(2006). Review of Financial Studies.. and Swanson. J. and Werner. T. and Diebold. J. and Boschan.. New York. Predicting stock returns. and Potter. C. Coincident and leading indicators of the stock market. A variance decomposition for stock returns. Ang. M. Journal of Econometrics.. USA. F. 19(4):1191– 1239. Nonparametric bootstrap procedures for predictive inference based on recursive estimation schemes. M. Journal of Business Finance & Accounting. M. (1991). C. 42 . Bry. Breaks and persistency: Macroeconomic causes of stock market volatility. Candelon. N. NY. 31(3):719–733. G. and Chordia. Journal of Banking & Finance. Review of Financial Studies. (2007). 101(405):157–179. A. and volatility dynamics. and Xing. C. Y. Journal of Economic and Social Measurement. (2007). (2006). T. Management Science. and Wermers. Beltratti. NBER. Chen. Investing in mutual funds when returns are predictable. direction-of-change forecasting. (1999). and Yu. Bohl.References Ang. 25(3/4):235–258. 48(1):67– 109. (2009). (2000). (2006). 7(1):87–111. J. R. P. and Bekaert. 15(4):1137–1187. Journal of Banking & Finance. On measuring synchronization of bulls and bears: The case of East Asia. 32(6):1022–1035. (2006). and Straetmans. A. and Morana. Predicting the bear stock market: Macroeconomic variables as leading indi- cators. V. A. Piplack. Journal of Empirical Finance.-S. 52(8):1273–1287. Chen. (1971). J. P. (2009). S. Y. S. (2006). L.. Avramov. Journal of Financial Economics. Financial asset returns. D. Economic Journal. F. (2008). T. 36(7-8):1007–1038. Chauvet. Christoffersen. Downside risk. Journal of Financial Economics. Liquidity provision of limit order trading in the futures market under bull and bear markets. Chauvet. M. Do central banks react to the stock market? the case of the Bundesbank. 33:211–223. Chiang. Corradi. Campbell. 82:387–415. Journal of Banking & Finance. Cyclical Analysis of Time Series: Selected Procedures and Computer Programs. Lin. Stock market fluctuations and the business cycle. International asset allocation with regime shifts. T. Avramov. D.. 131(1-2):151–177. International Economic Review.. (2002).. S. G. Siklos. X. 81(2):339–377. R. B..

Journal of Econometrics. Granger. 39(1):1–38. Guidolin. Journal of Finance. A. Diebold. Guidolin.. M. (2008a). (1977). (2003). Review of Economics and Statistics. Journal of Applied Econometrics. A. and St-Amour. Santa Clara University. Review of Financial Studies. 131:285–308. (2001). financial liberalization and volatility. X. pages 283–302. Non-Stationary Time Series Analysis and Cointegration. A. W. The economic value of volatility timing.. Kirby. C. 22(7):925–955. Spanish Economic Review. (1998). A. CA. 90(4):1019–1033. American Economic Review. editor. and Pérez de Gracia. and Uppal. Journal of the Royal Statistical Society: Series B. Systemic risk and international portfolio choice. Size and value anomalies under regime shifts. 31(11):3503–3544. and Pérez de Gracia. F. International asset allocation under regime switching. B. USA. S. An econometric model of nonlinear dynamics in the joint distribution of stock and bond returns. A preference regime model of bull and bear markets. C. Gordon. (1995). Journal of Financial Econometrics. Stock market cycles and stock market development in Spain. and Mariano. and Ostdiek. C. Fleming. Diebold.-H. F. and Timmermann. 13(3):134–144. (1989). P. 21(1):1–22. B. Lee. and Pesaran. R. Santa Clara. Guidolin. (2004). (2006a). Predicting US recessions: Financial variables as leading indicators.. M. (2003). C. Econometrica. H. D.. (2000). X. A. J. S. F. Oxford. S.Das. 6(1):1–48. P. N. and Mishkin. Gómez Biscarri. and Weinbach. (1994). 57:357–384. A new approach to the economic analysis of nonstationary time series and the business cycle. Journal of Forecasting. Comparing predictive accuracy. and Timmermann. Hamilton.. S. A. Journal of Economics Dynamics & Control. Journal of International Money and Finance. F. Laird. Term structure of risk under alternative econometric specifications. Gómez Biscarri. S. J. 6(2):127–151. Oxford University Press. G. 56(1):329–352.. Maximum likelihood from incomplete data via the EM algorithm. Dempster. M. 21(2):889–935. A. Economic and statistical measures of forecast accuracy. M. Guidolin. Asset allocation under multivariate regime switching. Stock market cycles. Journal of Business & Economic Statistics. M. and Timmermann. and Timmermann.. Estrella. 80(1):45–61. UK. Regime switching with time-varying transition probabilities. 19(7):537–560. (2000). Working paper.. Edwards. R. (2006b). J. In Hargreaves. and Rubin. J. R. (2007).. skew and kurtosis preference. J. and Timmermann. D. Guidolin. 43 . (2008b). J. M. M. F.

and Rockinger. Princeton University Press. Hamilton. L. 27(9):1695–1698. 45(5):38–45. D. and Schmukler. On information and sufficiency. and Pagan. G. T. J. Journal of Economic Dynamics and Control. Harvey. M. E. H. 22(3):253–273. A. (2004). Kaminsky. Harding. Stock market volatility and the business cycle. (1994). 22(1):79–86. E. Time Series Analysis. Journal of Applied Econometrics. D. Journal of Economic Dynamics and Control.. and Verbeek. S. R. Short-run pain. Comment on “a comparison of two business cycle dating methods”. J. (1990). Harding. Journal of Econometrics. A. and Pagan. Harding. (2006). D. C. and Lin. (2003). D. Journal of Monetary Economics. 18(1):100–112. 27:1681–1690. Hamilton. Dynamic linear models with markov-switching. 55(3):1263–1295. Harvey. J. Koedijk. (2006). Forecasts of economic growth from the bond and stock markets. (2000). and Leibler. Dissecting the cycle: A methodological investigation. 12(2):253–292. D. 60(1):1–22. J. Lunde. 27(9):1691–1693. R. Harding. Duration dependence in stock prices: An analysis of bull and bear markets. S. NJ. Rejoinder to james hamilton. (1951). Princeton. 44 . C. and Siddique. Financial Analysts Journal. (2002a). 45(1-2):39–70. A. D. long-run gain: Financial liberal- ization and stock market cycles. Kim. Optimal portfolio allocation under higher moments. 12(1):29–55. A. Conditional skewness in asset pricing tests. (1989). (2003a). Journal of Business & Economic Statistics. M. and Timmermann. A comparison of two business cycle dating methods. Kullback. (2008). M. Journal of Banking & Finance. R. Hamilton. Journal of Economic Dynamics and Control. A. D. J. European Financial Management. 27(9):1681–1690. Journal of Economic Dynamics & Control. Maheu. Identifying bull and bear markets in stock returns. (1994).-J. 11:573–593. 49(2):365–381. Portfolio implications of systemic crises. A. Journal of Econo- metrics. Journal of Business & Economic Statistics. and Pagan. USA. A. 30(8):2347–2369. Annals of Mathematical Statistics. (2003b). C. D. A comparison of two business cycle dating methods. Kole. Jondeau. A. G. Review of Finance. K. (2002b). (2000). Journal of Finance. Analysis of time series subject to changes in regime. and Pagan.Hamilton.. and McCurdy. L. (1996).

M. W. editors. Oxford. D. A. 89(428):1303–1313. Schwert. J. Granger. (2003). Journal of Forecasting. and Ullah. Marcellino. Basil Blackwell. and Hendry. Journal of Business. (1995). (1961). (2009). F. H. H. H. T. G. Pesaran. E. S. Moore. R. reprinted in: G. Journal of Finance. Mitchell. Journal of Applied Econometrics. The economic value of predicting stock index returns and volatility. heteroskedasticity and autocorrelation consistent covariance matrix. ch. In Clements. Netherlands. 118(2):639–669. G. ed. A simple framework for analysing bull and bear markets. (1994). 39(2):407–429. (2003). M. R. H. C. and Sossounov. Macro variables and international stock return predictability. L. 63(3):399–426. Handbook of Economic Forecasting. (2006).. Measuring the reaction of monetary policy to the stock market.. (1938). and West.s. 6. and Sack. UK. E.. Quarterly Journal of Economics. Sarno. University of Toronto. 23(8):541–557. Predictability of stock returns: Robustness and economic significance. and Timmermann. 55(3):1229–1262. A. CA. (2000). (1987). Amsterdam. Comparing the accuracy of density forecasts from competing models. Leading indicators: What have we learned? In Elliot.. A. Business cycle indicators. G. D. J. 18(1):23–46. Econometrica. 50(4):1201–1228. Marquering. A. Rapach. M. Pesaran. and Timmermann.. Working paper. M. J. Indexes of u.Maheu. Princeton NJ. (2005). M. Politis. Journal of the American Statistical Association. M. and Rangvid. and Valente. P. K. Perez-Quiros. Elsevier. (1990). USA. and Romano. International Journal of Forecasting. (1999). B. Journal of Financial and Quantitative Analysis. editors. W. Princeton University Press (Princeton). A simple. and Burns. A Companion to Economic Forecasting. Shiller. (2004). Journal of Finance. chapter 11. M. 45 . McCurdy. F. and Song. A. Irrational Exuberance... R.. A. Firm size and cyclical variations in stock returns. N. NBER (New York). W. positive semi-definite. Extracting bull and bear markets from stock returns. The stationary bootstrap. Y. and Verbeek.. 55(3):703–708. G. pages 879–960. Nonparametric Econometrics. A. D. Pagan. (2004). UK. Cambridge. J. C. W. K. and Timmermann. (2000). A. Pagan. Wohar. W. Newey. Cambridge University Press. Decision-based methods for forecast evaluations. Princeton University Press. Rigobon. D. stock prices from 1802 to 1987. Statistical indicators of cyclical revivals. P. and Skouras. pages 241–267. (2002). K. 21(1):137–166.

Stock market overreaction to bad news in good times: A rational expectations equilibrium model. D. Review of Financial Studies. M. (1993). and Goyal. and Watson. (2003b). (2003a). D. K. West. Journal of Econometrics. 89:71–90. (2000). 96(1):75–111. Veronesi. Stock. Forecasting output and inflation: The role of asset prices. Timmermann. Welch. A reality check for data snooping. 41(3):778–829. Edison. H. H. (2008). I. 64(5):1067– 1084. J.Stock. (1999). A comprehensive look at the empirical performance of equity premium prediction. 35:23–45. West. Journal of Economic Literature. Econometrica. 46 . J. A. W. Moments of markov switching models. M. How did leading indicator forecasts do during the 2001 recession? Federal Reserve Bank of Richmond Economic Quarterly. Econometrica. A. and Watson. and Cho. P. White. 12(5):975–1007. K. (2000). H. Journal of International Economics. Review of Financial Studies. H. 68(5):1097–1126. D. A utility-based comparison of some models of exchange rate volatility. 21(4):1455–1508. J.. (1996). W.. Asymptotic inference about predictive ability.

