SSRN Id2609814
SSRN Id2609814
Abstract
Revisiting the issue of return predictability, we show there is substantial predictive power in
combining forecasting variables. We apply correlation screening to combine twenty variables
that have been proposed in the return predictability literature, and demonstrate forecasting
power at a six-month horizon. We illustrate the economic significance of return predictability
through a walk-forward simulation, which takes positions in SPY proportional to the model
forecast equity risk premium. The simulated strategy yields annual returns more than twice
that of the buy-and-hold strategy, with a Sharpe ratio four times as large. To eliminate look-
ahead bias, we perform additional simulations including variables only as they are discovered in
the literature. Results show similar annual returns and Sharpe ratios. While a market-timing
strategy outperforms the market, it is difficult to implement.
* Blair Hull founded Hull Trading Company in 1985 and served as the firm’s chairman and chief executive
officer before selling it to Goldman Sachs in 1999. Hull is now founder and managing partner of Ketchum
Trading, LLC, a proprietary trading firm providing liquidity in options, futures, and equities. He also
manages his family office, Hull Investments, LLC. Xiao Qiao is a PhD candidate at the University of
Chicago Booth School of Business. We would like to thank Rick Anderson, Petra Bakosova, Mike Fearon,
and Jerome Pansera for data analysis, and John Halter and Brian Von Dohlen for contributing to the
original project. We thank Rick Anderson, Petra Bakosova, Dirk Eddelbuettel, Petri Fast, Alexios Galanos,
Jiahan Li, Jim Lodas, Jerome Pansera, and John Rizner for helpful comments. We are grateful to Chris
Jones and Matt Ringgenberg for sharing their data, and Martin Lettau and Robert Shiller for sharing their
data on their website.
Electroniccopy
Electronic copy available
available at:
at:https://ssrn.com/abstract=2609814
http://ssrn.com/abstract=2609814
1. Introduction
A central question in financial economics is the estimation of expected market returns.
Financial claims on real assets bear non-zero returns for two reasons. First, one dollar received
tomorrow is not equal to one dollar received one year from today, since investors demand
compensation for non-immediacy. The second source of returns comes from the fact that
many financial assets are risky, and investors are compensated for holding these risky assets.
For the aggregate equities market, this adjustment for risk is known as the equity premium.
It is well-known that the equity premium is difficult to estimate. Merton (1980) called attempts
to estimate the equity premium a “fool’s errand”: “Indeed, even if the expected return on the
market were known to be a constant for all time, it would take a very long history of returns to
obtain an accurate estimate. And, of course, if this expected return is believed to be changing
through time, then estimating these changes is still more difficult” (Merton 1980, p. 326).
Much of the empirical asset pricing literature up until Merton (1980) assumed a constant rate
of return for the market, while Merton anticipated the possibility of a non-constant equity
premium. Indeed, the equity premium may be time-varying, and move around depending on
prevailing business conditions.
If the equity premium is time-varying, then presumably given the appropriate information set it
can be forecasted. Early evidence of Fama and French (1988) and Campbell and Shiller (1988a,
1988b), inter alia, showed that market returns can be predicted using dividend yields. However,
evidence both for and against return predictability cropped up in the years following these
pioneering works. In an influential study, Goyal and Welch (2008) examined fourteen different
forecasting variables proposed by academics, and found that the predictors are unstable both
in-sample and out-of-sample. They concluded the variables would not have helped investors
profitably time the market. On the other hand, Cochrane (2008) put forth a strong defense of
return predictability by jointly examining the forecastability of returns and dividend growth.
There appears to be evidence for predictability over both the long and short term. At the one-
month frequency, Moskowitz, Ooi, and Pedersen (2012) document that past 12-month market
excess return is a positive indicator of the next month market return. Dividend yield (see
Campbell and Shiller 1988a, 1988b; Cochrane 2005) also has some forecasting power for next
month’s market returns, which becomes stronger at longer horizons of one to five years, as the
R-squared of the forecasting regression rises with forecasting horizon. We include a variety of
variables that the literature has demonstrated to work at various frequencies, and we combine
them to extract more information than univariate forecasting regressions.
