You are on page 1of 11

JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 24

Signal Weighting
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

RICHARD GRINOLD ctive strategies face the challenge level. Every strategy has an internal risk budget
was a global director of
research at Barclays Global
Investor (1994–2009),
director of research at
BARRA, (1979–1993), and
a professor at the University
A of combining information from dif-
ferent sources in order to maximize
that information’s after-cost effec-
tiveness.1 This challenge is evident for investors
who follow a systematic, structured investment
that shows how the aggregate risk is being allo-
cated among the various sources/themes that
we anticipate can add value. We should peri-
odically, say, every six months, re-examine and
possibly change that allocation. A similar recon-
of California in Berkeley, process. It is also an issue, although often unrec- sideration of the internal risk budget should
CA (1969–1989). ognized, for more traditional managers who take place if there is a significant change in the
must balance top-down views and the views signal mix (e.g., adding a new theme) or if there
of in-house and sell-side analysts. In any case, is a major change in the market environment.
a decision is made either through default, The method we present in this article
analysis, or whimsy. We favor analysis and to gives substance to the signal-weighting process
that end this article presents a simple and struc- by isolating the important aspects of the deci-
tured approach to the task. This task is com- sion and providing a structure that links those
monly called signal weighting, although risk features to a decision. The procedure we out-
budgeting seems to be a better description. line mimics the investment process. Each
The approach is based on standard methods signal is viewed as a sub, or source, portfolio
of portfolio analysis and their extension to that is a potential place to allocate risk. We
dynamic portfolio management. allocate risk to the source portfolios in a
We can consider an active investment manner that makes the most sense at the
product on three levels: strategic, tactical, and aggregate level. In this effort we are trying to
operational. At the strategic level are the target be comprehensive enough to depict the essen-
clients, fee structure, level of risk, amount of tials and simple enough to easily capture the
gearing, level of turnover, marketing channel, link between inputs and outputs. Thus, we
decision-making structure, and so forth. At the compromise between the desire to be rea-
operational level, we run the product: main- sonable and robust, on the one hand, and to
taining and replenishing data, executing trades, provide a clear box rather than a black box,
tracking positions and performance, informing on the other.
clients, and so on. At the tactical level, we allo- To preserve transparency we should err
cate resources to improve the product and its on the side of simplicity rather than on exces-
operations, and do some fine-tuning on such sive elaboration. There is a temptation to more
parameters as risk level and turnover level. The and more closely mimic the actual investment
signal-weighting decision belongs at the tactical process (e.g., backtesting). Our unadorned


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 25

approach is to assume the portfolio is run with the straight- cost-free world with multiple signals. This exercise allows
forward and easy-to-analyze method A, although we realize us to introduce many of the terms and concepts we will
that we actually use the complex and impossible-to-ana- use in the more difficult setting with transaction costs.
lyze method B. We have to trade off the benefits of get- The next step is to introduce transaction costs. Costs
ting sensible, comprehensible answers against the costs of are difficult to handle in an analytic way. Our approach is
departing from the operational reality. Any effect we want traditional; we make an assumption. We assume that the
to consider has to argue for a place at the table. The watch- portfolio manager acts as if he is following a simple rebal-
word is simplicity and the burden of proof is against making ance rule. The rebalance rule has three parts:
it complicated. This focus on simplicity also argues against
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

a profusion of sources. In our view, the number of themes • For each signal we construct a source portfolio that
The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