Bt ≡ τ =0 (1 + rτf ). 47 . we use the S&P500. Data have been taken from FRED at the Federal Reserve Bank of St. For the stock market index. Figure 1: Performance US Market 400 350 300 250 200 150 100 50 0 Jan-55 Jan-57 Jan-59 Jan-61 Jan-63 Jan-65 Jan-67 Jan-69 Jan-71 Jan-73 Jan-75 Jan-77 Jan-79 Jan-81 Jan-83 Jan-85 Jan-87 Jan-89 Jan-91 Jan-93 Jan-95 Jan-97 Jan-99 Jan-01 Jan-03 Jan-05 Jan-07 Jan-09 This figure show the weekly observations of the US stock market in excess of the risk-free rate over the period January 7. 2010 (1/7/1955 = 100). Louis and Schwert (1990). The excess stock market index is calculates as the ratio Pt /Bt . where Pt is the value of the stock market index and Bt is the cumulation of a riskless Qt−1 bank account. 1955 until July 2. The risk-free rates is the three-month T-Bill rate.

6 0. and panels d and f.8 0.5 200 0.2 50 50 0.9 350 350 0. Figure 2: Identification of Bull and Bear Markets 400 1 400 1 0. 48 . respectively.3 0.5.9 0.5 0. Bullish (bearish) regimes are indicated with white (pink) areas.5 0.4 150 150 0.4 0.6 200 0. Panels (c–d) and (e–f) reflect identification by two-state and three-state regime switching models.8 300 300 0.7 250 250 0.3 100 100 0.6 200 0.8 0.4 150 150 0.2 50 50 0.1 0. In panel (a). A think black line indicates the smoothed inference probability of a bull market (right y-axis).1 0 0 0 0 Jan-55 Jan-57 Jan-59 Jan-61 Jan-63 Jan-65 Jan-67 Jan-69 Jan-71 Jan-73 Jan-75 Jan-77 Jan-79 Jan-81 Jan-83 Jan-85 Jan-87 Jan-89 Jan-91 Jan-93 Jan-95 Jan-97 Jan-99 Jan-01 Jan-03 Jan-05 Jan-07 Jan-09 Jan-55 Jan-57 Jan-59 Jan-61 Jan-63 Jan-65 Jan-67 Jan-69 Jan-71 Jan-73 Jan-75 Jan-77 Jan-79 Jan-81 Jan-83 Jan-85 Jan-87 Jan-89 Jan-91 Jan-93 Jan-95 Jan-97 Jan-99 Jan-01 Jan-03 Jan-05 Jan-07 Jan-09 (a) LT (b) PS 400 1 400 1 0.9 0.2 0.4 0.6 200 0.1 0.1 0 0 0 0 Jan-55 Jan-57 Jan-59 Jan-61 Jan-63 Jan-65 Jan-67 Jan-69 Jan-71 Jan-73 Jan-75 Jan-77 Jan-79 Jan-81 Jan-83 Jan-85 Jan-87 Jan-89 Jan-91 Jan-93 Jan-95 Jan-97 Jan-99 Jan-01 Jan-03 Jan-05 Jan-07 Jan-09 Jan-55 Jan-57 Jan-59 Jan-61 Jan-63 Jan-65 Jan-67 Jan-69 Jan-71 Jan-73 Jan-75 Jan-77 Jan-79 Jan-81 Jan-83 Jan-85 Jan-87 Jan-89 Jan-91 Jan-93 Jan-95 Jan-97 Jan-99 Jan-01 Jan-03 Jan-05 Jan-07 Jan-09 (e) RS3C (f) RS3L This figure shows the identification of bull and bear periods for the US. and is indicated in purple (only for panels e–f).7 250 250 0.2 0.8 300 300 0.3 0.7 0.6 0.5.5 200 0. The thick blue line plots the excess stock market index (left y-axis). when the smoothed inference probability for regime 1 (2) exceeds 0.3 100 100 0.4 0.5 200 0.2 50 50 0.5 0.6 0. respectively).3 0.8 0.7 250 250 0.9 0.9 350 350 0. and in panel (b) by the PS-algorithm. The strong bear regime prevails when the smoothed inference probability for regime 3 exceeds 0.1 0 0 0 0 Jan-55 Jan-57 Jan-59 Jan-61 Jan-63 Jan-65 Jan-67 Jan-69 Jan-71 Jan-73 Jan-75 Jan-77 Jan-79 Jan-81 Jan-83 Jan-85 Jan-87 Jan-89 Jan-91 Jan-93 Jan-95 Jan-97 Jan-99 Jan-01 Jan-03 Jan-05 Jan-07 Jan-09 Jan-55 Jan-57 Jan-59 Jan-61 Jan-63 Jan-65 Jan-67 Jan-69 Jan-71 Jan-73 Jan-75 Jan-77 Jan-79 Jan-81 Jan-83 Jan-85 Jan-87 Jan-89 Jan-91 Jan-93 Jan-95 Jan-97 Jan-99 Jan-01 Jan-03 Jan-05 Jan-07 Jan-09 (c) RS2C (d) RS2L 400 1 400 1 0.7 0.7 0.8 300 300 0.4 150 150 0. bull and bear markets are identified by the LT-algorithm. based on the different approaches.2 0.3 100 100 0. A bull (bear) market prevails in the RS-models.9 350 350 0. These models can have constant or time-varying transition probabilities (panels c and e.1 0.

3 0.8 0.7 0.9 1 0 0.8 0.2 0.1 0.3 0.1 0.3 0.1 0.4 0.2 0.5 0.2 0.1 0.6 0.9 1 0 0.3 0. Figure continues on next page.5 0.7 0.8 0.1 0.9 1 .1 0.7 0.4 0.7 0.9 1 0 0.5 0.5 0. -30 -20 -10 -100 -100 10 20 30 40 50 60 100 100 150 200 250 300 -80 -60 -40 -20 -50 0 20 40 60 80 50 0 0 1-7-83 1-7-83 1-7-83 1-7-84 1-7-84 1-7-84 (a) LT.9 1 0 0.3 0.9 1 49 -30 -20 -10 100 -100 -80 -60 -40 -20 10 20 30 40 50 60 100 150 200 250 300 20 40 60 80 -50 0 50 0 0 1-7-83 1-7-83 1-7-1983 (b) LT.5 0.6 0.4 0.4 0.4 0.6 0. constant transition probabilities (c) PS. time-varying transition probabilities (d) PS.8 0.6 0.6 0.7 0.2 0.8 0.4 0. constant transition probabilities 1-7-85 1-7-85 1-7-85 1-7-86 1-7-86 1-7-86 1-7-87 1-7-87 1-7-87 1-7-88 1-7-88 1-7-88 1-7-89 1-7-89 1-7-89 1-7-90 1-7-90 1-7-90 1-7-91 1-7-91 1-7-91 1-7-92 1-7-92 1-7-92 1-7-93 1-7-93 1-7-93 (e) RS2C 1-7-94 1-7-94 1-7-94 1-7-95 1-7-95 1-7-95 1-7-96 1-7-96 1-7-96 1-7-97 1-7-97 1-7-97 1-7-98 1-7-98 1-7-98 Figure 3: Predictions and performance 1-7-99 1-7-99 1-7-99 1-7-00 1-7-00 1-7-00 1-7-01 1-7-01 1-7-01 1-7-02 1-7-02 1-7-02 1-7-03 1-7-03 1-7-03 1-7-04 1-7-04 1-7-04 1-7-05 1-7-05 1-7-05 1-7-06 1-7-06 1-7-06 1-7-07 1-7-07 1-7-07 1-7-08 1-7-08 1-7-08 1-7-09 1-7-09 1-7-09 0 0.7 0.8 0. time-varying transition probabilities 1-7-84 1-7-84 1-7-1984 1-7-85 1-7-85 1-7-1985 1-7-86 1-7-86 1-7-1986 1-7-87 1-7-87 1-7-1987 1-7-88 1-7-88 1-7-1988 1-7-89 1-7-89 1-7-1989 1-7-90 1-7-90 1-7-1990 1-7-91 1-7-91 1-7-1991 1-7-92 1-7-92 1-7-1992 1-7-93 1-7-93 1-7-1993 (f) RS2L 1-7-94 1-7-94 1-7-1994 1-7-95 1-7-95 1-7-1995 1-7-96 1-7-96 1-7-1996 1-7-97 1-7-97 1-7-1997 1-7-98 1-7-98 1-7-1998 1-7-99 1-7-99 1-7-1999 1-7-00 1-7-00 1-7-2000 1-7-01 1-7-01 1-7-2001 1-7-02 1-7-02 1-7-2002 1-7-03 1-7-03 1-7-2003 1-7-04 1-7-04 1-7-2004 1-7-05 1-7-05 1-7-2005 1-7-06 1-7-06 1-7-2006 1-7-07 1-7-07 1-7-2007 1-7-08 1-7-08 1-7-2008 1-7-09 1-7-09 1-7-2009 0 0.5 0.2 0.2 0.6 0.3 0.

5 30 0. The regime-switching models apply the standard filter technique with the last available model parameters.9 70 0. (15) and (16).1 -10 0. The pink areas indicate the periods of bear markets as identified based on the full sample as in Figure 2. (2) from the last extremum onwards. Figure 3: Predictions and performance – continued 70 1 80 1 60 0. m m with risk aversion parameter γ = 5.4 10 0. The PS-method starts this recursion 13 weeks prior to week t.8 60 0. giving a total of 1410 predictions. The values for the mean µt+1 and second moment ψt are calculated as in Eqs. To make a prediction for week t+1.6 40 0.2 -10 0. we calculate the probability of a bull market as the sum of the predictions for regimes with positive means. The blue line shows the cumulative result (in %) of this portfolio.5 20 0. The parameters in the Markovian logit models for predictions in the rules-based approaches. the state-dependent mean µm m s and second moment ψs .3 0 0. The asset allocation for method m at time t equals wtm = µm m t /(γψt ).6 30 0.7 50 0. and the parameters of the regime switching models are updated every 52 weeks.9 0.7 40 0.1 -20 0 -20 0 1-7-83 1-7-84 1-7-85 1-7-86 1-7-87 1-7-88 1-7-89 1-7-90 1-7-91 1-7-92 1-7-93 1-7-94 1-7-95 1-7-96 1-7-97 1-7-98 1-7-99 1-7-00 1-7-01 1-7-02 1-7-03 1-7-04 1-7-05 1-7-06 1-7-07 1-7-08 1-7-09 1-7-83 1-7-84 1-7-85 1-7-86 1-7-87 1-7-88 1-7-89 1-7-90 1-7-91 1-7-92 1-7-93 1-7-94 1-7-95 1-7-96 1-7-97 1-7-98 1-7-99 1-7-00 1-7-01 1-7-02 1-7-03 1-7-04 1-7-05 1-7-06 1-7-07 1-7-08 1-7-09 (g) RS3C (h) RS3L This figure shows for each method the predictions and the evolution of the investment performance over time.4 20 0. the LT-method follows the recursion in Eq. 1983 and the last for July 2.8 50 0. 2010.2 0 0.3 10 0. The first prediction is made for July 1. For the RS3-models. We plot the predicted probability of a bull market with a black line (secondary y-axis). 50 . All approaches use weekly explanatory variables up to week t and monthly explanatory variables up to the last month prior to week t.

duration 60 60 12 10 41 42 41 42 std. their average and median duration. the standard deviation of the duration. duration 8 3 std. and its maximum and minimum. For the three-state regime switching models. duration 405 276 336 337 314 314 313 313 min. duration 15 27 2 2 13 13 5 5 (mild) bear number 16 18 34 43 24 17 25 17 avg. duration 11 9 med. duration 97 61 77 90 90 90 91 91 max. duration 10 12 max. duration 7 15 1 1 8 8 5 9 strong bear number 7 8 avg. In the columns RS3C∗ and RS3L∗ . duration 1 2 This table shows for every method the number of spells of the different market regimes. duration 28 29 min. dev. the mildly and strongly bear markets have been aggregated. 51 . duration 119 95 64 52 102 102 98 98 med. a period is qualified as bullish if the smoothed inference probability for the bull regime exceeds 0. Table 1: Number and Duration of Market Regimes approach LT PS RS2C RS2L RS3C RS3C∗ RS3L RS3L∗ bull number 16 19 34 43 17 17 17 17 avg. For the two-state regime switching models. duration 62 61 21 16 45 69 46 72 med. duration 90 74 24 23 64 64 62 62 std. duration 44 30 17 23 23 66 32 79 max. dev. dev.5 and bearish otherwise. duration 187 132 78 78 91 261 163 327 min. the highest smoothed inference probability determines the prevailing regime.