We study return predictability along several novel dimensions. We utilize many predictors from
the predictability literature, and combine them to produce a better forecast. Many previous
1|Page
Electroniccopy
Electronic copy available
available at:
at:https://ssrn.com/abstract=2609814
http://ssrn.com/abstract=2609814
studies address predictability in isolation, running univariate forecasting regressions. We know
many candidate variables that may forecast the equity premium, but it is unclear if they all
carry different amounts of information, or if they approximate some small set of state variables
that govern future investment opportunities. We show that different predictor variables
contain different information about future returns, at various horizons.
By combining predictors with diverse characteristics, we can produce a superior return forecast.
Like Goyal and Welch (2008), we look at a number of different forecasting variables. Unlike
Goyal and Welch (2008), we examine the joint forecasting power of all of these variables, and
find multiple predictors outperform univariate forecasting regressions. Rapach, Strauss, and
Zhou (2010) argue forecast combination using multiple predictors outperforms the historical
average. Our paper is similar in that we also combine the information contained in multiple
variables, but we look at a broader set of variables (including technical indicators,
macroeconomic variables, return based predictors, price ratios, commodity prices, etc) and we
combine them using correlation screening (Hero and Rajaratnam 2011).
Through our implementation of the market-timing strategy, we stress the importance of taxes,
transactions costs, and other implementation difficulties. Most return forecasting articles stop
2|Page
There are several shortcomings to the current state of literature on return predictability.
Previous studies often restrict the return series to monthly data. Although higher frequency
data has been available for many years, it is messy to deal with data at different frequencies.
Previous work preferred clean statistical results over sacrificing some rigor to create a system
that works well in practice. Our primary focus is to create a system that is implementable, so
we willingly deal with predictors designed to capture different frequency returns. Many studies
examine return predictors in isolation. Some studies, such as Rapach, Strauss, and Zhou (2010),
have attempted to combine information across predictors, but they only use a small set of
predictors restricted to a similar time horizon. Instead, we look at a relatively large set of
predictors, and combine them in sensible ways to produce better forecasts than they do
separately. Previous studies often rely exclusively on ordinary least squares (OLS) in forecasting
regressions. In contrast, we use correlation screening to filter out the least significant variables
and combine predictors.
Many economic decisions require the input of an estimated equity premium. Superior
decisions can be made based on a better forecast of future market returns. Individual and
institutional investors both face the problem of asset allocation, for which a good estimate of
the equity premium is strongly desired. Traditional investment advice is that market timing is
hopeless, and investors should seek to keep a constant split between stocks and bonds instead
of strategically changing the proportions. At the 2013 Rebalance IRA Conference (Center for
Retirement Investing), Burton Malkiel stated “Don’t try to time the market. No one can do it.
It’s dangerous”.
Market timing is also related to active management. Passive funds often beat active ones, and
mutual fund managers who do well in one year are no more likely to do well in the following
year (Berk 2005, Carhart 1997). During the recent financial crisis, our investment fund adjusted
our portfolio by investing more in equities as the market declined, but our overall performance
was less than stellar. To market time, we need sufficient evidence that actively managing the
portfolio will beat passively investing in the index.
The rest of the paper proceeds as follows. Section 2 describes the forecasting variables we use,
and our data sources. Section 3 presents the forecasting results. Section 4 discusses the details
3|Page
For some of the variables, we use their raw values in forecasting returns. For others, we
transform the variables into an exponential moving average (EMA), or the log of the raw values
minus their EMAs. The EMA of a raw variable creates a persistent series that captures a slow-
moving component of market returns. Log of the raw value minus its EMA is similar to an
innovation, which may capture a short-term component in market returns. For all of the
variables, we examine the forecasting performance of the raw values and various
transformations, staying true to the form proposed in the original papers whenever possible.
The variables considered are the following:
1. Dividend-Price Ratio (DP): Fama and French (1988), Campbell and Shiller (1988a, 1988b),
and Cochrane (2008), among others, have shown the dividend-price ratio can be used to
forecast future market returns. If the current dividend-price ratio is high, future returns
are also likely to be high. We use the log of a twelve-month moving sum of dividends
paid on the S&P 500 index minus the log of S&P 500 prices.