should be more than two and certainly less than, say, seven. captures that signal in an efficient (i.e., most signal
If you believe there are a larger number of significant alpha per unit of risk) manner.
drivers of a strategy you should 1) look up “significant” • We construct a model portfolio that is a mixture of
and 2) aggregate the signals into broader themes. the source portfolios.
One benefit of the proposed approach is that it is • We stipulate a rebalance rule to fix the rate at which
based on a theory. There is a logical consistency that lets the portfolio manager tries to close the gap between
us track cause and effect. The assumptions are few and the portfolio the manager actually holds and the
transparent. It is not a sequence of ad hoc decisions and model portfolio.
jury-rigged improvisations where the often hidden
assumptions are many and their implications are obscure. In Appendix A, we show that this type of rebalance
This clarity makes it easy to link cause and effect. In par- rule is optimal under certain conditions.2 We feel that a
ticular, it facilitates the use of sensitivity analysis to study rule, that is optimal under one set of conditions, will be
the links between important inputs and outputs. reasonable in a much wider group of situations. In
Signal weighting is a forward-looking exercise. We Appendix B, we study the consequences of using such a
are planning for the future not over-fitting the past. Past decision rule. The signal-weighting model and results are
performance of signals can, of course, inform us about illustrated using a three-signal example.
each signal’s presumed future strength, how fast the infor-
mation dissipates, and how the signals relate to each other. OTHER WORK
However, we will usually be in a situation with old sig-
nals that are part of the implemented strategy and new sig- This article stems from the author’s previous work
nals that have only seen hypothetical implementation. In on dynamic portfolio analysis, Grinold [2006, 2007]. The
addition, we may feel strongly that some signals will lose most directly related paper is Sneddon [2008]. Similar
strength because they are being arbitraged away. What we techniques to those described in Appendix A are used by
see in a backtest is one sample from a nonstationary prob- Garleanu and Pedersen [2009].
ability distribution. By all means consider past perfor-
mance, but add a large pinch of salt. ONE SIGNAL AND NO
The analysis is designed for an unconstrained imple- TRANSACTION COSTS
mentation. There may be some idiosyncratic features of
a signal that will make it unsuitable for say a long only We start with a quick description of the framework
implementation. These are difficult to predict, so the with only one signal and then expand the analysis to cover
analysis here should still provide a starting point of any multiple signals. There are N assets whose return covari-
study of the interactions of signals and constraints. ance is described by the N by N covariance matrix V.
The basic material in the weighting exercise is an N ele-
ment vector a, called a raw alpha. It is scaled to have, on
PREVIEW average, an information ratio of one,

This article has the following structure. We initially a′ ⋅ V–1 ⋅ a = 1 (1)

consider a cost-free world with one signal, and then a


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 26

To obtain a weight, we need an assessment of the • it makes things much simpler when we introduce
strength of the signal. Let IR be the information ratio of transaction costs.
the signal, then
First, we will consider the multiple-signal case with
α = IR ⋅ a (2) no transaction costs.
We are done. With one signal we simply weight the
raw alpha to have the desired information ratio. J SIGNALS AND NO TRANSACTION COSTS
We can get to the same place by a longer route that
We have J signals (sub-strategies), each with a raw
approaches the problem from a portfolio perspective. The
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

alpha aj j = 1, …, J. Each aj is a vector with elements for

The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

portfolio route is worth the extra effort because it gen-

each of the N assets. In addition, we start with a level
eralizes in a natural way when there are multiple signals
playing field; each aj is standardized so that its informa-
and transaction costs.
tion ratio is one,3
We associate a source portfolio s with our signal,
a ′j ⋅ V −1 ⋅ a j = 1 (6)
−1 2
s = V ⋅ a ⇒ ω = s′ ⋅ V ⋅ s = 1
We can, as in Equation (3), associate a position with
The source portfolio is an efficient implementation each raw alpha. These positions have, on average, 100% risk
of the signal, that is, among all positions x, it maximizes and will be the building blocks of our investment strategy,
the ratio
a j = V ⋅ s j , s j = V −1 ⋅ a j
a⋅x x′ ⋅ V ⋅ x s ′j ⋅ V ⋅ s j = 1

As noted in Equation (3), the source portfolio has an Because we have multiple signals, we have to take
excessive risk level of 100%. Suppose we choose a more into consideration any correlation between the informa-
realistic risk level ω. This is akin to scaling the source port- tion sources. We define that correlation in terms of the
folio. If ω is the risk level, then the portfolio is q = ω ⋅ s. expected return covariance between any pair of positions,4
This portfolio has an alpha of IR ⋅ ω and a variance of ω 2.
If we have a risk penalty, λ, and we do the standard mean- ρ i , j ≡ si′ ⋅ V ⋅ s j (8)
variance optimization, we choose the risk level ω to
maximize, For example, if i is a value signal and j is a momentum
signal, they will tend to be negatively correlated. Recall
λ 2 that signal weighting is a forward-looking exercise; the cor-
IR ⋅ ω − ⋅ω (4)
2 relations ρi,j are predictions. Nonetheless, a historical analysis
of past correlations should be informative;5 often, a fairly
The solution is ω = IR λ . For example, if IR = 0.75, stable relationship exists. In other cases, the relationship
and we have a risk penalty of λ = 18.75, then ω = 4%. appears to depend on the caprice of the market. As a prac-
The alpha resulting from this is tical matter, if any of these correlations are very large (say,
above 0.5), it might be wise to combine the signals. If a
α = λ ⋅ V ⋅ q = λ ⋅ V ⋅ s ⋅ ω = {λ ⋅ ω} ⋅ a (5)
correlation is large and negative, then either you have found
which gets us to the same place as the straightforward a very good thing or, more likely, you have uncovered
approach in Equation (2) since λ ⋅ ω = IR. something that is too good to be true.
The reasons for the portfolio optimization approach We build a strategy by assigning risk levels, ωj, to
used in Equation (5) are that each of J sub-strategies. The vector representation is
ω = {ω j } j =1, J . The result will be portfolio
• it generalizes nicely when we have multiple signals,
and q( ω ) = ∑s j
⋅ωj (9)
j =1, J