2.46 (0.04) 1.24) 3.04) 0. columns RS2C and RS2L) and with three regimes (bull.53) 5.50 (1.30 (0.16) 2.16 (0.16 (0.04) 1.08) −0.47 (0.04) 0.04) vol.15) 3.08) −0.28 (0.27 (0. For the regime switching models. We report standard errors in parentheses. and strong bear.36 (0. For the LT.62) 52 vol.60 (0.04) 0. In the regime switching approach.06 (0.65 (1. .36 (0.16) 2. we estimate the model for the return distributions with two regimes (bull and bear.14) −0.34 (0.25) 2.54 (0.06) 1.30) This table shows the mean and the volatility (in % per week) for the different regimes under the different approaches. Table 2: Return Characteristics of Bull and Bear Markets regime LT PS RS2C RS2L RS3C RS3L bull mean 0. In the approach of LT and PS identification is based on peaks and troughs in the prices series. columns RS4C and RS4L).47 (0.04) 0.74 (0.38 (0. we use the Fisher information matrix to compute standard errors.95 (0.40 (0.16 (0.and PS-methods we calculate HAC consistent standard errors of Newey and West (1987) in a GMM setting.13) −0.82 (0. mild bear.08) vol.04) 0.07) 1.15 (0. we estimate the means and volatilities of the returns distributions.07) −0.43 (0.04) 1.73) −0. 1. Conditioning on the regimes.04) (mild) bear mean −0.38 (0.86 (0.15) strong bear mean −0.39 (0.04 (0. 5.

015. 0.454] [0.214.268 0. 0.355] [0. 0. 0.and PS-method. 0.081. we incorporate Eq.247 0. 0.311 0. (10). we aggregate the mild and strong bear regimes.171. and the RS2-models we use Eq.591] RS3C 0.032 0.068 0.365] [0.098. Table 3: Integrated Absolute Differences of Identification PS RS2C RS2L RS3C RS3L LT 0. based on two states.182 [0.505] PS 0. 0. (11) into the calculation. When the RS3-models are part of the comparison.168 0.034. 53 . 0. 0.088. 0.271] This table reports the integrated absolute distance between the identification of the different approaches.201.220.272 [0.084.154 0.187. For comparisons between the LT.277 0. 0. For the difference between the LT. 0. 0.200. We report the 5% and 95% percentiles between brackets for each statistic based on the 200 bootstraps of both the return and the explanatory variable series following Politis and Romano (1994).061] [0.368] [0.291 0.602] RS2L 0.313 [0.022. 0.141] [0.483] [0.and PS-method on the one hand and the RS-models on the other hand.234 0.486] RS2C 0.214.032 [0.564] [0. 0.220.555] [0.167 [0.355] [0.

96] [-12.63 -19.34. 2.42 20.6 11. 17. 21.62] [-2.59 0.25.52 1.3 30. we calculate the mean and volatility (in % per year).3 Sharpe ratio 1. -2.67.14 1.96. 1.3 10.94] [10.05.85] This table reports performance measures of investment strategies that use the different methods as input for an asset allocation and the fees an agent would be willing to pay to exchange two methods.045 0. 24. The state-dependent mean µm m s and second moment ψs are based on the full-sample and reported in Table 2.4 11. which are binary for the rules-based methods and smoothed inference probabilities for the regime-switching approaches.77 18.34] [-2. The state-dependent probabilities are taken as the identification at time t.10 20.01. -10.80.45.64] [-2.91.59 1.0 Volatility 30.85] [-1. -2.90. and the annualized realized utility as in Eq. (18). 1.048 (b) Fees LT PS RS2C RS2L RS3C RS3L LT 0 -0.11. 24.37.65] [-2.066 0. 22.40] [10.21] [-21.26 -18.08] [9.241 0.24 20.02. 5.02] [10. -9. 12.62 -2.15] [-23.25 0 [-21. 1.88 -20. The fee ηm.3 11. -9.38] PS 0.17.74 -17.07.71] [-22.25 [-22. . Table 4: Performance Measures and Fees Based on Full-sample Identification (a) Performance Measures LT PS RS2C RS2L RS3C RS3L Mean 48.49] RS2C -18.39 0 -0.34] [-17.37 54 [-23.03] [-1.44.13 18.80. We report the 5% and 95% percentiles between brackets for each statistic based on the 200 bootstraps of both the return and the explanatory variable series following Politis and Romano (1994). 1.29 20. 4.87 0. 24. 2. 23. We express the fees in % per year.02] [2.16 [-4.88] [-21.89 0.01 -1.86.87 0.58] [-16.69 0. 21.n to switch from strategy m (in rows) to strategy n (in columns) solves Eq.36.89 [-23.37 -1. The values for the mean µt+1 and second moment ψt are calculated as in Eqs.97.3 8. 22.93. (15) and (16). 11.57.30.63 18.14 0 -0.35.62 1.03 [-5.9 7.06] [-11.83 0.13 0 18.82] [0.07] [-12.1 48. the yearly Sharpe ratio.17] RS2L -17.90] RS3C -19.52 0 2. -10.76 -1.4 9. -0. 17.20] [-1. Based on the realized returns of the asset allocations. 12.34] [2.13.061 0.58.71 Utility 0.4 9. The asset allocation for method m at time t equals wtm = µm m t /(γψt ). with risk m m aversion parameter γ = 5. 2.45.92.17] [-1. (13). 2.32.29] RS3L -19.29.242 0.

The predicted probabilities are rounded.9% 81. “Bull (bear) correct” gives the number of true bullish (bearish) weeks that were correctly predicted.4% 59.4% 73.6% 38. when the models would always predict a bullish week.6% 19.8% IAD 0.4% 48.6% 68.4% 62. The Kuipers Score is calculated as the percentage of correctly predicted bull markets minus the percentage of incorrectly predicted bear markets.222 0. We apply Eq.197 0.8% 36.2% 68.8% 85. The predictions are constructed as in Figure 3. while their realizations are binary.318 This table shows the predictive accuracy of the different methods. 55 .1% 92.1% 41. The row “improvement” shows by how much a method’s percentage of correct predictions exceeds the “default bull” prediction. since their predictions are probabilities. Table 5: Predictive Accuracy LTC LTL PSC PSL RS2C RS2L RS3C RS3L bull correct 852 851 768 779 868 935 708 681 bull wrong 214 215 193 182 66 65 3 0 % bull correct 79.2% 66. For the regime switching models the resulting smoothed inference probabilities are rounded.0% bear correct 150 132 281 261 340 269 295 268 bear wrong 194 212 168 188 136 141 404 461 % bear correct 43.157 0.5% 18.3% improvement -4.3% default bull 75.1% 71.0% Kuipers Score 23.7% 74.9% 79. “Bull (bear) wrong” gives the number of true bullish (bearish) weeks that were wrongly predicted.4% 71.4% 65.1% 67. The percentages are calculated with respect to the number of true bullish (bearish) weeks.2% 70.9% 50. and compared with the identification that results from the full sample.5% 20.8% total correct 1002 983 1049 1040 1208 1204 1003 949 total false 408 427 361 370 202 206 407 461 % correct 71.5% 99. (10) to calculate the IAD for the rules-based methods. we calculate the probability of a bull market as the sum of the predictions for regimes with positive means.1% 69.7% 85. The row “default bull” reports the percentage of total correct predictions. The row IAD reports the Integrated Absolute Distance between the predictions and the realizations. For the RS3-models.2% 36.7% 19.194 0.2% 5.283 0.9% 6.8% 79.9% 93.6% 58. too.6% 42.2% 64.158 0.5% 39.177 0.4% 14.6% 100.5% -5.6% 75.2% 42.

374] 56 RS2C 0.328] PSL 0.118] RS3C 0.244 0.278.059. 0.364] [0.262 0.216.110] [0.374] [0. 0. Table 6: Integrated Absolute Differences of Predictions LTL PSC PSL RS2C RS2L RS3C RS3L LTC 0.239 0.126] RS2L 0. (10).239.276.379] [0. 0.114] [0. 0.103] [0.303] [0.189.071.077] [0.305 0. 0. 0.048.060 0.332] [0.275 0.386] [0. 0.384] [0. 0. 0.271. 0.060.082 0.282. When the RS3-models are part of the comparison. 0. We calculate the probability of a bull market as the sum of the predictions for regimes with positive means.237.279 0.352] PSC 0.288] [0. 0.269 0. .332] [0.212.356] [0. 0.239 0. based on two states. 0.061 0.345] [0.178.090 0.272 0.284 0.086 0.294. 0.367] [0.331 0. 0.338] [0.074.116 [0.029.228 [0.360] [0.203. 0.193.210.327 0. For the difference between all two-state methods we apply Eq. 0. 0.250 0. 0.145] [0.318 [0.054] This table reports the integrated absolute distance between the predictions of the different approaches.281 [0.236.324] LTL 0. 0.276 0.253.255 [0.034 [0.057.320] [0.332] [0. 0. 0. The predictions are constructed as in Figure 3.213 0.223. 0. We report the 5% and 95% percentiles between brackets for each statistic based on the 200 bootstraps of both the return and the explanatory variable series following Politis and Romano (1994).072.194. 0. we concentrate on the predictions of bullish regimes.230. 0.201.284 0.095 [0. 0. 0.