2. Price-to-Earnings Ratio (PE): In the classic work of Graham and Dodd (1934), PE was
used as an indicator of value. Campbell and Shiller (1988b) report that the PE ratio
explains as much as 40% of future returns. A high price-to-earnings ratio today indicates
a low equity premium. We use the price divided by earnings over the last 12 months.
3. Book-to-Market Ratio (BM): Pontiff and Schall (1998) propose using the book-to-market
ratio of the Dow Jones Industrial Average (DJIA) to predict market returns. They find the
DJIA book-to-market ratio contains information not captured by DP. A high current
book-to-market ratio indicates high future market returns. We use the book value of
the S&P 500 divided by the S&P 500 index, SPX.
4. Cyclically Adjusted Price to Earnings Ratio (CAPE): This is also known as the Shiller PE.
Shiller (2000) use CAPE, price divided by the average inflation adjusted earnings over the
4|Page
The bulk of the data we use comes from publicly available sources. We obtain the necessary
data to construct DP, PE, BM, BY, DEF, TERM, CAY, SIM, VRP, IC, BDI, PCR, MA, PCA-tech, OIL,
and SI. CAPE is constructed with data from Bloomberg and the Federal Reserve Bank of St.
Louis. NOS is from the U. S. Census Bureau, and CPI is from the Federal Reserve Bank of St.
Louis. Short interest data from Rapach, Ringgenberg, and Zhou (2015) is kindly provided by
Matt Ringgenberg, although in our results we use our own definition of SI. We use the
difference between the realized returns on SPX from Bloomberg and 90-day Treasury Bill as our
forecasting target.
Table I shows the summary statistics of the forecasting variables. We include daily data from
06/08/1990 through 05/04/2015. It is evident the variables we use have different
characteristics along several dimensions. Volatility, skewness, and kurtosis are three measures
that differ wildly across the board. We see all combinations of volatile/calm, positive/negative
skew, and fat/thin tails among the predictors. We try to incorporate variables that may contain
different information into one analysis to capture the most informative signal about future
market returns. The predictor variables in Table 1 were discovered in the literature at various
times. We discuss this issue in more detail in the next section.
7|Page
nobs Min 25th Mean Median 75th Max Stdev Skewness Kurtosis
DP 5576 0.012 0.018 0.023 0.021 0.025 0.047 0.007 1.183 1.142
PE 6280 10.132 16.461 19.611 18.255 23.106 31.134 4.408 0.622 -0.465
BM 6280 0.192 0.324 0.372 0.369 0.438 0.688 0.088 -0.009 -0.304
CAPE 6280 11.668 21.763 25.907 24.993 28.480 43.431 6.098 0.820 0.452
PCA-price 5576 -5.939 -0.277 2.055 1.104 3.153 12.533 3.618 1.037 0.547
BY 6280 -1.399 -0.271 -0.060 -0.070 0.148 1.129 0.312 -0.003 0.330
DEF 6280 0.430 0.700 0.954 0.870 1.050 3.500 0.413 3.158 13.230
TERM 6280 -0.989 0.916 1.866 2.001 2.790 3.863 1.157 -0.290 -0.981
CAY 6280 -0.032 -0.013 0.001 -0.002 0.014 0.031 0.016 0.124 -1.017
SIM 6280 0.000 0.215 0.471 0.469 0.723 0.954 0.289 0.001 -1.220
VRP 6279 0.000 0.000 0.642 0.000 0.170 10.036 1.460 2.777 8.438
IC 6280 -44.552 -4.864 0.502 0.000 5.604 43.436 9.128 0.015 1.288
BDI 6280 -0.688 -0.049 -0.003 0.000 0.054 0.740 0.117 -0.641 6.523
NOS 6280 -0.204 -0.011 0.009 0.012 0.033 0.142 0.043 -1.030 4.774
CPI 6280 -0.020 0.017 0.024 0.026 0.031 0.055 0.012 -0.652 1.089
