Professional Documents
Culture Documents
R
in Marketing
Modeling Dynamic Relations
Among Marketing and
Performance Metrics
Suggested Citation: Koen H. Pauwels (2018), “Modeling Dynamic Relations Among
Marketing and Performance Metrics”, Foundations and Trends
R
in Marketing: Vol. 11,
No. 4, pp 215–301. DOI: 10.1561/1700000054.
Koen H. Pauwels
Northeastern University, USA
koen.h.pauwels@gmail.com
This article may be used only for the purpose of research, teaching,
and/or private study. Commercial use or systematic downloading
(by robots or other automatic processes) is prohibited without
explicit Publisher approval.
Boston — Delft
Contents
1 Introduction 216
References 286
Modeling Dynamic Relations
Among Marketing and
Performance Metrics
Koen H. Pauwels1
1 Northeastern University, USA; koen.h.pauwels@gmail.com
ABSTRACT
Marketing and performance data often include measures
repeated over time. Time-series models are uniquely suited
to capture the time dependence of both a criterion variable
and predictor variables, and how they relate to each
other over time. The objective of this monograph is to
give you a foundation in these models and to enable
you to apply them to your own research domain of
interest. To this end, we will discuss both the underlying
perspectives and differences between alternative models,
and the practical issues with testing, model choice, model
estimation and interpretation common in empirical research.
This combination of marketing phenomena and modeling
philosophy sets this work apart from previous treatments
on the broader topic of econometics and time series analysis
in marketing.
216
217
Table 1.1: Comparing traditional and modern time series models of marketing and
performance
2.1 Introduction
219
220 Evolution, stationarity, and ARIMA
yt = µ + ϕyt−1 + εt
with
µ = a constant,
εt = a disturbance term. (2.1)
This model states that sales in period t are determined by sales in the
previous period t − 1. Depending on the value of ϕ we distinguish three
situations:
1. If |ϕ| < 1, the effect of past sales (and thus any ‘shock’ that has
affected past sales) diminishes as we move into the future. We call
such time series stationary, i.e. it has a time-independent mean
2.2. Autoregressive Processes 221
3. If |ϕ| > 1, the effect of past sales (and thus of past shocks) becomes
increasingly important. Such explosive time series behavior
appears to be unrealistic in marketing (Dekimpe and Hanssens,
1999, p. 5). Exceptions are typically stock variables, such as the
stock of cars in a country (Franses, 1994).
0.80
ACF
0.70
PACF
0.60
0.50
0.40
0.30
0.20
0.10
0.00
1 2 3 4 5 6 7 8 9 10
Figure 2.1: ACF and PACF patterns of an AR(1) process with ϕ = 0.7.
0.60
ACF
0.40 PACF
0.20
0.00
1 2 3 4 5 6 7 8 9 10
–0.20
–0.40
–0.60
–0.80
Figure 2.2: ACF and PACF patterns of an AR(1) process with ϕ = −0.7.
yt = µ − θ1 εt−1 + εt . (2.3)
This model is indicated as MA(1). Note that the past random shock
does not come from yt−1 , as in the AR(1) model, but it stems from the
random component of yt−1 . The ACF and PACF for the MA(1) model
are depicted in Figures 2.3 and 2.4. Here, the ACF shows a spike at
lag 1, which is negative if θ1 > 0 (Figure 2.3) and positive if θ1 < 0
(Figure 2.4) while the PACF shows exponential decay in the former
case, or a damped wavelike pattern in the latter.
0.80
ACF
0.70
PACF
0.60
0.50
0.40
0.30
0.20
0.10
0.00
1 2 3 4 5 6 7 8 9 10
Figure 2.3: ACF and PACF patterns of an MA(1) process with ϕ = −0.7.
224 Evolution, stationarity, and ARIMA
0.60
ACF
0.40
PACF
0.20
0.00
1 2 3 4 5 6 7 8 9 10
–0.20
–0.40
–0.60
–0.80
Figure 2.4: ACF and PACF patterns of a MA(1) process with θ = −0.7.
yt = µ + θq (B)εt
θq (B) = (1 − θ1 B − θ2 B 2 − · · · − θq B q ),
B = the backshift operator defined as before. (2.4)
The patterns observed for AR(1) and MA(1) in Figures 2.1 to 2.4
generalize to AR(p) and MA(q) processes. Specifically, we can recognize
AR(p) and MA(q) processes based on the ACF and PACF patterns:
0.80
ACF
0.60 PACF
0.40
0.20
0.00
1 2 3 4 5 6 7 8 9 10
–0.20
–0.40
Figure 2.5: ACF and PACF patterns of an ARMA(1) process with ϕ = θ = 0.5.