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 27

The information ratios of the J sub-strategies are EXHIBIT 1

given by the J element vector IR = {IRj}j=1, J. The J by Three-Signal Example
J correlation matrix with elements ρi,j is denoted as R.
With this notation the alpha expected by using the
weights ω is

α (ω ) = ∑ IR j
⋅ ω j = IR ′ ⋅ ω (10)
j =1, J Note: The input data are the information ratio, IR, and the correlation.
The optimal result is in the column labeled “Risk.”
The variance of the portfolio, q(ω), with weights ω is
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

σ 2 (ω ) = ∑ω ⋅ ∑ ρ i i, j
⋅ω j = ω′ ⋅ R ⋅ ω (11)
i =1, J j =1, J The following is an example with three sub-
strategies. In anticipation of the next section, we have
We choose ω to maximize, classified the signal by the rate of information turnover.
The sources are SLOW, INTERMEDIATE, and FAST.
α (ω ) − ⋅σ 2 ( ω ) (12) The goal is to mix these three sources to get an ideal port-
2 folio with 5% risk.
The data in Exhibit 1 represent our best estimates for
The solution is obtained by solving the first-order
a reasonably long planning horizon, say, a year or more.
Referring to Equation (16), the net effect of these data is an
aggregate information ratio of IRQ = 1.11. If we want an
IRi = λ ⋅ ∑ ρi , j ⋅ ω j overall risk of ωQ = 5.00%, we can set the risk penalty as
j =1, J
λ = IRQ/ωQ = 1.11/0.05 = 22.17 and achieve the desired
result. The final column is the optimal risk level derived
IR = λ ⋅ R ⋅ ω (13) from Equation (14). We can interpret this as running the
INTERMEDIATE sub-strategy at 3.78% risk, the SLOW
The maximizing vector of weights is given by sub-strategy at 1.96% risk, and the FAST sub-strategy at
4.10% risk.
ω= ⋅ R −1 ⋅ IR (14)

We will refer to the optimal portfolio (a.k.a., the Signal weighting, like portfolio management, is an
ideal portfolio) as Q with positions exercise in risk budgeting. We can express the result in
terms of the resulting risk budget. The variance of the
q( ω ) = ∑s ⋅ ωj (15) ideal portfolio Q is
j =1, J
i. ω Q2 = ∑ω ⋅∑ ρ j j ,k
The information ratio and risk of the ideal port- j =1, J k =1, J

folio Q are or
 ω j   ω k 
IRQ ≡ IR ′ ⋅ R −1 ⋅ IR , ω Q = ω ′ ⋅ R ⋅ ω , IRQ = λ ⋅ω Q ii. 1 = ∑ ω  ⋅ ∑ ρ j ,k ⋅   (17)
j =1, J  Q  k =1, J ω Q 
j ω ω
If we allocate a fraction {ωQ } ⋅ Σk =1, J ρ j ,k ⋅ {ωQk } of the
Note that IRQ is independent of the risk penalty λ,
risk to alpha source j, we are in effect taking the two
and thus we can choose λ to get the desired level of active
covariance terms for sources j and k, ωj ⋅ ρj,k ⋅ ωk + ωk ⋅
risk, ωQ.
ρk,j ⋅ ωj, and splitting the baby, thus attributing ωj ⋅ ρj,k ⋅
ωk to signal j and ωk ⋅ ρk,j ⋅ ωj to signal k.