PCR 6280 -2.606 -1.681 -1.370 -1.428 -1.087 -0.315 0.441 0.271 -0.409
MA 6280 0.000 1.000 0.780 1.000 1.000 1.000 0.414 -1.350 -0.178
PCA-tech 6280 -7.420 -1.439 0.198 1.972 2.165 2.299 2.983 -1.382 0.498
OIL 6280 -1.206 -0.085 0.010 0.025 0.112 0.860 0.184 -0.724 4.290
SI 6280 -2.633 -0.453 0.065 0.041 0.756 2.547 1.063 -0.257 0.039
R_1M 6258 -0.354 -0.016 0.006 0.011 0.033 0.202 0.046 -1.053 4.877
R_3M 6215 -0.527 -0.014 0.019 0.028 0.065 0.328 0.078 -1.254 4.604
R_6M 6150 -0.644 -0.005 0.039 0.054 0.107 0.433 0.118 -1.563 5.005
R_12M 6028 -0.670 0.028 0.074 0.104 0.183 0.522 0.172 -1.353 2.260
8|Page
PCA- PCA-
DP PE BM CAPE BY DEF TERM CAY SIM VRP IC BDI NOS CPI PCR MA OIL
price tech
DP
PE -0.38
BM 0.48 -0.76
CAPE -0.59 0.75 -0.96
PCA-price -0.63 0.92 -0.93 0.97
BY -0.03 0.08 -0.12 0.13 0.10
DEF 0.15 -0.39 0.53 -0.48 -0.41 -0.16
TERM 0.22 -0.12 0.49 -0.52 -0.44 0.17 0.25
CAY 0.42 0.15 -0.02 -0.07 0.03 0.07 -0.08 0.12
SIM -0.15 0.04 -0.07 0.05 0.07 0.21 -0.03 0.06 0.07
VRP 0.05 -0.08 0.22 -0.17 -0.07 -0.19 0.54 0.04 0.16 -0.13
IC 0.12 -0.16 0.07 -0.14 -0.15 -0.23 0.36 -0.06 0.12 0.01 0.38
BDI -0.09 0.06 -0.07 0.06 0.08 0.11 -0.12 -0.03 0.05 -0.03 0.11 -0.09
NOS -0.14 -0.19 -0.15 0.15 0.00 -0.01 -0.32 -0.32 -0.32 -0.05 -0.39 -0.04 -0.12
CPI 0.08 0.06 -0.20 0.16 0.04 -0.09 -0.21 -0.18 -0.13 -0.02 -0.39 -0.04 -0.15 0.35
PCR -0.65 0.60 -0.84 0.86 0.87 0.05 -0.21 -0.36 -0.16 0.03 0.02 0.03 0.02 0.02 -0.05
MA 0.00 0.11 -0.21 0.25 0.12 0.17 -0.54 -0.16 -0.09 0.00 -0.41 -0.41 0.03 0.23 0.11 0.05
PCA-tech 0.02 -0.05 -0.07 0.13 0.02 0.25 -0.48 -0.15 -0.06 0.05 -0.38 -0.38 0.00 0.22 -0.05 -0.06 0.80
OIL -0.19 0.08 -0.09 0.11 0.14 0.35 -0.21 0.00 -0.01 0.14 -0.09 -0.11 0.29 0.00 -0.06 -0.09 0.04 0.06
SI 0.14 -0.15 0.18 -0.22 -0.17 -0.10 0.34 -0.05 0.03 -0.03 0.13 0.21 -0.01 0.05 0.23 -0.34 -0.31 -0.25 0.05
Table 2 presents the pairwise correlations of the forecasting variables. We have highlighted in
shades of green the positive correlations, and shades of red the negative correlations. This
table further highlights the diversity of our variables. Some variables apparently carry
information that is weakly correlated with other variables. For example, OIL is positively
correlated to BY and BDI, but only mildly correlated with most of the other variables. BDI is
positively correlated with OIL, and has low correlations with the rest of the variables. BY is also
not highly correlated with most of the variables, except OIL. The upper-left corner of price
ratios are all highly correlated, or highly negatively correlated, depending on if price is in the
numerator or the denominator. Since the four price ratios, DP, PE, BM, and CAPE, contain
similar information, in our forecasting models we have tried including all four separately, or
including the first principal component of these series, PCA-price, in place of the four series.
MA and PCA-tech are both technical indicators, and are 0.80 correlated. Interestingly, PCR
(ratio of stock price to commodity price) is highly correlated with the price ratios. Commodity
price in the denominator appears to serve a similar role as the fundamental variables in the
price ratios — dividends, earnings, or book value. VRP and IC appear to contain information
about credit markets, as they are 0.54 and 0.36 correlated with DEF. The technical indicators
MA and PCA-tech are negatively correlated with these variables.