226 Evolution, stationarity, and ARIMA
0.60
ACF
0.40 PACF
0.20
0.00
1 2 3 4 5 6 7 8 9 10
–0.20
–0.40
–0.60
–0.80
Figure 2.6: ACF and PACF patterns of an ARMA(1) process with ϕ = θ = −0.5.
The orders (p, q) of an ARMA process are the highest lags of yt and
εt , respectively, that appear in the model.
As to marketing applications, Dekimpe et al. (1999) provide an
overview of published univariate time series results. They identify 44
studies between 1987 and 1994 that have published results from such
models, on a wide variety of durable (trucks, airplanes, furniture),
nondurable food (cereal, catsup, beverages) and nonfood (detergents,
toothpaste, cleaning aids) product categories and services (holidays,
passenger transit, advertising). Even earlier applications include Geurts
and Ibrahim’s (1975) forecasting of Hawaii tourist numbers with 4
different ARMA models.
zt = µ + ϕ1 zt−1 − θ1 εt−1 + εt
(2.5)
zt = yt − yt−1 .
The ACF of a nonstationary process decays very slowly to zero.
Therefore it is difficult to determine the AR and MA components
from the ACF and PACF. However these functions for zt (i.e. after
differencing), display the typical patterns for ARMA processes, as
illustrated above. In order to formulate the general ARIMA(p, d, q)
process, integrated of order d, we define the differencing operator
∆d = (1 − B)d . For example, if d = 1 this amounts to taking
zt = (1 − B)yt = yt − Byt = yt − yt−1 . For d = 2 we have quadratic
differencing zt2 = (1 − B)2 yt = (yt − yt−1 ) − (yt−1 − yt−2 ). Trends in
marketing are often linear and sometimes quadratic, so that d in rarely
greater than 2. The ARIMA(p, d, q) process can then be defined as:
The most common unit root test starts by redefining the AR(1)-
model (2.1) as follows:
zt = µ + γyt−1 + εt
zt = yt − yt−1 , and (2.7)
γ = ϕ1 − 1.
Then the unit root null hypothesis is H0 : γ = 0.
The reparameterized model (2.7) applied to data yields a t-statistic
for γ̂, which can be used to test H0 . This test is known as the Dickey–
Fuller test. However, the t-statistic that is obtained cannot be evaluated
with the regular tables of the t-distribution. Instead, special tables need
to be used. The generalization of the Dickey–Fuller test to an AR(p)
process yields the Augmented Dickey–Fuller test. This test is based on
a reformulation of the AR(p) process as:
The Augmented Dickey–Fuller (ADF ) test can be used to test the null
hypothesis γ = 0. A large number of lagged first differences should be
included in the ADF regression to ensure that the error is approximately
white noise. In addition, depending on the assumptions of the underlying
process, the test may be performed with or without the µ term in the
model. Enders (2003) offered an iterative procedure to implement these
different test specifications, as implemented in several recent marketing
papers (e.g. Slotegraaf and Pauwels, 2008; Srinivasan et al., 2004).
While the Augmented Dicky–Fuller (ADF) method is the most
popular unit root test in marketing, it is only one of the many developed
in light of the low power of the test under certain conditions (see Maddala
and Kim, 1996 for an excellent elaborate discussion). Because a unit
root is the null hypothesis of the ADF-test, particularly appealing are
alternatives that have stationarity as the null hypothesis, such as the
Kwiatkowski et al. (1992) KPSS test. Unfortunately, Maddala and Kim
(1996) find little convergence of test results for the typical series analyzed
in economics research. In contrast, the two applications in marketing
(Pauwels and Weiss, 2008; Pauwels et al., 2011) show that the ADF and
the KPSS test most often lead to the same conclusion for marketing and
2.6. Testing for Evolution versus Stationarity 229
Evolution Stationarity
(i) Technology diffusion factors
Technological breakthroughs (Christensen, 1997; D’Aveni, 1994) Lack of technological innovation (Pauwels and D’Aveni,
2016)
Growth stage in the product life cycle (Day, 1981) Maturity stage in the product life cycle (Lilien and Yoon,
1988)
Changing marketing effectiveness over product life cycle (Arora, Stable marketing effectiveness over product life cycle
1979) (Wildt, 1976)
(ii) Macroeconomic and regulation factors
Co-movements of performance with economy boosts, recessions Performance isolated from economic ups and downs or
and disruptions (Clark et al., 1984; Deleersnyder et al., 2007) related to stable economy
Regulations that open markets and stimulate competition Regulations that limit company growth and/or
bankruptcy options
(iii) Behavior of consumers and channel partners
Consumer learning (Alba et al., 1991) Consumer forgetting (Little, 1979)
Changing customer tastes/base (Abell, 1978) Consumer inertia (Ehrenberg, 1988)
Retail distribution and performance positive feedback loop Retailer inertia and assortment commitment (Broniarczyk
(Bronnenberg et al., 2000; Ataman et al., 2008) et al., 1998)
(iv) Competitive behaviour
Performance feedback in evolving business practice (Dekimpe and Normative marketing decision models (Kalish, 1985)
Hanssens, 1999)
Error-correcting behavior towards desired performance (Salmon, Satisficing managers (March and Simon, 1958)
1982; 1988)
Competitive quest for more (Hunt, 2000) incorporated in budgets Fixing marketing budgets as sales ratios in mature markets
Breakthrough changes in marketing strategy or practices, which Interdependent adaptation with competitors (Johnson
are not quickly matched by competitors (Hermann, 1997) and Russo, 1994) enables fast competitive cancellation of
marketing effects (Bass and Pilon, 1980)
Evolution, stationarity, and ARIMA
2.7. Unit Root Tests with Structural Breaks 231
Dekimpe et al. (1999) apply unit root tests and other time-series
methods to identify empirical generalizations about market evolution.