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 28

In our example, over 52% of the risk comes from the EXHIBIT 2
FAST signal, 34% from the INTERMEDIATE signal, Three Views of the Solution to the Signal/Source-
and only 14% from the SLOW signal. Weighting Problem with No Transaction Costs


We can view the result in terms of alphas as well as

of portfolios. According to Equation (15), our ideal
portfolio is
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

q( ω ) = ∑s
j =1, J

The previous analysis showed how to successfully

The alphas that lead to this ideal are7 approach the source/signal-weighting problem when there
are no transaction costs. But the presence of transaction
αQ = λ ⋅ V ⋅ q = ∑a
j =1, J
⋅ (λ ⋅ ω j ) (18) costs introduces three new obstacles. These obstacles are

• Transaction costs are paid either directly through

WEIGHTS commissions, spread, taxes, and so forth, or indirectly
by demanding liquidity and shelling out too much
Equation (18) can be interpreted in terms of weights for purchases and getting too little for sales (a.k.a.,
(i.e., numbers that sum to one), as follows: market impact).
• Transaction costs are intimidating. They keep you
  from fully exploiting your information, which cre-
i. αQ = scale ⋅  ∑ a j ⋅ weight j  , where
 j =1, J  ates an opportunity loss.
• Transaction costs are levied on changes in positions,
ii. scale = λ ⋅ ∑ ω j
so the initial position is a crucial part of the analysis.
j =1, J (19)
iii. weight j = The last point implies a multi-period perspective is
∑ ω
i =1, J i required. Our approach will be to look for strategies or
decision rules that can, in some sense, be considered
The scale factor is to ensure that the resulting alphas optimal in a multi-period setting. To do this, we make
have the correct information ratio, for example, assumptions that allow us to abstract from the reality of
the day-to-day portfolio management environment and
α Q′ ⋅ V −1 ⋅ α Q = IRQ = 1.11 (20) still capture something of its essence. This is, after all, both
the power and the Achilles’ heel of any economic analysis.
The weights in our examples are 38% for INTER- As Solow [1956] said, “All theory depends on assumptions
MEDIATE, 20% for SLOW, and 42% for FAST. Note that that are not quite true.”
these are similar, but not equal, to the risk budgeting num-
bers mentioned previously.
The following three ways of presenting the results THE PORTFOLIO’S LAW OF MOTION
are illustrated in Exhibit 2: An active portfolio is an object in motion. Call it
portfolio P. If we look at portfolio P periodically, say, every
• risk of each source ∆t years,8 then we see a sequence of positions. The changes
• risk budget in these positions, defined as
• weights that sum to one
∆p(t) ≡ p(t) – p(t – ∆t) (21)
Now we turn to the more challenging and realistic
situation where we consider transaction costs. determine the cost.


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 29

The challenge is to capture these changes in a useful A MEASURE OF SIGNAL SPEED

way. To do this we will specify a rule, called the law of
motion, to show how portfolio P changes in response to We can gauge the rate of change in the signal j
changes in the source portfolios. We will do this in two between times t and t – ∆t by measuring the correlation
steps, as follows: of the positions sj(t) and sj(t – ∆t) using the asset covariance,
V(t), at time t. To do the calculation in a sensible way, we
1. Combine the source portfolios into a model port- should tailor the time interval, calling it ∆tj, on a signal-
folio M. by-signal basis so that a reasonable amount of change takes
2. Show how the portfolio P attempts to track the place resulting in a value for γj(t) in Equation (24) in the
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

model portfolio. range of 0.8 to 0.975. This will avoid spurious results if
The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

∆t is either too long or too short. Even if we keep daily

data we might want to calculate the correlation over one
week, two weeks, or one month so that we can see sig-
The model portfolio is a mixture of the source port- nificant changes.
folios. Its holdings at time t are This past correlation is given by

m( t ) ≡ ∑ s j ( t ) ⋅ ω j ⋅ ψ j (22) s ′j (t ) ⋅ V(t ) ⋅ s j (t − ∆t j )
j =1, J
γ j(t ) =
s ′j (t ) ⋅ V(t ) ⋅ s j (t ) ⋅ s ′j (t − ∆t j ) ⋅ V(t ) ⋅ s j (t − ∆t j )
The risk levels ωj are the same levels we used to (24)
define the ideal portfolio; see Equations (13) and (14).
The parameters ψj, where 0 < ψj < 1, are explained in We can eliminate the effects of the tailored time
detail in the next section. interval selection, ∆tj, by calculating a rate of change,