9|Page
Table 3 shows correlations among the predictor variables and future market returns. The
correlation is one measure of how well each predictor variable would do in univariate
forecasting regressions. Two key observations are evident. First, predictor variables are related
to future returns in different ways. Some variables, including DP, BM, CAY, VRP, BDI, MA, and
PCA, have positive correlations with future market returns at all horizons. Other variables
including PE, CAPE, SIM, NOS, CPI, PCR, and SI, are negatively correlated with future market
returns at all horizons. Still, some other variables such as BY, DEF, TERM, IC, and OIL, may have
positive or negative correlation with future returns depending on the horizon. Second,
different variables forecast returns at different horizons. The slow-moving price ratios DP, PE,
BM, and CAPE all have stronger correlations at longer horizons. Other predictors such as CAY,
NOS, MA, PCA, OIL, and SI exhibit the same pattern. However, some predictors appear to work
better for shorter horizons and their forecasting power weakens at longer horizons: DEF, SIM,
10 | P a g e
Many previous papers on return predictability tend to focus on expected returns at the
business-cycle frequency — one to five years — and forecasting variables are designed to
capture this variation. For example, DP, PE, BM, and CAPE forecast one year returns more
strongly compared to one month returns. DEF and TERM are specifically chosen to coincide
with business cycle peaks and troughs and the equity premium variation associated with those.
Less focus has been put on short-term variables that attempt to capture expected return
variation in the next days or weeks, because short-term returns contain much more noise, and
reliable statistical evidence is harder to establish.
Information contained in inflation measures is also helpful in forecasting equity returns, but not
necessarily contained in the variables measuring the macroeconomy. We augment our
information set by including a transformation of CPI as a measure of inflation. Of course, the
real economy and inflation are not independent, so it is possible CPI may contain information
about the macroeconomy as well, and the macro variables contain information about inflation.
Variables that contain direct information about financial markets will enlarge the forecasting
information set. We include SIM, VRP, IC, PCR, MA, PCA, and SI to gauge the performance of
future returns from a different perspective than the macroeconomy. VRP and IC use
information in derivative markets, and MA and PCA-tech are technical indicators. PCR
incorporates information from commodity markets. SI examines investor behavior by looking
at how bearish they are. Finally, we include information about international trade with BDI.
11 | P a g e
We run simulations of the portfolio performance based on our market-timing model1. There is
potentially a look-ahead bias, since some variables were only discovered after the simulation
start date. Including those variables at the beginning of the simulation would assume prescient
knowledge of these return predictors. We repeat our analysis only including a return predictor
after its discovery, to alleviate any look-ahead bias.
In combining return predictors, we uncover better forecasting results, and larger economic
significance, compared to using return predictors individually. Of course, we are not the first to
combine return predictors. Rapach, Strauss, and Zhou (2010) combine individual return
forecasts, and find the combination delivers statistically and economically large out-of-sample
gains compared to the historical average. We use a larger set of return predictors that likely
cover a broader information set, and illustrate the large economic gains from timing the market.
We look for medium-term return forecasts. The forecast target is the upcoming 130-day
market return. We first determine the best transformation of each forecasting variable by
maximizing the correlation between the transformed variable and the forecasting target.
Transformations include the raw value, an exponentially-weighted moving average, and log
value minus its exponentially-weight moving average. Specific transformations are determined
by the maximal correlation, using previous published work as a guideline. Every 20 days
beginning June 2001, we use a training period of 10 years to estimate model coefficients, either
with fixed variable transformations or transformations that maximize correlations with 130-day
future returns subject to sign constraints (Campbell and Thompson 2008). For the next 20 days,
we calculate expected returns using the estimated coefficients, and take a position eight times
the expected equity premium. The parameters we use (20 days, 10 years, 130 days, and eight
times expected returns) are robust: We have tried other combinations that give us similar
results2.
12 | P a g e
1. Details on how to replicate the results of this paper are available at
http://www.hullinvest.com/HI/wp-content/uploads/2015/06/How-to-replicate-
A-Practitioners-Defense.pdf
2. Robustness resultsElectronic
are available
copyupon request.
available at: https://ssrn.com/abstract=2609814
Our first forecasting model is a simple kitchen sink regression, which includes all of the return
predictors in Section 2 except the price ratios, which we replace with PCA-price, for a total of 16
variables.