Based on data from hundreds of published studies, their meta-analysis
for performance measures is given in Table 2.2.
The models are classified into sales models and market share models.
The results show that evolution occurs for a majority of brand sales
series, and that a vast majority of the market share time-series models
are stationary. This is consistent with arguments of Bass and Pilon
(1980) and Ehrenberg (1994) that many markets are in a long-run
equilibrium. The relative position of the brands is only temporarily
affected by marketing activities. Even for brand sales (and category
sales) series, later studies suggest that stationarity, and not evolution is
the norm for mature brands in mature categories (e.g. Nijs et al., 2001;
Pauwels et al., 2002). Emerging markets and brands show a substantially
higher potential for evolution (Osinga et al., 2010; Pauwels and Dans,
2001; Slotegraaf and Pauwels, 2008). Future research is needed to update
the relative likelihood of evolution versus stationarity in today’s markets,
characterized by disruption of competitive advantages and redefinitions
of price-quality relationships (Pauwels and D’Aveni, 2016).
The earliest unit root tests assume no structural break in the time
series; i.e. if a unit root is found, it is assumed that any shock to
the series has a permanent effect. In contrast, it could be that most
(small) shocks have no permanent effect, while one or a few big shocks
do (Perron, 1989). Such big shocks are called ‘structural breaks’ and
232 Evolution, stationarity, and ARIMA
page views and brand choice for each newspaper. The total number of
online news visits initially evolves, but later stabilises. In contrast, page
views continue to evolve as usage depth increases over time. Finally,
brand choice is stable and proportional to the brand equity borrowed
from the printed newspaper. Thus, evolutionary growth first comes from
free riding on the wave of increasing online news readership, but then
needs to come from increases usage depth, which is brand specific.
What if there may be more than 1 structural break? The testing
procedure by Ben-David and Papell (2000) tests, for consecutive values
of M, the null hypothesis of M breaks against the alternative hypothesis
of M + 1 breaks. This testing follows the spirit of Chow (1960) tests for
parameter stability and aims to identify any remaining ‘irregular’ events
that could change a part of the model (Ben-David and Papell, 2000).
However, because Ben-David and Papell (2000) stop once stationarity
is established, Kornelis et al. (2008) propose the identification of all
unknown breaks as the stopping rule. They quantify the impact of
entrants in the Dutch TV advertising market and find that the steady-
state growth of incumbents’ revenues was slowed by these entries.
The ARIMA(0, 1, 1)s is the most common seasonal model and it provides
an exponentially weighted moving average of the data. This simple model
is written in two equivalent forms:
yt = µ + yt−s − θs εt−s + εt
(2.11)
∇s yt = µ + (1 − θs B s )εt .
Seasonal processes can be identified from the ACF and PACF functions,
similarly to the nonseasonal ARIMA processes above, except that the
2.9. Seasonal Processes 235
patterns occur at lags s, 2s, 3s, etc., instead of at lags 1, 2, 3, etc. Thus,
for these purely seasonal processes:
2. the ACF decays at multiples of lag s and the PACF has spikes
at multiples of lag s up to lag P × s, after which it is zero, for
seasonal AR processes;
or
237
238 Relating Marketing to Performance
yt = µ + v0 xt + v1 xt−1 + εt . (3.1)
yt = µ + vk (B)xt + εt
µ = a constant,
vk (B) = v0 + v1 B + v2 B 2 + · · · + vk B k , the transfer function,
(3.2)
B = the backshift operator, and
k = the order of the transfer function, which is to be
determined.