ln(γ j (t )) − g j ( t )⋅∆t j
g j (t ) = − or e = γ j (t ) (25)
The changes in portfolio P are governed by a linear ∆t j
decision rule of the form
The data will show how the signal has moved in the
i. p(t ) = δ ⋅ p(t − ∆t ) + (1 − δ ) ⋅ m(t ) past; for example, seasonality (e.g., quarterly or semi-annual
release of earnings) or a trend (e.g., getting faster or slower)
or may be present. The challenge is to use this analysis to
choose a point estimate that we will call gj, that will deter-
∆p(t ) (1 − δ ) mine how fast, on average, we believe the signal will move
ii. = ⋅ {m(t ) − p(t − ∆t )} (23)
∆t ∆t in the future. One way to consider this question is in terms

The parameter delta, 0 < δ < 1, is discussed later.

In Appendix A, we demonstrate that the rule described EXHIBIT 3
in Equations (22) and (23) is optimal under special cir- Values of gj for Various Half-Lives
cumstances. The leap of faith is to assume that the rule is
at least “reasonable” under more general conditions.
The parameters δ, ψj mentioned in Equations (22)
and (23) depend on the rate of change of the signals as
represented by either the raw alphas aj(t) or the source
portfolios sj(t) and on the change in the portfolio p(t)
itself. We now turn our attention to measuring those
changes and determining values for these parameters.


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 30

of a half-life, that is, the time it will take for the correlation, lags M. Both P and M are based on information that has
as calculated by Equation (24), to drop to 0.5. Exhibit 3 flowed in over previous periods. The age-weighted expo-
shows how gj depends on the half-life. sure of P to this information is about 1/d years longer
In our example, we will use nine months as the half- than the age-weighted exposure of M to the same infor-
life for the SLOW signal, one month for the FAST signal, mation. Slow trading leaves P under exposed to recent
and six months for the INTERMEDIATE signal. information and relatively over exposed to older stale
information. Exhibit 4 shows lags of one to six months
THE PACE OF PORTFOLIO CHANGE and the corresponding values for d.
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

The parameter delta in Equation (23) governs the

The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.


rate of change of portfolio P. We will express delta as
δ = e–d⋅∆t, where d measures how fast we attempt to close It is possible to get a rough estimate of d by exam-
the gap between where we start, p(t – ∆t), and where we ining the history of the portfolio and the source portfo-
would like to be, m(t). The choice of d, whether it is made lios. As we indicate in Appendix B, we can use a regression,
explicitly or implicitly, is an attempt to balance two costs: or portfolio optimization, approach to estimate parame-
ters δˆ(t ) and βˆ j (t ), so that
1. Transaction costs: A larger d (thus lower δ ) leads to
more trading and higher costs.
p(t ) = p(t − ∆t0 ) ⋅ δˆ(t ) + ∑ s (t ) ⋅ βˆ (t ) + εˆ(t )
j j
2. Opportunity loss: A smaller d (thus higher δ ) implies j =1, J
a less efficient implementation; the fraction of port-
folio P’s risk budget directed toward alpha declines and the unexplained variance, εˆ ′(t ) ⋅ V(t ) ⋅ εˆ(t ) , is mini-
and the amount of uncompensated risk increases. mized. If things are going along relatively smoothly (i.e.,
no radical changes of policy) and the time interval ∆t0 is
In Appendix A and in Grinold [2007], we show selected in a reasonable manner,9 then estimates δˆ(t ) will
how, under special circumstances, the optimal choice of tend to be positive and less than one. In such cases, we get
d is possible. In what follows, we first discuss the inter- an estimate of d as
pretation of d, which should isolate a reasonable range of
values, and then we consider procedures that will help us ln(δˆ(t )) ˆ − dˆ( t )⋅∆t0
dˆ(t ) = − , δ (t ) = e (27)
discover an implied value of d by examining the past ∆t0
behavior of the portfolio.
The analysis will also produce implied risk levels,
INTERPRETATION OF d βˆ j (t )/(1− δˆ(t )), and a measure of the fitting error,
{εˆ ′(t ) ⋅ V(t ) ⋅ εˆ(t )}/{p ′(t ) ⋅ V(t ) ⋅ p(t )}.
Portfolio P chases the model M. It is always behind From these estimates of implied past policy in
the curve. Indeed, 1/d is a measure of just how much P addition to the intuition that might be gained from
Exhibit 4 and future policy plans, we settle on values for
the parameters d and gj, j = 1, …, J. Then, for any rebal-
EXHIBIT 4 ance interval ∆t, we can calculate the risk adjustments ψj
Relationship between Trading Rate d and Lag promised in Equation (22),10
between Average Age of Information in Model
Portfolio M and Tracking Portfolio P − g j ⋅∆t
i. δ = e − d⋅∆t , γ j = e
1 −δ (28)
ii. ψ j ≡
1 −δ ⋅ γ j