60.30. 3.74 -1.51] [8. -8. 8. 1.77 [7.98] [-1.55 -17. We express the fees in % per year.013 -0. 25.36] [7.50.15.64] [-25. -7.86] [8.74 -3.56 -4.7 32.53] [-26.80] [-6.50 0.4 31.9 Sharpe Ratio 0. the yearly Sharpe ratio.44.80.82] [-0.45 [-8.15 0.34. -8.75 -16.02.87 0.61 3.92.62 0.05 0. The optimal portfolio is constructed similarly as for Figure 3.48 2. -9.1 9.60.56.64 Mean 3.43.28.26.13 1.28.26] RS3L 15.56 0 -0. 11.84.73.04.26 0 1.53] 57 PSC -0.68] [-3.64] [-24. 26.05] PSL -2. 1.66 Mean Bear .30.17] [-6. 2. 1.73 15. 22.28 16.60 -16. 1. 6. -8.97] [-27.82.43] [-2. we report the statistics of a long position in the market.82.12] [-26. 25. Based on the realized returns of the asset allocations.50] [-23.62 -1. For the purpose of comparison.52 -0. 23.15] [-3. 1.07.42] [9.0 30.90 1. 24.39 -0. 6.15.19 18.9 28. -10.26 1.34 6. Mean Bull (Mean Bear) is the annualized mean during the subperiods identified ex post as bull (bear) markets. 23.11.29 0.36] [-1. 5. 8.44] [-1.98 -17.12.54] [-11.27 -16.70 16.86 [-1. and the annualized realized utility as in Eq.97 5.48] [-1.34 1.37.23.86] [-7.50 -3.61 0.93 -2.82.09] [-25. 0.62 4.01 -2.31 0 2.71.8 10.73 7.44 8.07 -16.40 -17.n to switch to strategy m (in rows) from strategy n (in columns) solves Eq.69 2.64.036 -0.33] [-26.176 -0.5 10.19] [-8.52 [8.018 (b) Fees LTC LTL PSC PSL RS2C RS2L RS3C RS3L LTC 0 0.68.30.24 -16.86.76] [8.96 -18.82 0.97 19.32 1.18] LTL -0.02] [-25. 8.22 0.181 -0.94] [5.78 [8. 25.30 -1.42 17.97 1.66 0.67] [-25.36. 25.82] This table reports performance measures of investment strategies that use the different methods as input for an asset allocation and the fees an agent would be willing to pay to exchange two methods. -8.35] [-2.66. Weight .90 0 -0.38. 5. 12. 6.37.70.35] RS3C 14.75 [-9. 21. The fee ηm.13. 0. (13).30 0.25 -15.26 0.50 16. 8.91] [6.74] [7.69] [-23.06 [-5.59.42] [-22.45.85] [5.48 17. 24.12] [-5.50 2.32] [-25.68] [-26.27 0. Table 7: Performance Measures and Fees Based on Prediction (a) Performance Measures market LTC LTL PSC PSL RS2C RS2L RS3C RS3L Av.17 0.70 1. we calculate the mean and volatility (in % per year).17 Mean Bull . (18).18.77] [-12.62.32.87] [-24. -6.21 0.28 -19.97. -5.93 15.63 4. .03 0.96 10.01 16.90.1 10.82 1. 1. -8. 2.010 -0.05.51. 22.24 -5. 9.10] [-2.70 18. -7.15. We report the 5% and 95% percentiles between brackets for each statistic based on the 200 bootstraps of both the return and the explanatory variable series following Politis and Romano (1994).026 -0.10 0.51.94 -17.21 17. 25.67 -16. -8. -8.91 -4.78 0.14.96 Volatility 16. -5.69.86 -15. 24.203 -0.87] [10.88. 25.93] [8. -8.96] RS2L 16.98.66 -16.25 0 -19.52 0.184 -0. -7.77 0 [7. 1. Abs.64] [-1.68] [-0. 7.39] [-6.81 0 -0.08 0.11.44] [-25. 24.78] [7.56 0. 3.82.11 Utility -0. 2.92] RS2C 16.

(25) ∂ℓ(B) X X T 1 ∂πςq = ξςq.2 we propose a regime switching model with time-varying transition proba- bilities.1 Estimation To estimate the parameters βsq . (1977).A Multinomial logit transitions In Section 2.t ≡ πsq (zt−1 ) ≡ Pr[St = s|St−1 = q. we can focus on the part of the log likelihood function related to the parameters βsq . ∂βsq t=1 ς∈S πςq ∂β sq 58 .t log πsq. The transition part of the expectation of the likelihood function is given by X T XX ℓ(B) = ξsq. These probabilities are based on the complete data set of returns and predictive variables ΩT . s. Therefore.t . (1994) consider estimation when the transition probabilities are linked via a standard (binomial) logit transforma- tion. The probability of a transition from regime q to regime s at time t is linked to predicting variables zt−1 by a multinomial logit transformation ′ eβsq zt−1 πsq.t . In the expectation step the set of smoothed inference probabilities is determined. In the maximization step new parameters values are calculated that maximize the likelihood function. We have dropped the model-superscript m for notational convenience.t ≡ Pr[St = s|St−1 = q. zt−1 ] = P βςq ′ zt−1 . based on the EM-algorithm by Dempster et al. q ∈ S} is the set of all parameters βsq and ξsq. (25) t=1 s∈S q∈S where B = {βsq : s. q ∈ S. ΩT ] is a smoothed inference probability. and are calculated with the method of Kim (1994). We derive the first order conditions that apply to βsq by differentiating Eq. we extend the approach of Diebold et al. (24) ς∈S e with ∃s ∈ S : βsq = 0 to ensure identification. Diebold et al. A. We maintain the attractive feature of the EM-algorithm that the expectation of the complete-data log likelihood can be split in terms related to only a subset of the parameter space. (1994).

we calculate the marginal effect of the change in one variable zi . Numerical techniques can be used to find parameters βsq that solve this system. A.t−1 πsq. ∂βsq  −πςq πsq zt−1 if ς 6= s Combining these two expressions yields the first order condition X T (ξsq.t ) zt−1 = 0 ∀q. To solve this problem. For each departure state q the set of the first order conditions for the different s ∈ S comprise a system that determines the set Bq = {βsq : s ∈ S}. evaluated at specific values for all variables z̄. the coefficients on the explana- tory variables cannot be interpreted in a straightforward way. The marginal effect is given by the first derivative of (5) with respect to zi : . (24) we find   πsq (1 − πsq )zt−1 ∂πςq if ς = s = .t − ξq. s ∈ S (26) t=1 where ξq.2 Marginal Effects Because the multinomial logit transformation is non-linear.Based on Eq.t = Pr[St = q|ΩT ].

! ∂πsq (z) .

.

(27) ∂zi . X = πsq (z̄) βsqi − πςq (z̄)βςqi .

πsq (z̄)(1− πsq (z̄))βsqi . the above expression reduces to the familiar expression for marginal effects in logit models.org/fred2/ 59 . When only two regimes are available. except the unemployment rate and the 8 See https://research.z=z̄ ς∈S where βsqi denotes the coefficient on zi . Since the probabilities for the destination states should add up to one.stlouisfed. a marginal increase in one probability should be accompanied by decreases in the other probabilities. It is easy to verify that the sum of this expression over the destination states s is equal to zero. B Predictive variables The predictive variables that we use in this study are based on data from the Federal Reserve Bank of St. Louis available via FRED8.

10%.27%. The AR(1)-coefficient and ADF-test are based on this smaller subsample. Rapach et al. The yearly change in unemployment is close to zero with a standard deviation of 1. Because a unit root is not rejected.23%. seasonally adjusted).05. The resulting difference is on average zero and shows a standard deviation of 1. Series ID: WGS10YR) minus the 3-month T-Bill rate. (2005). The results are reported in Table B.14%. For each predictive variable we test whether it has a unit root. We reject a unit for the inflation rate at conventional confidence levels. we transform the series by taking a yearly change.1. together with the mean and standard deviation of the (transformed) series. we transform this series by taking the difference with respect to the yearly moving average as in Campbell (1991). so we do not transform this series. If so. We determine the credit spread as the difference between the yield on BAA-rated and AAA-rated corporate bonds (as calculated by Moody’s. We use the monthly series for the 10-year government bond and 3-month T-Bill rate to construct the observations from 1955 to 1962.06% with a standard deviation of 5. We obtain the unemployment rate from the Bureau of Labor Stastics (seasonally adjusted. They indicate stationarity. We construct the inflation rate as the monthly relative change in the seasonally adjusted consumer price index (All Urban Consumers All Items.] The T-Bill rate corresponds with a maturity of three months and is again taken from FRED (Series ID: WTB3MS). we transform the data to create a stationary time series.78% per year with a standard deviation 1. Series INDPRO). We construct the term spread as the yield on a 10-year government bond (constant maturity.04%. and find an average inflation rate of 3.dividend-to-price ratio.42% with a volatility of 1. A weekly series for the 1-year yield becomes available from 1962 onwards.1 about here. As weekly data become available starting in 1962 again. The average term spread is 1. and has an average of 3. we follow the same 60 . Series ID: LNS14000000). [Table B. Because of the high AR(1)-coefficient and the p-value for the ADF-statistic close to 0. These three variables are constructed at the monthly frequency. and assume that the difference stays constant within a month. It does not show evidence of a unit root. Series ID: CPIAUCSL). We construct the growth rate of industrial production as the yearly relative change in industrial production (seasonally adjusted.

The price index for the S&P500 is spliced together from Schwert (1990) and FRED.34%. To calculate the D/P-ratio for time t (in weeks).htm for more information. As the resulting series shows evidence of a unit root. we divide the dividends over the last 12 months prior to the month corresponding with time t by the current price level. This difference is on average zero and has a volatility of 0. 61 .46% for the whole sample.yale. 9 See http://www. we take again the difference with respect to the moving yearly average.econ. and find an average spread of 0.9 The series consists of monthly observations of the moving total dividends over the past twelve months. The dividend-to-price ratio is constructed from several series. We reject the null-hypothesis of a unit root for the subsample with weekly data. We use the dividends se- ries for the S&P500 as in Shiller (2000). which are available on Robert Shiller’s homepage.procedure as for the term spread.edu/~shiller/data. as explained in Section 4.99% and a volatility of 0.

623 -3.46 62 D/P ratio weekly 0. The dividend series.997 -3.001 3.36 0. 2010.23 Credit spread weekly 0.20 < 0.031 yearly change 0.78 1. which is a twelve month moving total.152 change to yearly average -0.998 -2. is then divided by the price series. The last two columns show the mean and standard deviation of the (transformed) variables in %.05 0.998 -1.27 Unemployment monthly 0.001 1.06 5. 1957) and again from FRED (from that date onwards). we have taken the dividend series from Shiller (2000) and spliced the prices series for the S&P500 together from Schwert (1990) (up to January 4.28 < 0.42 1.003 0. the ADF test-statistic and the p-value for the hypothesis of the presence of a unit root.14 Tbill rate weekly 0.993 -3.02 0.991 -4. Dev. All data series except the D/P ratio are obtained from the Federal Reserve Bank of St. Table B. Prod. The monthly series run from December 1954 to May 2010.966 -5. yearly growth rate monthly 0.017 3. Louis via FRED.08 1. the weekly series from January 7.01 1.34 This table shows the set of predicting variables with its source and frequency.390 change to yearly average -0. ..1: Characteristics of Predicting Variables Frequency AR(1) ADF p-value Transformation Mean Std. To construct the D/P ratio. If this hypothesis is not rejected.10 Ind. For each variable we conduct an adjusted Dickey-Fuller test. Inflation monthly 0. 1955 until June 25. the next column shows the transformation that is applied to the variable.99 0.78 0. We report the first order autocorrelation coefficient.26 0.04 Term spread weekly 0.81 0.