The transfer function is also called the impulse response function, and the
v-coefficients are called the impulse response weights. In the example,
if sales do not react to advertising in period t, but only to lagged
advertising, v0 = 0. In that case, the model is said to have a ‘dead
time’ (or ‘wear-in time’) of 1. In general, the dead time is the number of
consecutive v’s equal to zero, starting with v0 . An excellent visualization
of the different shapes these impulse response functions can take is
provided by Helmer and Johansson (1977, Figure 1) and applied to the
advertising-sales relationship.
Before we discuss how the order of the impulse response function
can be determined, we consider a well-known special case: the Koyck
model. In this model, the impulse response weights are defined by
vi = αvi−1 , for i = 1, . . . , ∞. Thus, the response is a constant fraction
of the response in the previous time period, and the weights decay
exponentially. It can be shown that Koyck’s model is equivalent to:
yt = v0 xt + αyt−1 + εt (3.3)
(1 − αB)yt = v0 xt + εt (3.4)
For the identification of these models from data, two problems arise.
The first is to find a parsimonious expression for the polynomial vk, ` (B),
and the second is to find an expression for the time structure of the
error term, in the form of an ARIMA model. An important tool in
the identification of transfer functions is the cross-correlation function
(CCF ), which is the correlation between x and y at lag k: ρ(yt , xt−k ).
The CCF extends the ACF for the situation of two or more series
(those of x and of y), with similar interpretation: spikes denote MA
parameters (in the numerator in (3.6)), and decaying patterns indicate
AR parameters (in the denominator in (3.6)).
We note that the simultaneous identification of the transfer function
and the ARIMA structure of the error in (3.3) is much more complex
than it is for a single ARIMA process. An example that occurs frequently
in marketing is one where the original sales series shows a seasonal
pattern, for which an ARIMA(1, 0, 1)(1, 0, 0)12 model is indicated.
240 Relating Marketing to Performance
described above for the estimation of the impact of such events. Put
another way, intervention analysis extends dummy-variable regression
to a dynamic context. The interventions may have two different effects:
a pulse effect, which is a temporary effect that disappears (gradually),
or a step effect that is permanent once it has occurred. A strike could
have a pulse effect on the production or distribution of a product.
The introduction of a new brand may permanently change sales of an
existing brand. The dummy variables for these two types of effects can
be represented as follows:
t Time t Time
ω 0 xtp
(c) ν (B)xtp = (ω 0 + ω 1 B) xtp (d) ν (B)xtp = α <1
yt (1 – α B (
yt
t Time t t +3 Time
t Time tt +1 Time
4.1 Introduction
245
246 Modern (Multiple) Time Series Models: The Dynamic System
VAR in Differences Dekimpe and Hanssens (1999) in the long run and short run, accounting for
Vector Error Correction model Baghestani (1991) the unit root and cointegration results?
4. Policy simulation analysis
Unrestricted impulse response (Section 4.5) What is the dynamic impact of
Pesaran and Shin (1998) marketing on performance?
Restricted policy simulation Pauwels et al. (2002)
Pauwels (2004) Which actors drive the
dynamic impact of marketing?
5. Drivers of performance
Forecast Variance Error (Section 4.6) What is the importance of each
Decomposition (FEVD) Hanssens (1998) driver’s past in explaining
Srinivasan et al. (2004) performance variance?
Generalized FEVD Nijs et al. (2007) Independent of causal ordering?
6. Policy recommendations Sims (1986) How do model results yield
(Section 4.7) Sims (1986) advice to policy makers?
Franses (2005)
247
248 Modern (Multiple) Time Series Models: The Dynamic System
and Wittink, 1992), daily marketing mix data where managers cannot
change marketing spending day-by-day (Wiesel et al., 2010) and hourly
advertising data (Tellis et al., 2005).
An important caveat is that Granger causality tests are pairwise, i.e.
can indicate variable x is Granger causing y while instead they are both
being driven (x earlier than y) by a third variable z. Thus, Granger
causality does not mean that x is the ‘ultimate’ causing variable of
interest; the researcher should further consider what Granger causes x
to develop a full understanding of the web of Granger causality and
how performance can be affected.
The most common test procedure for Granger causality goes back
to Granger (1969) himself, running the regression:
∞
X ∞
X
Yt = c + ρi Yt−i + βj Xt−j + ut . (4.2)
i=1 j=1
two fast moving consumer goods, Pauwels and Joshi (2016) find
that Granger causality tests reduce a set of 99 potential ‘key
performance indicators’ to only 17 performance-leading indicators.