Now let’s examine how this would work with the

example introduced earlier. The data in Exhibit 5 are


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 31

Assumed Half-lives of the Three Signals and Corresponding Values of gj, γj, and ψj based on ∆t = 1/52
and δ = 0.926
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.
The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

based on a value of d = 4 and a rebalance interval of one EXHIBIT 6

week, ∆t = 1/52, thus δ = 0.926. Risk Budgets and Weights before (Q) and after (M)
In Exhibit 5, the gj are based on the assumed half- Adjustment for Trading
lives. If HLj is the half-life in months, then gj = 12 ⋅ {ln
(2)/HLj}. The γj are calculated using item (i ) of Equation
(28) and the ψj are derived from item (ii) of Equation (28).
The column “Risk Q” repeats the optimal risk levels, ωj,
of the ideal portfolio previously reported in Exhibits 1
and 2. The final column, “Risk M,” contains the risk levels
ψj ⋅ ωj associated with the model portfolio, as in Equa-  
i. α M = scale ⋅  ∑ α j ⋅ weight j  , where
tion (22). The risk levels of all the signals fall, although the  j =1, J 
decrease, evident in the “psi” column, is more dramatic for
the FAST signal. The risk11 of model portfolio M is 3.06% ii. scale = λ ⋅ ∑ {ψ j ⋅ ω j } (30)
j =1, J
compared to the 5% risk level for the ideal portfolio Q.
To raise the risk level of the model portfolio to, say, 5%, {ψ j ⋅ ω j }
iii. weight j =
merely decreases the risk penalty lambda from 22.17 to ∑ i =1, J
{ψ i ⋅ ω i }
{3.06/5} ⋅ 22.17 = 13.57 and repeats the calculation. This
will scale up the risk, but keep the risk budget allocation
and weights the same. Exhibit 6 compares the results pre-adjustment (ideal
Q) and post-adjustment (model M) for transaction costs.
The adjustment for transaction costs works for alphas
as well as for portfolios. According to Equation (22), our The adjustment from ideal portfolio Q to model
model portfolio is portfolio M effectively lowers our sights from one unob-
tainable goal to another slightly less unobtainable goal. In
the process, we have left some potential value-added on
m= ∑s
j =1, J
⋅ {ψ j ⋅ ω j }
the table.
If we measure the value-added from any portfolio,
The alphas that would lead to this portfolio are say F with positions f, as

α M = λ ⋅ V ⋅ m = λ ⋅ ∑ a j ⋅ (ψ j ⋅ ω j ) (29) U F = αQ′ ⋅ f − ⋅ ω2 (31)
2 F
j =1, J

then the loss we incur by lowering our sights from Q

Equation (29) can, as before, be interpreted in terms λ
of weights (i.e., numbers that sum to one), as follows: UQ − U M = ⋅ ω2 (32)
2 Q−M


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 32

where ωQ–M is the risk of a position that is long the ideal EXHIBIT 7
portfolio Q and short the model portfolio M.12 In our Risk Budgets for the Portfolios: Ideal Q, Model M,
example that risk is ωQ–M = 2.58% and the corresponding and Actual P
loss in value-added is UQ – UM = 0.74%.