Bull markets tend to be slightly more persistent than bear markets. A rise in the D/P-ratio increases the likelihood of a bull market in the next period in the LT.1 Constant versus Time Varying Transition Probabilities The results in Section 5 indicate that the difference between constant and time-varying transition probabilities in the regime switching models are minor.89. time-variation can be related to a few economic variables.2.C Additional Results on Identification C. PS and RS3L-models. This results comes as no surprise. which is more puzzling. We then use these labeled periods as input for the estimation of the Markovian logit model in (6). Only the strongly bearish regime of is somewhat less persistent.1 about here. Bull and bear markets are quite persistent with probabilities of around 0. It is similarly puzzling than an increase in the D/P-ratio strengthens the persistence of the strongly bearish regime in the RS3L-model. The D/P ratio is selected for all models.] As an alternative to constant transition probabilities. In the RS2L-model the D/P-ratio decreases the likelihood of a bull market. [Table C. which makes them time-varying. In the rules-based approaches.3.1 reports the transition probabilities under the assumption that they are con- stant over time. We use a significance level for the likelihood ratio test of 10%. For the four-state regime switching model. Table C. though its probability of continuation for another week is still 0.or PS-algorithm to label periods as bullish or bearish. the lo- gistic transformation is extended to the multinomial logistic transformation in (5). we first use the LT. 63 . According to Table C. The two- state regime switching model uses the same logistic transformation to link the transition probabilities to predicting variables. we link the probabilities to pre- dicting variables.95 or higher that the current state prevails for another week. since an increase in the D/P ratio can point at higher future expected returns. We determine the variables to include for specific departure-destination combinations by the specific-to-general procedure proposed in Section 4. Here we investigate in more detail how different the results are.

the probability of a switch from a bear to a bull market can still double. the marginal effect of variable zi with coefficient βi is given by π(1 − π)βi . in the PS-model.04. We report these probabilities for the different regimes and models in the last row of Table C. As a reference point we use the average forecast probability PT Pr[St+1 = s|St = q. the marginal effect of a one-standard deviation rise in the D/P-ratio increases the probability a bear-bull switch from 0. The marginal effects are calculated from the reference probability π and the coefficient β. In this expression.02 to 0. Pr[St = q|Ωt ].2 about here. When logit transformations. They confirm the strong persistence reported in Table C.2. marginal effects are small. Overall. (28) t=1 Pr[St = q|Ωt ] where Ωt denotes the information set (predicting and dependent variables) up to time t.1.01– 0. To determine by how much the transition probabilities change when the predicting variables change. while in the RS2L-model they increase the probability that a bull market continues. In the rules based approaches. For instance. with changes of around 0. zt−1 ] of a switch from state q to state s is weighted by the likelihood of an occurrence of state q at time t.02. we calculate the marginal effect that a one-standard deviation change in a predicting variable has on a reference probability π. [Table C. In the PS-approach. but has no effect when the market is bearish. For the multinomial logit transformation we derive the marginal effects in Appendix A. However. a higher term spread or a lower credit spread increases the probability of a switch from a bear to a bull market. In the regime-switching approaches the weights are the so-called inference probabilities.] 64 . each forecast probability Pr[St+1 = s|St = q. Other variables show up less consistently. An increase in the unemployment rate during a bull market leads to a higher probability of continuation in the PS and RS3L-model. the weights are either zero or one. zt−1 ] Pr[St = q|Ωt ] π̄sq = t=1 PT . PS and RS3L-models. A higher T-bill rate decreases the probability of a bull market in the LT-model. A rise in the inflation rate negatively affects the persistence of bull markets in the LT.

However. We conclude that the evidence favoring time- varying transition probabilities is at best marginal.2 Model choice To judge the quality of the different models. the evidence so far supports a model with three regimes. which puts a heavier penalty on additional parameters. the Akaike Information Criterion favors the models with time-variation in the transition probabilities over the models where they are constant. Both information criteria prefer a model with three regimes over one with only two. The evidence for time-variation is limited.3 about here. these improvements are all significant. it is not surprising that other economic variables fail to predict the stock market sentiment well. we can determine the added value of predictive variables in the transition probabilities. When transition probabilities are constant. we conclude that all methods identify persistent bull and bear markets. As our interest is not in selecting the best statistical model we do not conduct this test here. the statistic does not follow a standard χ2 distribution and simulations can be used. due to presence of nuisance parameters under the null-hypothesis of two regimes. However.50). The D/P- ratio which is closely related to expected returns in the stock market is most consistently selected. we calculate and compare log likelihood values in Table C. However.3.2. 65 . Given the size of the LR-statistic and the values of the information criteria. Also. the model with two regimes is nested in the model with three regimes. the improvements of the likelihood are not enough to improve the Bayesian Information Criteration. if the sentiment of the stock market is a good predictor of other economic processes. an additional regime leads to large improvements in the likelihood values. By construction. Combining Tables C. [Table C. For both cases (constant and time-varying transition probabilities). C. For all models. and we could conduct a likelihood ratio test (the LR-statistic equals 105.] For the regime-switching models we can also evaluate the added value of an extra regime.1 and C.

570 (mild) bear 0.348 0.001 bear 0.984 0.010 0.27 0.615 0.948 0.001 bear bull 0. As a second step we estimate the probabilities.396 strong bear 0.652 0.052 0.014 crash < 0.009 crash bull < 0. Table C. In the approaches of LT and PS.992 0. We assume that the probabilities are constant over time.894 (b) Unconditional Regime Probabilities LT PS RS2C RS3C bull 0.008 0.972 crash 0.019 0. 66 . we first apply their algorithms to identify the sequences of bull and bear markets. For the regime switching models the probabilities result directly from the estimation.985 0.990 0.981 0.986 bear 0. The unconditional probabilities π̄ satisfy π̄ m P m = π̄ m .019 bear 0.1: Constant Transition Probabilities (a) Probability Estimates from to LT PS RS2C RS3C bull bull 0.730 0. The regime switching model can either have 2 regimes (RS2C) or 3 regimes (RS3C).385 0.106 crash 0.034 This table shows the transition probabilities between the different regimes under the different approaches and the resulting unconditional probabilities.016 0.015 0.

97 0. the marginal effect is given by (5).01 0.35 −0. For the two-state approaches.010] [0.207] π̄sq 0.003] [−0.96 < 0.020] [−0. In the two-state regime switching model. In the second step we estimate a Markovian logit model as in (1).18 inflation −0.27 0.40 −0.01 1.015] [0.33 88.95 − − − − [0.16 −5.02 0. RS2L.031] [0. (multinomial) logit models model LT PS RS2L RS3L from bull bear bull bear bull bear bull mild bear strong bear to bull bull bull bull bull bull bull mild bear bull mild bear bull mild bear constant 5. The procedure stops when the remaining variable-transition combinations fail to produce an increase in the likelihood function that is significant on the 10%-level. we first apply the algorithms to identify bullish and bearish periods. For the three- state regime switching model. In the approaches of LT and PS.009] credit spread − − − −0.14 [0. The marginal effects in brackets are calculated for the average forecast probability π̄sq reported in the last row of the table.77 − − − − − − − − − − − [−0.32 − − − − − − − [0.013] prod.2: Time-varying transition probabilities.73 − − − − −0.68 − − − − − − − 67 [−0.99 0.32 5.72 −63. RS3L.006] term spread − − − 0. The average forecast probability is calculated as in (28).04 − − − − − [−0.014] [−0. Table C. The variable-transition combinations that subsequently produce the biggest increase in the likelihood function are included in the models.62 −2. growth − − − − − − − − − − − − unempl. − − 0.007] [0.013] [0.85 − − − −1.008] [−0.05 1.006] [−0.35 3.02 0. the marginal effect of variable i is calculated as π̄sq (1 − π̄sq )βqi .58 83. For the three-state approaches. the logistic transformation in (6) is used to link the predicting variables to the transition probabilities. The predicting variables have been standardized by subtracting their full-sample mean and dividing by their full-sample standard deviation.02 0.018] D/P ratio 0.25 − −0.011] [0.23 −5.99 0.08 0.01 0.76 1. the multinomial logistic transformation in (5) is used.43 1.006] [0.99 0. In that case the coefficients for a switch to the strong bear regime have been fixed at zero.70 0. where the coefficients depend on the departure state. .014] [−0.012] [−0.34 − −0.29 3.24 This table shows the estimated coefficients and marginal effects of the predicting variables.012] t-bill rate −0. For the RS3L-model we allow a maximum of four combinations to be included.14 − − − − −1. when they are linked to the transition probabilities by (multinomial) logit models.

01) 13.12 -192.104 BIC 0.155 4.110 0.28 This table shows the log likelihood values of the different models.2). The last row reports 1% critical values of the corresponding χ2 distribution.28 13.09 df 4 6 4 4 Pr(0.136 4.96 17. In the row labeled “LR” we report the likelihood ratio statistic for time-varying vs.133 4.123 0.96 AIC 0. Based on the log likelihood values we calculate the Akaike and Bayesian Information Criterion (AIC and BIC).60 47. 68 .133 time-varying # parameters 6 8 11 18 log L -153. For the regime switching models with two or three states.150 4.28 16.139 4. we report the log likelihood values of the Markovian logit models as in (1). The transition probabilities can be constant (corresponding with Table C.1) or time-varying (corresponding with Table C.75 -5927.87 -5916. constant transition probabilities.3: Log likelihood values and information criteria of different model specifi- cations LT PS RS2 RS3 constant # parameters 2 2 7 14 log L -170.78 20.119 0.122 4. For the rules-based approaches LT and PS.61 -5979.139 4.123 0.82 -168.01 AIC 0.137 LR 32.81 13.134 4.63 -5970.Table C. which has a χ2 distribution with degrees of freedom listed in the row below. we report the likelihood of the complete model.100 BIC 0.

and a higher volatility. The probability of remaining in a bull state never falls below 0. The methods with two states produce transition prob- abilities that are also quite stable over time. but the probability of continuation almost always exceeds 0. Generally. The characteristics of the bearish regime match quite well with those of the bearish regime in the RS2-models. one bullish and two bearish regimes show up. In the regime-switching approaches. As we concluded in the full-sample analysis. the same applies to the bear regime. Figure D. The evolution of the transition probabilities when assumed constant within an estima- tion window are in Figure D. we see that the means and volatilities are stable. while the other is more mild. It stresses the exceptional behavior of the US stock market during the credit crisis.1 shows the evolution of the means and volatilities of the different regimes.2. In the rules-based approaches.95. For the rules-based ap- proaches. When data until May 2008 is used. [Figure D.] The evolution of the means and volatilities of the regimes in the RS3-models in Fig- ures D. the difference between the mean for the bull and for the bear regime is more extreme for the rules-based approaches. with a mean slightly above zero.1f indicates the influence of the credit crisis.] 69 . Persistence is high for both regimes. After 2008.1 about here.D Additional Results on Predictions In the experiment to compare the predictions of the various methods. In the RS2C model.1e and D. two bullish and one bearish regime are identified.90. a bear market seems slightly less persistent. and estimate means and volatilities based on that. we estimate means and volatilities directly. Here we show how the different param- eters evolve. One regime has strongly bullish characteristics with a high mean and a low volatility. while the difference between the volatilities for these two regimes is more extreme for the RS2-models. Besides aiding our understanding of the predictions. while the other has much stronger bearish features. we reestimate the parameters of the different models every 52 weeks.2 about here. it also shows how robust the parameters and characteristics are when more information becomes available. One bearish regime has mild characteristics. [Figure D. we first do the identification.

] 70 .] [Figure D. The D/P-ratio is consistently present for all switches in both the LT and PS-methods. the selection and the parameters are quite stable. these are sometimes selected. In the rules-based method.3 (continued) about here. The probabilities that a specific regime continues are high (around 0. The parameter dynamics of the models for time-varying transition probabilities in Fig- ure D. In the LT-approach. The T-bill rate is selected for predictions from the bullish regime in both models. [Figure D. The transition probabilities in the RS3C-model vary a bit more than in the two-state models. the corresponding coefficient is stable. In the PS-method.90 or more) and quite stable. the inflation rate.04 to 0. but also whether the variable selection is stable. given that a variable is selected. the probability of a switch from a (strongly) bearish to a mildly bullish/bearish regime ranges from 0.3 about here.11. For example. the unemployment rate and the term spread also show up consistently. In the RS-models the selection shows a more haphazard pattern. in particular the probabilities for a switch to another regime.3 indicate whether the parameters are stable. However.