4000
3500 ADVERTISING (K $)
3000 SALES (K $ )
2500
2000
1500
1000
500
0
1907 1912 1917 1922 1927 1932 1937 1942 1947 1952 1957
Figure 4.1: Long-term equilibrium between Lydia Pinkham’s advertising and sales.
y =a+b∗x+ε (4.3)
4000
3500
3000
2500
Sales
2000
1500
1000
500
0
0 500 1000 1500 2000 2500
Advertising
Marketing effort
Temporary Sustained
Sales Temporary Business-as-usual Escalation
response
Permanent Hysteresis Evolving business
practice
the four cells with unit root tests and panel VAR models (Sismeiro
et al., 2012). The largest group of U.K. physians combined stationary
prescribing behavior with sustained marketing increases (escalation),
which makes them less attractive for pharmaceutical companies.
(I − Φ1 B − Φ2 B 2 − · · · − Φp B p )yt = µ + (1 − θ1 B − θ2 B 2 − · · · − θq B q )εt
(4.5)
with yt the n × 1 vector of the j time series, and εt the n × 1 vector of
white noise errors, which are typically interdependent and identically
4.4. Dynamic System Model 259
2. Its reduced form (Eq. (4.10)) plays a key role in its estimation
and in the derivation of impulse response functions.
3. Its moving average form (Eq. (4.12)) plays a key role in deriving
the forecast error variance decomposition.
4. The AB-model, which nests both the K and the C model. For
instance, Keating (1990) bases himself on rational expectations
models to impose a set of nonlinear restrictions on the off diagonal
elements of B0 .
Note that none of these impose restrictions on the number of lags or the
exact form of dynamic effects; practices typical for dynamic structural
models in economics. A key critique of such restrictions is that economic
theory is uninformative on lags and that it is heroic to assume that
effects decay with the same fraction over time (Amisano and Giannini,
1997; Sims, 1980). Instead, VAR-modelers appear more willing to impose
restrictions on the contemporaneous relations and the error covariance
matrix.
Structural VARs have seen many applications in economics, but few
till recently in marketing (DeHaan et al. 2016; Freo, 2005; Gijsenberg
et al., 2015; Horváth et al., 2005; Kang et al., 2016). Freo (2005) applies
the AB model to study the response of store performance to sales
promotions. He causally orders the variables by decreasing Granger
exogeneity test statistics and deleting coefficients not significantly
different from 0. He finds that promotions in heavy household items
increase store revenues, but promotions in the textile category decrease
them instantaneously. Horváth et al. (2005) show that the inclusion
262 Modern (Multiple) Time Series Models: The Dynamic System
with all variable labels as before, and the square of the number
of endogenous variables K 2 also known as the number of freely
estimated paramaters (adding a lag requires the estimation of K 2 more
parameters). Lütkepohl (1993, p. 129) shows the exact relation between
AIC and FPE and concludes that they are equivalent for moderate and
large T . Indeed, experience shows that both criteria typically select the
same lag order p in marketing applications.
When the main researcher goal is selecting the correct VAR order
p, the FPE and AIC have the issue that they are not consistent; i.e.
they do not uncover the correct lag order p as the number of time series
observations goes to infinity (see Quinn, 1980 for the proof). Consistent
information criteria were developed by Schwarz (1978) and by Hannan
and Quinn (1979).
Schwarz (1978) derives his criterion from a Bayesian perspective, it
is hence often called the Bayesian Information Criterion (BIC). The
BIC criterion for a VAR of lag order p is given by:
ln(T )
BIC(p) = ln|MLCov(p)| + pK 2 . (4.15)
T
Compared to the AIC, the BIC has a stronger punishment for increasing
lag order p, because ln(T ) > 2 in virtually any time series application.
Finally, the Hannan–Quinn (HQ) criterion for a VAR of lag order p
is given by:
ln[ln(T )]
HQ(p) = ln|MLCov(p)| + pK 2 . (4.16)
T
Thus, Hannan–Quinn yields a stronger punishment than the AIC does,
but a lesser punishment than the BIC does for increasing the lag order.
In typical applications (moderate to large T ), the BIC and the HQ
criteria will choose a similar lag order, with the BIC sometimes choosing
a smaller lag orer than the HQ. (1993, p. 132) shows the BIC and the
HQ criteria are consistent, i.e. select the correct lag order p as the
number of observations T goes to infinity.
In practice though, researchers may care more about small-sample
properties than about the asymptotic properties of the lag selection
criteria. Simulations for T = 30 and T = 100 show that all criteria,
but especially the BIC, tends to underestimate the true lag order (set
268 Modern (Multiple) Time Series Models: The Dynamic System
What should the lag order of the ‘focus’ model be in the presence of
such robustness checks? Of course, the researcher can choose to stick to
a specific information criterion based on the above-mentioned benefits.