Where
We fit our model every 20 days to obtain parameter estimates, which we use with updated
return predictors each day for the following 20 days to produce expected equity premium
forecasts. We then take positions in SPY proportional to our return forecasts. This process is
repeated every 20 days. Figure 1 shows the variables selected for the kitchen sink model. All of
the variables are included by construction.
13 | P a g e
Figure 2 plots the wealth evolution of $1 invested in a market-timing strategy based on the
kitchen sink model, cash, or buy-and-hold SPY. At the end of our sample, kitchen sink and buy-
and-hold have similar cumulative returns. It is notable that through the two large market
downturns we experience during this period, in 2002 and 2008, the kitchen sink model would
have kept us from large drawdowns as the overall market experienced. In fact, during the large
downturns the kitchen-sink based strategy adjusts the position to be negative as the six-month
forecast implies medium-term future returns are likely to be low or negative.
The lower panel in Figure 2 displays the positions taken by the kitchen sink model. 100%
indicates a buy-and-hold strategy. As expected, the market exposure is generally positive, as
the market tends to go up on average. During the recent financial crisis there was an extended
period in which the position taken by the model was negative, indicating our model was able to
capture the falling market as it was happening. In our implementation, we do not adjust our
14 | P a g e
Figure 2. Wealth Accumulation and Positions of the Kitchen Sink Model. The top panel plots
the cumulative returns ($1 compounded) of the market timing strategy (black solid line) from
the kitchen sink model, buy and hold SPY (blue dotted line), and cash (green dashed line). The
bottom panel plots the changing positions of the strategy. The strategy is capped at 150% long
and 50% short SPY.
Overall, the market-timing model based on the kitchen sink model does not outperform the
buy-and-hold market in this period in terms of returns. This is probably because naively
dumping all of the variables into a linear model increases the likelihood of overfitting in sample,
15 | P a g e
Where
16 | P a g e
Correlation screening is a practical and simple way to narrow down variable with the highest
predictive power (Hero and Rajaratnam 2011). We use a threshold of 10% to select variables
that have the highest predictive power for future returns. In Appendix A, we examine other
threshold values. Figure 3 illustrates the variable selection for the correlation screening model.
Although all of these variables have had univariate forecasting power in the literature, when
they are included simultaneously not all variables provide the same forecasting power.
Variables such as CAY, VRP, DEF, PCA-price, and NOS are almost always selected, indicating a
high predictive power throughout the sample. Some variables including PCR, BDI, IC, and TERM
are picked up by the model intermittently. Some variables were strong predictors in the first
half of the sample, but were not selected in the second half (PCA-tech, MA, and SIM). A small
set of variables such as SI, BY, and OIL is almost never selected.
17 | P a g e
Figure 4 shows the cumulative wealth of $1 invested in the correlation screening model, cash,
or buy-and-hold SPY. The market-timing strategy based on the correlation screening model
outperforms the buy-and-hold strategy. The cumulative returns of correlation screening from
2001 to 2015 are more than twice that of the buy-and-hold strategy. We do not suffer large
negative returns in the two large market downturns in 2002 and 2008. The lower panel shows
the position in SPY of the correlation-screening strategy. Positions undergo large changes
through time, and are negative in 2002 and 2008 when the overall market had large negative
18 | P a g e
Most of these variables were discovered before our simulation start date of 06/08/2001, but
some were discovered afterwards. To guard against look-ahead bias, we repeat our simulation
using only variables that are known at the time, and add variables after they have been
discovered. BDI has been known at least since January 2011. NOS was discovered in December
2008 (private correspondence with Chris Jones). OIL was first mentioned as a predictor of
equity returns in 2005. PCR was found in late 2014. PCA-tech was first used as a return
predictor in 2010.
Figure 5. Variable Selection, Real-Time Correlation Screening. The top figure plots the number
of variables that are chosen in the model at any given time for the real-time correlation
screening model. The bottom figure shows which variables are selected.