Figure D. and constant or time-varying transition probabilities.5 0 -0. In the approaches of LT and PS. The means and volatilities of the regimes follow directly from the estimation.6 6 0.5 1.5 1.2 0.4 0. 1955 – June 24.4 5 0. The first window comprises the period January 7.6 2. we first apply their algorithms to identify the sequences of bull and bear markets for each estimation window. 71 .1 0.5 0.8 3 1.3 -0.3 -0.6 0 -0.4 2 -0.7 3 4 5 6 7 88 89 90 91 92 93 94 95 96 97 98 99 00 01 02 03 04 05 06 07 08 09 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 8 8 8 8 8 8 8 8 8 8 8 8 9 9 9 9 9 9 9 9 9 9 0 0 0 0 0 0 0 0 0 0 n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju (a) mean of bull regimes in two-state models (b) mean of bear regimes in two-state models LT PS RS2C RS2L LT PS RS2C RS2L 2 3.1 -0.2 -0. and is continuously expanded with 52 weeks until we reach the end of the sample (July 2.2 4 0 -0.8 -1 0 3 4 5 6 7 88 89 90 91 92 93 94 95 96 97 98 99 00 01 02 03 04 05 06 07 08 09 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 8 8 8 8 8 8 8 8 8 8 8 8 9 9 9 9 9 9 9 9 9 9 0 0 0 0 0 0 0 0 0 0 n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju (e) mean of regimes in RS3-models (f) volatility of regimes in RS3-models This figure shows the evolution of the means and volatilities when estimated with an expanding window (end date on the x-axis). As a second step we calculate means and volatilities per regime.1: Evolution of means and volatilities per regime LT PS RS2C RS2L LT PS RS2C RS2L 0.2 1. The regime switching models can either have two regimes or three regimes.6 1 -0. 1983 (1485 observations). 2010).5 1 1 3 4 5 6 7 88 89 90 91 92 93 94 95 96 97 98 99 00 01 02 03 04 05 06 07 08 09 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 8 8 8 8 8 8 8 8 8 8 8 8 9 9 9 9 9 9 9 9 9 9 0 0 0 0 0 0 0 0 0 0 n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- n- ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju ju (c) volatility of bull regimes in two-state models (d) volatility of bear regimes in two-state models C bull L bull C mild L mild C (strong) bear L (strong) bear C bull L bull C mild L mild C (strong) bear L (strong) bear 0.4 2 1.2 3 -0.4 -0.

93 0.91 0.84 0.1 0.94 0.99 0. The regime switching models can either have two regimes (RS2C) or three regimes (RS3C).94 0. we first apply their algorithms to identify the sequences of bull and bear markets for each estimation window.97 0.08 0. and is continuously expanded with 52 weeks until we reach the end of the sample (July 2.02 0.82 0 jun-83 jun-84 jun-85 jun-86 jun-87 jun-88 jun-89 jun-90 jun-91 jun-92 jun-93 jun-94 jun-95 jun-96 jun-97 jun-98 jun-99 jun-00 jun-01 jun-02 jun-03 jun-04 jun-05 jun-06 jun-07 jun-08 jun-09 (b) RS3C This figure shows the evolution of the transition probabilities when estimated with an expanding window (end date on the x-axis). For the methods with two states. 72 .96 0. we plot the probabilities of a bull-bull and a bear-bear switch in Panel (a). Figure D.9 0. For the RS3C-model in Panel (b) we indicate the transition in the legend above the subfigure. For the regime switching models the probabilities result directly from the estimation. 2010).06 0.98 0.86 0.001.92 0. We do not show transition probabilities that never exceed 0.96 0. We assume that the probabilities are constant within each estimation window.12 0.98 0. The first window comprises the period January 7.9 jun-83 jun-84 jun-85 jun-86 jun-87 jun-88 jun-89 jun-90 jun-91 jun-92 jun-93 jun-94 jun-95 jun-96 jun-97 jun-98 jun-99 jun-00 jun-01 jun-02 jun-03 jun-04 jun-05 jun-06 jun-07 jun-08 jun-09 (a) 2-state models bull  bull mild  mild (strong) bear  (strong) bear bull  mild mild  bull mild  (strong) bear (strong) bear  bull (strong) bear  mild 1 0. 1955 – June 24. As a second step we estimate the probabilities.04 0.88 0.95 0.2: Evolution of constant transition probabilities LTC bull LTC bear PS bull PS bear RS2C bull RS2C bear 1 0. 1983 (1485 observations). Dashed lines correspond with the secondary y-axis. In the approaches of LT and PS.92 0.

5 1 1.5 1 1.5 -1 -0.5 jun-83 jun-84 jun-84 jun-84 jun-85 jun-85 jun-85 jun-86 jun-86 jun-86 jun-87 jun-87 jun-87 jun-88 jun-88 jun-88 jun-89 jun-89 (f) RS2L.5 0 0. growth inflation D/P ratio credit spread term spread t-bill rate unempl. -2 -1.5 0 0.5 0 0. growth inflation D/P ratio credit spread term spread t-bill rate unempl.5 1 1.5 -2 -1.3: Evolution of parameters in logit models jun-95 jun-95 jun-95 jun-96 jun-96 jun-96 jun-97 jun-97 jun-97 jun-98 jun-98 jun-98 jun-99 jun-99 jun-99 jun-00 jun-00 jun-00 jun-01 jun-01 jun-01 jun-02 jun-02 jun-02 jun-03 jun-03 jun-03 jun-04 jun-04 jun-04 jun-05 jun-05 jun-05 jun-06 jun-06 jun-06 jun-07 jun-07 jun-07 jun-08 jun-08 jun-08 jun-09 jun-09 jun-09 D/P ratio credit spread term spread t-bill rate unempl. growth inflation D/P ratio credit spread term spread t-bill rate unempl. from bear to bull jun-90 jun-91 jun-91 jun-91 jun-92 jun-92 jun-92 jun-93 jun-93 jun-93 jun-94 jun-94 jun-94 jun-95 jun-95 jun-95 jun-96 jun-96 jun-96 jun-97 jun-97 jun-97 jun-98 jun-98 jun-98 jun-99 jun-99 jun-99 jun-00 jun-00 jun-00 jun-01 jun-01 jun-01 jun-02 jun-02 jun-02 jun-03 jun-03 jun-03 jun-04 jun-04 jun-04 jun-05 jun-05 jun-05 jun-06 jun-06 jun-06 jun-07 jun-07 jun-07 jun-08 jun-08 jun-08 jun-09 jun-09 jun-09 D/P ratio credit spread term spread t-bill rate unempl.5 1 1. from bull to bull (c) PS. growth inflation D/P ratio credit spread term spread t-bill rate unempl. prod.5 -1 -0.5 -2 -1. prod. growth inflation . prod.5 1 1.5 -1 -0.5 0 0.5 jun-83 jun-83 jun-83 jun-84 jun-84 jun-84 jun-85 jun-85 jun-85 jun-86 jun-86 jun-86 jun-87 jun-87 jun-87 jun-88 jun-88 jun-88 jun-89 jun-89 jun-89 (e) RS2L. from bull to bull jun-90 jun-90 jun-90 (a) LT. prod.5 -2 -1. from bear to bull jun-90 jun-90 (b) LT. prod. growth inflation 73 -2 -1.5 jun-83 jun-83 -2 -1.5 0 0. from bear to bull jun-89 (d) PS. prod.5 -1 -0.5 -1 -0.5 1 1. Figure note on next page. from bull to bull jun-91 jun-91 jun-91 jun-92 jun-92 jun-92 jun-93 jun-93 jun-93 jun-94 jun-94 jun-94 Figure D.5 0 0.5 -1 -0.

The predicting variables have been standardized by subtracting their full-sample mean and dividing by their full-sample standard deviation. we first apply the algorithms to identify bullish and bearish periods in the subperiod under consideration. 1983 (1485 observations). from bull to mild bear 3 3 2 2 1 1 inflation inflation prod. from strong bear to mild bear This figure plots the evolution of the coefficients in the (multinomial) logit transitions for the predicting variables in Table B. In the second step we estimate a Markovian logit model as in (1). unempl. where the coefficients depend on the departure state. For identification. Figure D. unempl. growth prod. growth unempl. growth unempl. when estimated with an expanding window (end date on the x-axis). growth unempl. we only plot the variables that have been selected at least once. In the RS3L. 2010).3: Evolution of parameters in logit models – continued 3 3 2 2 1 1 inflation inflation prod. In the approaches of LT and PS. t-bill rate t-bill rate 0 0 term spread term spread credit spread credit spread D/P ratio D/P ratio -1 -1 -2 -2 -3 -3 jun-83 jun-84 jun-85 jun-86 jun-87 jun-88 jun-89 jun-90 jun-91 jun-92 jun-93 jun-94 jun-95 jun-96 jun-97 jun-98 jun-99 jun-00 jun-01 jun-02 jun-03 jun-04 jun-05 jun-06 jun-07 jun-08 jun-09 jun-83 jun-84 jun-85 jun-86 jun-87 jun-88 jun-89 jun-90 jun-91 jun-92 jun-93 jun-94 jun-95 jun-96 jun-97 jun-98 jun-99 jun-00 jun-01 jun-02 jun-03 jun-04 jun-05 jun-06 jun-07 jun-08 jun-09 (g) RS3L. all coefficients for a switch to the (strong) bear regime have been fixed at zero. t-bill rate 0 term spread credit spread D/P ratio -1 -2 -3 jun-83 jun-84 jun-85 jun-86 jun-87 jun-88 jun-89 jun-90 jun-91 jun-92 jun-93 jun-94 jun-95 jun-96 jun-97 jun-98 jun-99 jun-00 jun-01 jun-02 jun-03 jun-04 jun-05 jun-06 jun-07 jun-08 jun-09 (k) RS3L. from bull to bull (h) RS3L. The variable-transition combinations that subsequently produce the biggest increase in the likelihood function are included in the models. . 1955 – June 24.1. In the RS2L-model the logistic transformation in (6) is used to link the predicting variables to the transition probabilities. In each subfigure. from mild bear to bull (j) RS3L. from mild bear to mild bear 3 2 1 inflation prod. t-bill rate t-bill rate 0 0 term spread term spread credit spread credit spread D/P ratio D/P ratio -1 -1 -2 -2 -3 -3 jun-83 jun-84 jun-85 jun-86 jun-87 jun-88 jun-89 jun-90 jun-91 jun-92 jun-93 jun-94 jun-95 jun-96 jun-97 jun-98 jun-99 jun-00 jun-01 jun-02 jun-03 jun-04 jun-05 jun-06 jun-07 jun-08 jun-09 jun-83 jun-84 jun-85 jun-86 jun-87 jun-88 jun-89 jun-90 jun-91 jun-92 jun-93 jun-94 jun-95 jun-96 jun-97 jun-98 jun-99 jun-00 jun-01 jun-02 jun-03 jun-04 jun-05 jun-06 jun-07 jun-08 jun-09 (i) RS3L. The procedure stops when the remaining 74 variable-transition combinations fail to produce an increase in the likelihood function that is significant on the 10%-level. the multinomial logistic transformation in (5) is used. The first window comprises the period January 7. and is continuously expanded with 52 weeks until we reach the end of the sample (July 2. growth prod.

we follow LT and also consider these combinations of lower thresholds. Comparing these difference with those in Table 3 shows that they are slightly smaller than the difference between the standard LT and the PS identifications. 0. though the result is less pronounced than in LT.2 about here.E Result of Robustness Checks E.10). [Table E.011 and 0. So from a statistical point of view. λ2 ) are (0. the number of cycles increases from 16 in the original (0. that as a consequence will last briefer. 0.20. Other combinations they consider for (λ1 . 0.15) setting to 24 for the lowest thresholds (0. Their average difference varies between 0. Since we conclude that a quick identification of the current state is important when making predictions. This is due the fact that the longest bull market is unaffected by the choice of thresholds.15) and (0.10.10).1 Robustness of the LT-method Lunde and Timmermann (2004) consider four combinations for the values of the thresholds λ1 and λ2 to identify switches between bull and bear markets. 0.2a shows that the integrated absolute difference between these series is quite small. (0. Lower thresholds lead to a more rapid alternation of bull and bear markets.059. Average and median durations decrease when we work with lower threholds.1 about here. Table E.] We apply the techniques of Section 3 to determine how different the identifications and predictions are that results from the LT-method with different thresholds.15. the type of method matters more than the specific thresholds. A lower value of 15% for λ2 subsequently accounts for the positive drift of the stock market. Lowering both thresholds has a much stronger effect than lowering only one of the thresholds.10). [Table E. 0.] 75 .15. while the upper bound of the 90% confidence intervals is around 0.15. When we consider identification based on the full sample. They argue that a value of 20% for λ1 is conventionally used.20. and much smaller than the differences between the standard LT-method and the RS-methods.