Alternatively, the researcher can consider all criteria, and choose a
lag order that performs rather well on each, typically in between the
FPE/AIC recommended and the BIC recommended ‘optimal’ lag.
In marketing applications, most researchers follow the BIC advice by
Hafer and Sheehan (1989) as it yields shorter-lagged models. Horváth
et al. (2005) use the Forecast Prediction Error for lag selection. The
remaining marketing researchers typically use the AIC criterion to
ensure against residual autocorrelation (e.g. Naik and Peters, 2009) or
check whether additional lags do not improve the model nor lead to
different conclusions (e.g. Nijs et al., 2001).
The trade-off between accuracy and degrees of can also be addressed
in different ways, such as pre-selecting endogenous variables and
estimating submodels (e.g. Srinivasan et al., 2004) and/or setting some
of the model’s parameters to zero. As to the latter, Gelper et al. (2016)
adapt the Lasso technique (Ren and Zhang, 2013; Tibshirani, 1996),
which minimizes the least squares criterion penalized for the sum of
the absolute values of the regression parameters. They demonstrate
the superior model performance (compared to Bayesian methods in
a simulation study) and forecasting accuracy of this sparse VAR
estimation of cross-category marketing effects (Gelper et al., 2016).
Such techniques hold much promise as the number of potential variables
4.4. Dynamic System Model 269
25
Immediate effect
Promotional elasticity
20
15 Cumulative
10 effect
5 Permanent effect = 0
0
–5
–10
1 2 3 4 5 6 7 8 9 10 11 12 13
Weeks
Figure 4.3: Impulse response of price promotion on sales for mature brand close-up.
4.00
Promotional elasticity
Immediate effect
3.00
0.00
1 2 3 4 5 6 7 8 9 10 11 12 13
Weeks
Figure 4.4: Impulse response of price promotion on sales for new brand Rembrandt.
Pauwels, 2008). This sales benefit is even permanent; i.e. the baseline
of sales has increased.
For a more detailed understanding of the impulse response function,
we start from the VAR-system in Dekimpe and Hanssens (1999) and
Pauwels (2004) with log sales (s), log of focal marketing action (f m), log
of other own marketing actions (om) and log of competitor marketing
actions (cm). Changes to sales represent consumer action and changes
to competitive marketing represent competitive action, while changes
to focal marketing and other marketing represent how the company
respectively sustains and support the focal marketing action. If all
variables are stationary, we may write our model in terms of deviations
from the steady-state means:
s∧ 0 0 0
t+p = β12 f mt+p + β13 omt+p + β14 cmt+p
1 1
+ β11 st+p−1 + β12 f mt+p−1 + · · · . (4.20)
Starting from the steady state, the updated forecast for the sales
deviation at time t becomes:
s∧ 0 0 0 0 0 0 0 0
t = β12 f mt + β13 omt + β14 cmt = β12 + β13 β32 + β14 β42 (4.21)
and the forecast for the sales deviation 1 period in the future:
s∧ 0 0 0 1
t+1 = β12 f mt+1 + β13 omt+1 + β14 cmt+1 + β11 st
1 1 1
+ β12 f mt + β13 omt + β14 cmt
4.5. Impulse Response Functions 275
40
35 sales change
30 sales
25
20
15
10
5
0
1 2 3 4 5 6 7 8 9 10
Figure 4.5: Full hysteresis in sales change and how it translates into sales.
0 0 0 0 0 1 0 1 0 0 1 0 0
= (β12 + β13 β32 + β14 β42 )(β21 β12 + β21 β13 β32 + β21 β24 β42
1 0 1 0 0 1 0 1 0 0
+ β23 β32 + β14 β42 ) + β13 (β31 β12 + β31 β13 β32
1 0 0 1 1 0 1 0
+ β31 β14 β42 + β32 + β33 β32 + β34 β42 )
0 1 0 1 0 0 1 0 0 1
+ β14 (β41 β12 + β41 β13 β32 + β41 β14 β42 + β42
1 0 1 0
+ β43 β32 + β44 β42 )). (4.22)
40
sales change
35
sales
30
25
20
15
10
0
1 2 3 4 5 6 7 8 9 10
–5
–10
Figure 4.6: Partial hysteresis in sales change and how it translates into sales.
effects. Hanssens et al. (2001) call this ‘partial hysteresis’ i.e. only
part of the immediate impact becomes permanent as market players
adapt (e.g. consumers become less excited and/or competitors imitate
the marketing action). Figure 4.6 illustrates this pattern, with the
immediate effect of 10, and a negative effect in the next period of −7,
leading to a net permanent sales increase of 3. Note that accumulating
the impulse response function coefficients for sales change (10 − 7 = 3)
gives us the net permanent effect on sales.