In Appendix B, we examine the results of following

the rebalancing strategy portrayed in Equation (23). If we
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

let ω i ,M ≡ Σ j =1, J ρi , j ⋅ ψ j ⋅ ω j be the covariance of source

The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

portfolio i with the model portfolio M, then we can easily

determine the alpha, risk, and implementation efficiency
of the implemented strategy P, as follows:

i. α P = λ ⋅ ω M2 = λ ⋅ ∑ ψi ⋅ ω i ⋅ ω i ,M If the target risk level for portfolio P is, say, 4.00%,

i =1, J
then we can alter the analysis to get that outcome. If we
ii. ω P2 = ∑ψ 2
⋅ ω i ⋅ ω i ,M (33) change the risk penalty lambda from 22.17 to 22.17 ⋅
i =1, J
{2.58/4.00} = 14.30, we will scale up the risks. The risk
iii. ω M2 −P = ∑ ψ ⋅ (1 − ψ ) ⋅ ω ⋅ ω
i i i i ,M
budget numbers in Exhibit 7 will not change.
i =1, J

As previously discussed, opportunity losses are
The opportunity loss just described was due to drop- incurred by lowering our sights from portfolio Q to port-
ping our sights from portfolio Q to portfolio M. Because folio M, 0.74%, and by, in fact, lagging behind M, 0.69%.
of trading costs, the portfolio we actually hold, P, will not We also will incur direct costs in the form of fees and
capture all the potential value added to the model portfolio commissions or indirect costs through market impact.
M. It is possible to take the analysis one step further (see We can estimate these costs as well (see Appendix A).
Appendix B) and look at the efficiency of portfolio P, the Let’s call the difference between portfolios M and P the
portfolio that results from using the updating rule described backlog. This is the trade we would make if we had one
in Equations (22) and (23). In our example, portfolio P has period free of transaction costs. We can measure the size
2.58% risk. Its correlation (transfer coefficient) with Q and of this backlog by its variance, ω M2 −P . The estimated annual
M is 0.727 and 0.857, respectively. The loss in potential transaction costs, cP, for portfolio P are a function of that
value-added moving from the ideal portfolio Q to the actual backlog variance
portfolio P is 1.43%. Given that we lost 0.74% by lowering
our sights from the ideal Q to the model M, we lose another cP = ⋅ ω M2 −P (34)
0.69% because we cannot exactly track M. That means a cer- 2
tain fraction of portfolio P’s risk budget will be residual Based on the data, this cost is 0.30%. In addition to
(i.e., not correlated with any source of alpha) and thus that the 1.43% opportunity loss, transaction costs bring the
risk is taken with no expectation of return. Exhibit 7 con- total bill to 1.73% and the annual value added after costs
tains the risk budgets for the three portfolios Q, M, and P. to 1.04%.
Exhibit 7 shows the magnitude of the compromises
forced upon us by trading frictions. With no trading costs,
we would allocate 51.74% of our risk budget to the FAST
signal. When costs are taken into consideration, however, One of the attractive features of our approach to
we allocate only 2.93% to the FAST signal. This is a huge signal weighting is the ability to do sensitivity analysis. The
shift. equations used in this article can be coded in a spreadsheet


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 33

EXHIBIT 8 of the characteristics of the resulting

Alpha Weights for the Model Portfolio M as Trading Rate d Varies strategy, including level of risk, risk budget,
and Other Parameters Are Held Constant opportunity loss, and annual trading cost.
Finally, our approach can change per-
spective by viewing each signal as a sub-
strategy and the weights as active risk levels
for the sub-strategy. These risk levels, in
turn, allow the calculation of risk budgets
that are a better measure of the importance
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

of each signal. It changes perspective by

The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

focusing on the dynamics of the signals

in a matter of minutes. As an example, we calculated the and the portfolio. The exercise of trying
weights of the model portfolio M for various values of the to measure the rate of change of the signals, gj, and the
trading rate parameter d, as shown in Equation (30). trading rate, d, of the portfolio requires a novel, dynamic
The cases in Exhibit 8 bracket the base case with look at the signals and the portfolio.
d = 4 that is summarized in Exhibit 6. We can also con-
sider the ideal portfolio Q as the case with very large d. ENDNOTES
The weight for the ideal portfolio is shown in Exhibit 6
as 19.95%, 38.42%, and 41.64% for SLOW, INTERME- In this article, the following terms—sources, themes, and
DIATE, and FAST, respectively. signals—are used as synonyms.
The technical appendices, A (Optimization) and B
(Policy Analysis), are available by writing to the author at
It is best practice to have this standardization work in a
We have presented a relatively straightforward time-average fashion. This allows a natural emphasis on a theme
approach to the signal-weighting problem in the face of when more information is available and, in particular, avoids
transaction costs. Our approach has several attractive attrib- trying to make something out of nothing when less informa-
utes. First, it is portfolio based. It looks at the signal- tion is on hand.
weighting problem as an investment problem by linking 4
Note this also implies that ρi , j = ai′ ⋅ V −1 ⋅ a j = ai′ ⋅ s j = a ′j ⋅ si .
each signal to an investment strategy and then looking 5
If there is a history of the past source portfolios sj (t), and
for the optimal mix of strategies. Second, the approach is dated covariance matrices, V(t), then ρ (t ) ≡ si′(t ) ⋅ V(t ) ⋅ s j (t )/