The fees to switch from (0. with maxima around 0. and it is not obvious that working with lower thresholds leads to better predictions or allocations ex ante.3% per year.2b and E. at the cost of a slight increase in the volatility from 30.] The IAD-measures in Table E.15) are all positive. A comparison with Table 6 shows that the predictions by LT- method with different thresholds still have more in common than the standard LT-method with the other methods. An investor would be willing to pay large fees up to 9. A faster identification of bull and bear markets leads to a considerably higher mean. Logi- cally.1% to 66.4 about here.67% to switch to lower thresholds.3% to 35. as confidence intervals are substantially bounded from zero. 0. We conclude from these results that an exceedance of a lower threshold suffices to mark a definite switch between bull and bear markets. The lowest thresholds produce the largest hit rate and Kuipers score and the lowest IAD with the final identification.4 indicate that the differences between predictions based on the different threshold combinations are larger than those between identification.15) to (0.20. In line with earlier results.2c shows larger differences. including the PS method. The economic comparison in Tables E. the fees in Table 4b may un- derestimate the preference of the standard LT-method over the regime switching methods. IADs are larger when both threshold values are different. However. the use of predictive variables only leads to marginal differences in the predictions. these fees correspond with identification ex post.3 about here.10) are significant. We turn to the predictions made with different thresholds in Table E. Confidence intervals indicate that the differences are significant.20. both the Sharpe ratio and the realized utility increase. Since fees to switch from the standard (0.2%. [Table E. These numbers put the LT-method with lowest thresholds at par with the two state regime switching models in Table 5.16. rising from 48. 0. meaning that 16% of the predictions are different. Lowering thresholds leads to a higher predictive accuracy.15. As a consequence.] 76 . 0. Only the joint lowering of thresholds commands a significant fee. In particular bear markets are better predicted with lower thresholds. [Table E.3.

Lower thresholds lead to better identification.10) com- mands a fee of 4–6% per year. but the confidence intervals are quite wide and stretch beyond zero. we conclude that lowering the thresholds improves identification and predic- tions. 0. [Table E. the economic comparison shows that the false alarms actually become costlier. we compare the performance of investment strategies that are based on the LT- methods with different thresholds in Table E.5a is still lower that the utilities for the regime switching models in Table 7a. and the overall balance in terms of economic performance is negative.20. the results are more mixed. The magnitude of the fees is small compared to the fees reported in Table 7. Since utility is still considerably lower than the level produced by the regime-switching models.5 about here. the threshold combination (0. Part of the increases come from a larger position in futures contracts. lower thresholds improve the performance of the LT-methods to a level comparable to the two- state regime switching models. Overall. while a lowering of the λ1 threshold actually leads lower utility because of the increase in volatility. Judged by Sharpe ratio and realized utility.5b indicate that the economic improvement of a strategy by one threshold combination compared to another combination is still relatively small and mostly insignificant. which in the end makes a slower.5. Nonetheless. these models lead to predictions that are economically more valuable.10) combination with constant transition probabilities. We already concluded that the rules-based methods are better suited for identification. 0. which rise from around 8% to 12%. 77 .20. For pre- dictions. more cautious strategy more attractive.] The fees in Table E. Changing from the standard (0. The highest utility in Table E. lowering the threshold λ2 makes sense as a switch to this setting commands a positive fee and leads to higher utility. The results for lower thresholds indicate an even larger difference.15) works best. The only significant fees correspond with a switch from the (0. both in a statistical and an economic performance. Finally. Lower thresholds lead to higher means. Lowering the thresholds leads to better identification but also to more risk-taking.15) combination to the (0. However.20. If hits and false alarms are equally important. but also to higher volatilities. 0.15. 0.

E.2 Robustness of the PS-method
Pagan and Sossounov (2003) base their algorithm on the algorithm for business cycle
identification of Bry and Boschan (1971), and adjust some of the original settings based
on early literature on bull and bear markets, the so-called Dow theory after Charles Dow.
They do not investigate how robust their findings are to changes in these settings. We
conduct a small robustness investigation, where we change one parameter at a time. In
comparison with the LT-method, the PS-method identifies a few more cycles, so we consider
only a relaxation of the constraint on cycle length, which we lower to 52 weeks instead of
70. For the constraint on the length of a phase, we consider a lower value of 12 weeks and a
higher value of 20 weeks. A higher value may lead to less false alarms. The bound on price
change that must be crossed to overrule the phase constraint is currently at 20%. As in the
robustness checks for the LT-method, we consider a value of 15%. Censoring (currently
13 weeks) will not influence the identification much, but may have considerable impact on
predictions. We consider a lower value of 7 weeks and a higher value of 26 weeks, which
corresponds with the setting of PS.
We find that the identification is quite robust to these changes. In the basic setting,
we identify 18 bull and 18 bear markets (see Table 1). The results do not change at all,
when we change the phase constraint or the price change bound. Censoring 7 weeks at
the beginning and end leads to one extra switch to a bear market at the end of of the
sample. Censoring 26 weeks does not lead to different results. Lowering the cycle to 52
weeks leads to 19 bull and 19 bear markets. Because results change at most slightly, the
further analyses based on identification will also change only slightly. Our conclusion that
the PS-method works well for identification remains unaffected.
The predictions and allocations based on the PS-method change more when the pa-
rameters are changed. A prediction for week t + 1 based on the PS-method uses only
information up to time t. The identification over the subsample until week t may differ
from the full-sample identification and may be more sensitive to parameter choices. This
larger sensitivity then carries over to the predictions.
Table E.6 show the sensitivity of predictive accuracy for the restrictions. The hit rate
for bull markets changes a bit (at most 10%), while the hit rate for bear markets shows

78

more variation. However, an improvement in correctly predicting bull markets is offset by
a deterioration for bear markets. The overall hit rates remain steadily between 70–75%.
This result contrasts with the results for the LT-method in Table E.3, which shows more
variation.

[Table E.6 about here.]

Table E.7 shows a further analysis of the impact of parameter changes. Censoring
leads to predictions that are quite different from the predictions in the basic setting, with
IADs of around 0.l7 when 7 weeks are censored and 0.22 when 26 weeks are censored.
Less censoring leads to a more aggressive allocation with weights larger than two, while
more censoring leads to a less aggressive allocation with weights close to one. This pattern
is caused by predictions being closer to the steady state distribution of bull and bear
markets, when more observations are censored. Effectively, more censoring necessitates
a prediction for more steps ahead. More censoring and less aggressive investments lead
to a better performance: the mean increases while the volatility decreases. Less censoring
shows the opposite effect. It means that the predictions are too extreme, and that shrinking
them towards their long-term average makes sense. The effect of more censoring is quite
impressive with a Sharpe ratio that increases to 0.37 when predictive variables are used.
This ratio is similar in magnitude to the best LT-method in the previous subsection.
Moreover, utility increases compared to the basis strategies. The resulting utility is only
slightly lower than the utility the results from using two-state regime switching models.
Because more censoring leads to higher utility, an investor is willing to pay significant fees
of around 14% per year to switch from the basic PS-method.

[Table E.7 about here.]

All other parameter choices do not matter much. Changing the phase constraint does
not give different results. When we put the constraint at 12 or 20 weeks, the predictions
are exactly the same as in the basic setting. Therefore we have not include the phase con-
straint in a further analysis. A relaxation of the constraints on the duration of a cycle or
the price change to trigger a switch do not lead to large statistical or economic differences

79

in Table E.7. The characteristics of the investment strategies are largely similar. Conse-
quently, fees for switching differ only a little bit from zero. The confidence intervals for
the IADs and the fees encompass zero, indicating that the differences are not significant.

80

Table E.1: Number and Duration of Market Regimes for Different Thresholds in the
LT-method
λ1 0.20 0.20 0.15 0.15
λ2 0.15 0.10 0.15 0.10
bull number 16 18 18 24
avg. duration 119 98 108 77
med. duration 90 59 84 50
std. dev. duration 97 97 97 90
max. duration 405 405 405 405
min. duration 15 15 6 5
bear number 16 18 18 24
avg. duration 62 63 53 44
med. duration 60 55 60 39
std. dev. duration 44 49 29 30
max. duration 187 187 91 91
min. duration 7 7 7 4
This table shows the number of spells of the different market regimes, their average and median duration,
the standard deviation of the duration, and its maximum and minimum for different thresholds λ1 and λ2.

81

92.15) 2. -5. 0.15.Table E. 0.7 35.15) 0.45. 0. 0.59 1.259 0. The statistics are calculated as in Tables 3 and 4.00] [-16.10) 1.15.034. 4.57.61.20.72 7. λ2 ) (0.15) 0 -1.087] (0.20.10) (0.000.10) (0.20.15) (0.8 52.41.10) 9.15.011 0.112] (0.66.15) (0. 12.37.17 0 [5.83.264 0.54 -7.126] [0.047 0. 0.059 0.98] [-12.089] (b) Performance Measures (λ1 . 0. 82 . 0.102] [0.89] (0.15. 0.59 [-7.15) 0. λ2 ) (0.1 51.017. 0.4 31.00.54 0 -7.15.014.09] [3. 6.36] (0.031.94 0 -0.73.20. 0.2 Sharpe ratio 1.47] [-13.10) 0.62] [-4.83] This table shows the statistical and economic comparison of the identification that results from the LT- method when different thresholds λ1 and λ2 are used.10) Mean 48.10) (0.15. 0. 0.67 1.033 [0.20. 0. -3. 0.66 [0.241 0. 0.15. 0. 0.10) (0.80.20.15) (0.12 [0. 0.8 66.67 7.045 [0.88 Utility 0. 0.96.15.2: Statistical and Economic Comparison of the Identification by the LT- method for Different Thresholds (a) Integrated Absolute Distances (λ1 . 0. -3.48 -9.038 [0.20.10) (0. 4.059] [0. 0.15.65 1.76] [3.77] [-4. 7.331 (c) Fees (λ1 . λ2 ) (0. 0.15) (0. -0.15) (0. 13.20.3 31. 0. 0.49 0.017.41] [-6.94 -2. 0.20.3 Volatility 30.78] (0. 17.