What happens when the impulse (i.e. marketing) is the evolving
variable? Again, we include the marketing action (e.g. price) in first
differences in the model, and its IRF therefore shows the performance
impact not of a temporary shock but of a permanent change (e.g.
reduction in regular price). We need to keep this in mind when
interpreting the result.
responses of the other variables for five periods. Second, Pauwels (2004)
decomposes the net performance impact of a marketing action into
the actions of consumers, retailers, competitors and own performance
feedback.
The answer depends on the endogenous variables that we allow to be
affected. If we allow only consumer (sales) response, we restrict future
own and competitive marketing actions to remain in steady state, i.e.
to have zero deviations (omt = fmt = 0). In that case, the forecast
expressions in (4.21) and (4.22) greatly simplify to respectively:
s∧ 0
t = β12 (4.23)
s∧
t+1 = 1 0
β11 β12 + 1
β12 . (4.24)
s∧ 0 0 0
t = β12 + β14 β42 (4.25)
s∧
t+1 = 0
β14 1 0
(β41 β12 + 1 0 0
β41 β14 β42 + 1
β42 + 1 0
β44 β12 . (4.26)
While impulse response functions track the effect of one shock to the full
system, forecast error variance decomposition (FEVD) reveals how the
forecast error variance in one variable (e.g. sales) can be explained by
its own past shocks and all the shocks of the other endogenous variables.
Analogous to a ‘dynamic R2 ’, it shows the relative importance of each
variable in having contributed to the variation in the performance
variable. An important issue in standard FEVD is the need to impose a
causal ordering for model identification purposes (see Hanssens (1998)
for a marketing application of FEVD). Like the Generalized Impulse
Response function, the Generalized FEVD (Pesaran and Shin, 1998)
does not require such a priori causal ordering. The GFEVD for time
4.7. Software 279
4.7 Software
While the other softwares are better for your own coding, Eviews
and OX/PcGive offer a low-threshold click-and-find software. Moreover,
Eviews has the most typical option as the default in its software and
updates regularly, adding the most recent tests and deleting those found
to be less desirable. Its student version is a low-cost option to get
started.
5
Conclusion and Extensions
Our discussion of econometric and time series models has focused on the
dynamic relations among marketing and performance metrics. Other
fascinating uses of time series analysis occur in the analysis and impact
of business cycles (see Dekimpe and Deleersnyder, 2018 for a thorough
review) and in the frequency domain, with the spectral analysis of
phenomena such as the periodicity of competitive pricing (Bronnenberg
281
282 Conclusion and Extensions
5.2 Policy Analysis Based on Modern Time Series Models and the
Lucas Critique
their expectations and behavior to the new policy (Lindé, 2001, p. 986).
In contrast, the modern time series models discussed in this monograph
are backward-looking summaries of past correlational patterns that are
assumed to persist in the future, regardless of policy changes (ibid).
Van Heerde et al. (2005) correctly argue that the Lucas critique is not
worrisome for ‘any change in policy’ (managers typically consider only
incremental actions, such as adding or dropping 1 price promotions
or direct mail in their marketing plan), but only for a major policy
change such as P&G ‘Value Pricing’ strategy substantially cutting
price promotions (Ailawadi et al., 2005) or Albert Heijn announcing a
permanent regular price reduction across categories (Van Heerde et al.,
2008).
But even for major policy changes, it is important to realize that the
above critique does not necessarily disqualify forecasing models for policy
analysis – instead we should compare their identifying interpretations
from those of alternatives (Sims, 1986). A currently popular alternative
to data-driven (reduced form) models are assumption-driven models,
also known as ‘structural models’. Not to be confused with ‘structural
VARs’ (Section 4.3), economic models are called ‘structural’ to the extent
that they formalize and estimate parameters of ‘objective functions of
agents’, described as characterizing the agents’ ‘tastes and technologies’
invariant to the policy under consideration (Hansen and Sargent,
1980). In marketing applications, the meaning of ‘structure’ has been
narrowed to consumers and managers ‘making optimal decisions based
on a maximimition of an objective function subject to constraints’
(Bronnenberg et al., 2005, p. 23). To be able to predict reactions to
policy changes, the claim is that the structural model specification
reflects all relevant aspects of these tastes and technologies. For any
model, the more the contemplated action differs from those undertaken
in the past, the more difficult it is to predict its consequences (Sims,
1986).