an extension of the solution of the signal-weighting

{ si′(t ) ⋅ V(t ) ⋅ si (t ) ⋅ s ′j (t ) ⋅ V(t ) ⋅ s j (t )} gives a measure of the
problem in the absence of costs. And third, it is forward
correlation at time t.
looking and depends on three traits of signals: predicted 6
If one or more of the ωj turn out to be negative, it just
information ratio, IRj; predicted return correlation, ρi,j;
means that we are willing to short that particular sub-strategy
and rate of change of each signal, gj. In addition, our or replace sj with –sj. This is not of any technical concern
approach depends on two investment strategy parame- although it definitely is a danger sign. λ ⋅ω j
ters: 1) trading rate of the portfolio, d, a surrogate for 7
Another way to say this is α Q = Σ j =1, J {IR j ⋅ a j } ⋅ { IR j }. In
turnover, and 2) a risk penalty, λ, that controls the strate- this case we have scaled each signal j to have the correct information
gy’s level of risk. ratio, IRj. If the signals are not correlated, ρi,j = 0 for i ≠ j, then
Additional attributes include the fact that it is easy the expression λ ⋅ ωj/IRj equals one for all signals. In the more
to perform sensitivity analysis with our approach, a capa- general case, we see λ ⋅ ω j /IR j = ω j /Σ k =1, J ρ j ,k ⋅ ω k as an adjust-
bility that is vitally important in establishing a firm under- ment for the correlations among the signals.
standing of the model and its results and in testing the In other words, monthly ∆t = 1/12, weekly ∆t = 1/52,
robustness of the answers to reasonable adjustments in the and so forth.
As mentioned earlier, even if the portfolio were to change
inputs. Our approach also lends itself to reverse engi-
on a daily basis, we would likely want ∆t to represent a longer,
neering in that it is relatively easy to adjust the risk penalty
say, one- or two-week timeframe, so that more substantial
and trading rate to attain the desired level of turnover and changes take place.
active variance. Furthermore, the approach predicts some


JPM-GRINOLD:JPM-GRINOLD 7/16/10 11:52 AM Page 34

As ∆t goes to zero ψj → d/{d + gj}.
If ω̂ j = ψ j ⋅ ω j , then ω M2 = Σ iωˆ i {Σ j ρi , j ⋅ ωˆ j }.
The portfolio Q – M has positions q − m = Σ j =1, J ,
s j ⋅ω j ⋅ (1 − ψj ).


Garleanu, N., and L. Pedersen. “Dynamic Trading with Pre-

dictable Returns and Transactions Costs.” Working Paper, Haas
It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

School of Business at the University of California, Berkeley,

The Journal of Portfolio Management 2010.36.4:24-34. Downloaded from by IMPERIAL COLLEGE on 09/07/10.

February 24, 2009.

Grinold, R. “A Dynamic Model of Portfolio Management.”

Journal of Investment Management, Vol. 4, No. 2 (2006), pp. 5-22.

——. “Dynamic Portfolio Analysis.” Journal of Portfolio Man-

agement, 34 (2007), pp. 12-26.

Sneddon, L. “The Tortoise and the Hare: Portfolio Dynamics

for Active Managers.” Journal of Investing, Vol. 17, No. 4 (2008),
pp. 106-111.

Solow, R. “A Contribution to the Theory of Economic

Growth.” Quarterly Journal of Economics, Vol. 70, No. 1 (1956),
pp. 65-94.

To order reprints of this article, please contact Dewey Palmieri at or 212-224-3675.