0% 82.4% 45.0% 73.6% Total correct 1002 983 1079 1068 1120 1114 1166 1173 Total false 408 427 331 342 290 296 244 237 % Correct 71.7% 3.7% 79.15) (0.0% 76.160 0.6% 73.2% 48.1% Improvement -4.7% 76.6% 63.10) Transition C L C L C L C L Bull correct 852 851 816 819 983 977 934 941 Bull wrong 214 215 214 211 89 95 111 104 % Bull correct 79.7% 52. A C in the row transition means that transition probabilities are constant.15. and an L means they are modelled with the Markovian logit-model.5% 18.1% Kuipers Score 23.7% 91.2% Default bull 75.0% 8.127 See Table 5 for explanation.20.9% 53.10) (0. λ2 ) (0.5% 63.0% Bear correct 150 132 263 249 137 137 232 232 Bear wrong 194 212 117 131 201 201 133 133 % Bear correct 43.6% 75. 0.153 0.6% IAD 0.5% 2.5% -5.171 0.9% 79.Table E. 0.5% 75.4% 79.194 0.5% 40.0% 76.15.1% 89.6% 9.9% 3. 83 .2% 65.2% 31. 0.15) (0.0% 74.1% 74.222 0.5% 91.4% 3.5% 40.4% 69.20. 0.8% 79.1% 69.6% 38.3: Predictive Accuracy based on the LT-method with Different Thresholds (λ1 .159 0.4% 90.132 0.2% 79.0% 32.7% 83.

0.217] (0. 0. 0. 0. 0.083.096.166] [0.15.127 0.052 0.162 [0.172] (0.20. 0.063 0.20. where the first two elements indicate the values for λ1 and λ2 and the third element indicates constant (C) or time-varying (L) transition probabilities.204] [0.165 0.104. 0. 0. 0.180] [0.207] [0.15.055.121.101 [0. 0. 0. 0.134] [0.10. C) (0.247] [0.15.15.20.161] [0. 0. 0. 0.15.108 [0.135] 84 (0.116] [0.094. L) (0.108.099] [0.057. C) 0. C) (0. C) (0. 0. 0. 0. 0. L) 0. 0.092.052. .110] [0.15. 0.15.107.213] [0. 0.122.105] [0.10. Table E.096.10.181] [0. C) 0.206] [0. L) (0. L (0.10.121 0.116 [0.077 0.15.150 [0.116 0.15.136] [0.15.093] This table reports the integrated absolute distance between the predictions of the LT-method constructed with different thresholds and with constant or time-varying transition probabilities.089.165] (0.20.15.104] [0. 0.113 0.073.10.139.067 0. 0.060 0.098 0. 0. 0.163 0.10.142.038.086.15. C) (0.080 0.161 0.082 0. 0. L) 0.20. 0. C) 0.113 0.055.041. 0. 0. 0.101.20.056.162] [0. 0. 0. 0.097.239] (0.067 0.035. 0. L) 0. L) 0. The settings are reported in the headings in parentheses.118 0. 0.4: Integrated Absolute Difference of Predictions based on the LT-method with Different Settings (0. See Table 6 for further explanation.181] [0.20. 0.119.15.155 0.192] (0.10.242] [0.128 0.109 [0. 0.106 0.119] [0.032. 0.185] [0. 0. 0.065.15. 0.054 [0.

33 14.26.66] [-13.27 15.90] (0.92 [-3.13. 1. 5.18 10. 0.05] [-11.67] (0.46.98 -5.96 2.93 1.79] [1. 5.71.8 33. 0.90 1.10.32 0. 14.87 14.39.15) (0.08.18. 8. 0.38.86.20.15.99.97.06] See Table 7 for explanation.77] [-13.53] [-13. 1.03 [-3.39 Utility -0.16 -10.15.41 0. 0. -1.87 6.52] (0.87 4.99 4.08] [-11. 0. 12.21.16] (0.06] [-0. -2.91 1.77 -6. L) -0. Abs.82.20.15. 6.82.136 -0.36 12. 3.15.86] [-5.70] [-5.10. 11. 0.11] [2.9 32.10) (0. C) -5.10. 6.47] [-9. 13.91 2. which can be constant (C) or time-varying (L).15. 13.15.76 0.76 3. 7. C) 0 0.20.195 (b) Fees (λ1 . 3. 7.20.36] [1. 0.00] [-6.50.66.83.83] [-16.72.73] [-11.89 9.10 5.89 18.66.92 5.87.43 0.84 10. -1.14] [0.81 0 -0.92 -3.12 -5.194 -0.73. 5.4 31. 0.17 0.76 0.20.95] [-12. 3.22.21] [-8.90 9.16 -2.72] [-3.87 0 6.7 36. 1. 2.25.27 0.184 -0. 4.15.94] [-5.92 -0.22 0. 5. 8.15] [-8.89.10) (0.11 1.86.95 -5.89 -4. 0. 0.30 12.71.15) (0.45.235 -0.15.78.03.04 1.15 5.87 [-2.92.00 -10.92 5.67] [-20.73] [-7.82 -3.17 11.89.76.03 0 [-10. 8.31 0.75] [-4. Table E.89. 0. 13.20.27 0.77 -0.96 -6.42.81. Weight 1.81] [1.39.60.96 5.80 -5.20.26 Mean 8.56. . 10.03 [-16.93.7 33.85] [-19.22.08 0 -4.10.90] [-0. 2.15) (0. 11. 8.15.96 0 6.10 0.26] [1. C) -1.01. L) 0. 2.61. 0. -0. 0.94 1.127 -0.176 -0.04 4.44] [-1.81 0 -4.20.72] [-18. 11. 3.17] [-5. The different settings in the LT-method are indicated by the values for λ1 and λ2 and the choice for transition probabilities.82 1.64] [-2.175 -0.10) Transition C L C L C L C L Av.28 15.76.00] [-3.35] [-1. λ2 ) (0.87.04 2. 0.15 2.58.87] [-2.19 10. -1.77.87] [-14. 16. 0.15.69 16.01 -2.13 13. 16.95 -0.97 -4.10 9.99.82] [-5.90.96. 9.38] [-8.3 37.8 Sharpe Ratio 0. L) 4.94 -1.73 0 -1.39] (0.97.60 -0. 18.68] [-3. C) 3.81] [-2.25.71 4. 1.16 [-7.89] [-6.98] [-11.8 32. λ2 ) (0. 6. L) -1.74] [4.85] [-6.96 10.99 -4.17 Volatility 32.15) (0. 20.10) Transition C L C L C L C L (0.15.93 85 [-5. -4.73 7. 0.91] [-2.20 -6.13.60. 0. 19.77 -0. 0. 2. 2.06. 3.20] [-8.15.07 Mean Bear 8. -1.5: Performance Measures and Fees when Predictions are based on the LT-method with Different Settings (a) Performance Measures (λ1 .10.55] [-2. 5.88.11 [-1. 2.04 14. 5.36] (0.25] [-1.36 Mean Bull 8.31 13.85 1. 11.95 5.50] (0.

1% 32.3% 31.177 0.4% 2. See Table 5 for further explanation.2% 68. 86 .3% 62. All other parameters are set to their basic value.4% 28. It requires cycles to be longer than 70 weeks.9% Default bull 68.194 0.8% 71.6: Predictive Accuracy of the PS-method with Different Settings Setting basic censoring: 7 censoring: 26 cycle: 52 change: 15% Transition C L C L C L C L C L Bull correct 768 779 706 714 760 792 768 874 770 781 Bull wrong 193 182 245 237 201 169 193 87 191 180 % Bull correct 79.239 0.0% 23.6% 3.3% Bear correct 281 261 295 281 235 207 261 145 281 261 Bear wrong 168 188 164 178 214 242 188 304 168 188 % Bear correct 62.1% 79.0% 70.2% 67. Phases should be longer than 16 weeks unless the price changes with more then 20% since the last peak or trough.5% 36.2% 68.5% 38.2% 68.2% 42.1% 64.4% 67.161 0.176 0.4% 5.3% 46.2% Improvement 6.2% 38. It censors 13 weeks at the beginning and end.5% 73.9% 81.6% 70.7% 39. Table E.1% 58.5% 39.196 This table show the predictive accuracy of the PS-method when one parameter is changed. The basis setting is used throughout the paper.7% Kuipers Score 42. The column headings indicate which value is used for a parameter.2% 68.210 0.9% 80.4% IAD 0.6% 70.4% 79.1% 6.144 0.6% 58.2% 68.2% 68.2% 52.1% 81.3% 61.1% 74.7% 4.3% 74.9% 90.6% 58.2% 75. Transition probabilities can be constant (indicated by a C) or time-varying (indicated by an L).1% 82.4% 68.5% 3.262 0.0% 72.9% 73.2% 5.1% 2.4% 73.197 0.8% 4.1% Total correct 1049 1040 1001 995 995 999 1029 1019 1051 1042 Total false 361 370 409 415 415 411 381 391 359 368 % Correct 74.

91.56 28.335 -13.89 2.175 [0.21.84 -5.001 [0.181 0 Basic. C 0. Table E. L 0.02 -0.32 -9.009 1.22. the yearly Sharpe ratio. C 0 1.01 -5. The column fee reports the maximum fee that an investor is willing to pay to switch from a basic setting to another setting in the PS-method.22 [1. We report the 5% and 95% percentiles between brackets for the IAD and the fee based on the 200 bootstraps of both the return and the explanatory variable series following Politis and Romano (1994).7 0. we calculate the mean and volatility (in % per year). C 0.000.0 0.67 0. 2.08 6.0 0.03] Cycle: 52.02 Cycle: 52.20 5.181 0.91 24.03] Change: 15%. We do not report confidence intervals for the different cycle restrictions.42 9. -1. 0.70 1. 0.06 -0. We report the average absolute weight of each strategy.158.13 -3.87 2.175 0. the setting for the transition probabilities (constant or time-varying) in the basic strategy corresponds with the setting in the alternative method.65] The column IAD gives the Integrated Absolute Distance between the predictions based on a specific setting and the basic setting for the PS-method.1 0.04 -5.87 3. When calculating the IAD or the fee. L 0.33 19.185] 2. Abs.56 28.25 87 Change: 15%.05 16.57 1.82 -4.71 1.1 0.028 1. -3.98 28.21 29. and the annualized realized utility as in Eq.37 -0.7: Statistical and Economic Comparison of Predictions based on the PS-method with Different Settings Setting IAD Av.33 7. Mean Bull (Mean Bear) is the annualized mean during the subperiods identified ex post as bull (bear) markets. L 0.08 -0.195 0.38 1.11 -0.65 -0.4 -0.7 0.9 -0. 0.186.91 30.31 [-22.222 [0. (13).188] 2.228] 1. Mean Mean Mean Vol.16. We express the fees in % per year.57] Censoring: 26. 0.91 30.023] 1.70 3. .82 [-0.26 5. 2.97 2. as changing cycle minima cannot be combined with a stationary bootstrap.000.001 [0.05 -0.48 0.212 [0. 0.169 [0.203 0 Censoring: 7.91] Censoring: 7.181.07 -7.74 32.0 0.243] 1. 19.73 [6.23.324 -14.10 1.029] 1. L 0.62 [-0.27 -0. C 0.19 [-22.061 14.16 -0.81 0.043 13. C 0.6 0.95] Censoring: 26.07 34.61 -4.216 -1. 16.2. The next columns subsequently report performance measures of investment strategies that use the different predictions as input for an asset allocation as discussed in Section 3.70 2. Based on the realized returns of the asset allocations.48 4. L 0 1. 0.54.10 -0.153. Sharpe Utility Fee Weight Mean Bull Bear Ratio Basic.08 -0.