So how can the researcher address these issues? As detailed in
Franses (2005) any model used for policy analysis should first pass the
tests for descriptive and forecasting models. In addition, models for
policy simulation should have a constant-parameter equation for the
policy instrument – in which case the Lucas critique does not hold
284 Conclusion and Extensions
(Ericsson and Irons, 1995). If there are any changes in the parameters of
the instrument model, such as due to an announced permanent change
in price, the model for price should include such level shift, as should the
model for sales (Franses, 2005). Likewise, an assumption-driven model
can be formulated to explicitly incorporate expectations and endogeneity
(Bronnenberg et al., 2005). In the end though, the proof of the pudding is
in the eating: the ‘true test is predictive validity for data in which there
is a major change in policy’ (Van Heerde et al., 2005). Ailawadi et al.
(2005) do so by showing how their game-theoretic predictions outperform
those of reaction function models after P&G Value Pricing policy change.
Likewise, Wiesel et al. (2011) re-estimate their VAR-model after the
company agreed to a field experiment where it increased paid search
spending by 80% and reduced direct mail (its main marketing activity)
by 56%. The parameter estimates of the experimental period are like
those from the earlier period, with the exception that direct mail has a
stronger impact when it was reduced (as expected given diminishing
returns). An important research opportunity is to perform such an
exercise directly contrasting the predictive accuracy and outcomes of
an assumption-driven versus an evidence-based model. For instance,
in macroeconomics, virtually no evidence is found in support of the
empirical relevance of the Lucas critique (Ericsson et al., 1998).
In sum, the modeler faces the trade-off between strong identifying
assumptions (which typically reduce forecasting performance) and the
benefits of explicit and theory-derived primitives of rational consumers
and managers (such as forward looking behavior). After understanding
their relative strengths and weaknesses, the choice basically comes
down to personal preference and model purpose. For instance, New
Empirical Industrial Organization (NEIO) models allow us to uncover
the company’s costs (which are typically not available to researchers)
from its actions, if we are willing to assume profit maximizing behavior
and a rather extensive information set from the decision makers.
As Bronnenberg et al. (2005) state, assumption-driven (structural)
models are expected to have worse forecasting performance than
evidence-based models, but the latter ‘do not ensure that predictions
are immune to the Lucas critique’ (p. 23). However, neither does any
estimated ‘structural’ model, because the structure can be misspecified.
5.2. Policy Analysis Based on Modern Time Series Models 285
286
References 287
Pauwels, K. (2007). “Price War: What is it Good For? Store Visit and
Basket Size Response to the Price War in Dutch Grocery Retailing”.
Marketing Science Institute, Report 07–104.
Pauwels, K. (2014). It’s Not the Size of the Data – It’s How You Use
It: Smarter Marketing with Analytics and Dashboards. Amacom.
Pauwels, K., Z. Aksehirli, and A. Lackmann (2016). “Love the Ad or
Love the Brand?” International Journal of Research in Marketing.
33(3): 639–655.
Pauwels, K., I. Currim, M. G. Dekimpe, D. M. Hanssens, N. Mizik,
E. Ghysels, and P. Naik (2004a). “Modeling Marketing Dynamics
by Time Series Econometrics”. Marketing Letters. 15(4): 167–183.
Pauwels, K. and E. Dans (2001). “Internet Marketing the News:
Leveraging Brand Equity from Marketplace to Marketspace”. The
Journal of Brand Management. 8(4): 303–314.
Pauwels, K. and R. D’Aveni (2016). “The Formation, Evolution
and Replacement of Price–quality Relationships”. Journal of the
Academy of Marketing Science. 44(1): 46–65.
Pauwels, K., S. Erguncu, and G. Yildirim (2013). “Winning Hearts,
Minds and Sales: How Marketing Communication Enters the
Purchase Process in Emerging and Mature Markets”. International
Journal of Research in Marketing. 30(1): 57–68.
Pauwels, K. and D. M. Hanssens (2007). “Performance Regimes and
Marketing Policy Shifts”. Marketing Science. 26(3): 293–311.
Pauwels, K., D. M. Hanssens, and S. Siddarth (2002). “The Long-term
Effects of Price Promotions on Category Incidence, Brand Choice,
and Purchase Quantity”. Journal of Marketing Research. 39(4):
421–439.
Pauwels, K. and A. Joshi (2016). “Selecting Predictive Metrics for
Marketing Dashboards – An Analytical Approach”. Journal of
Marketing Behavior. 2(2–3): 195–224.
Pauwels, K., P. S. Leeflang, M. L. Teerling, and K. E. Huizingh (2011).
“Does Online Information Drive Offline Revenues? Only for Specific
Products and Consumer Segments!” Journal of Retailing. 87(1):
1–17.
298 References