19 | P a g e
Where
Figure 5 illustrates the variable selection for the real-time correlation screening model. BDI,
NOS, OIL, PCR, and PCA-tech were not known at the beginning of the sample so they could not
be selected. The number of selected variables in the first half of the sample is lower compared
to the correlation screening model, because these five variables were mechanically left out. In
the second half of the sample, the number of variables selected and the identity of the selected
variables are virtually identical in the correlation screening model and the real time model.
20 | P a g e
Figure 6 shows the wealth accumulation of the real-time correlation screening model and its
positions in SPY. The wealth accumulation process for the real-time correlation screening
model is highly similar to that of the correlation screening model (Figure 4), except the
positions taken prior to 2005 for the real-time model are more conservative compared to the
correlation screening model. This indicates the look-ahead bias for the correlation screening
model is small.
Figure 7. Actual Versus Forecast Returns, 06/08/2001-05/04/2015. Actual returns are on the
vertical axis, and forecast returns are on the horizontal axis. The left panel is the kitchen sink
model; the right panel is the correlation screening model. The black line is the 45-degree line
that the data points would lie if forecast returns exactly coincided with realized returns. The
dashed green line is the best line fit of actual returns on forecast returns.
21 | P a g e
In contrast, expected return forecasts from the correlation screening model does capture
considerable variation in actual returns, shown on the right panel in Figure 7. We see forecast
returns are positively correlated with actual returns, and the dashed green line has a large
positive slope. The green line is much closer to the black line compared to the left panel,
indicating expected returns from this model are much closer to a perfect forecast compared to
the kitchen sink mode. It is evident the correlation screening model is able to pick up important
information about future returns. Data points tend to bunch up in the middle because actual
returns are more volatile compared to forecast returns.
4. Implementation Details
We invest in two assets, SPY and cash. Every day, we take positions in SPY based on our
expected market return forecasts. Every 20 days, we refit our model and keep its parameters
constant for the next 20 days, at which time the procedure is repeated. Each day at 3:55 pm
EST, we download the data and use our fitted model to produce a forecast of expected returns.
Our desired position is proportional to the expected return forecast, with a cap of 150% (long)
and -50% (short) SPY. Orders are submitted to the closing auction at NYSE Arca. Orders for the
closing auction must be submitted by 3:59 pm EST. As a result of using the closing auction, we
receive the settlement price. Our assumption is that we are able to get the closing price at
4:00pm EST, and can execute our trade on market close.
Although we could use a number of equity indexes to calculate market excess returns, we
implement our strategies with SPY because the S&P 500 market including futures contracts is
the most liquid equity market in the world. In 2014 the SPX futures traded $145 billion and SPY
traded $21 billion per day. The closing auction at NYSE Arca alone averaged more than $422
million per day. At such depth, it is unlikely that slippage will significantly degrade the returns
from our strategies.
Table 4 shows the performance of the three market-timing strategies we consider, along with
the buy-and-hold strategy. The correlation screening model, our benchmark, yields an annual
return of 12.11% from 2001 to 2015, compared to 5.79% for the buy-and-hold SPY. The Sharpe
22 | P a g e
The real-time correlation screening model, which only adds variables as they are discovered in
the literature, performs almost equally well: 11.66% annual returns, 0.88 Sharpe, and a slightly
smaller drawdown. The similarity in performance between the correlation screening and real-
time correlation screening model provides evidence that the look-ahead bias in our correlation
screening model was small to start with. The kitchen sink model, which naively includes all of
the forecasters, has returns similar to that of the buy-and-hold strategy, but a Sharpe ratio
about twice as high, and a smaller drawdown. It is clear that the market-timing strategies give
superior performance compared to the buy-and-hold strategy.
KS CS RTCS SPY
Return 5.89% 12.11% 11.66% 5.79%
Sharpe Ratio 0.41 0.85 0.88 0.21
Max Drawdown 26.44% 21.12% 21.83% 55.20%
Table 5 presents the annual returns from each of the market-timing strategies as well as for the
buy-and-hold SPY. With large positive and negative annual returns, the buy-and-hold strategy
is more volatile compared to any of the market-timing strategies. The correlation screening
strategy underperforms the market in seven different years, and the real-time correlation
screening strategy also underperforms the market in six separate years. It would not be wise to
compare their short-term performance against the buy-and-hold strategy and conclude the
market-timing strategies are of no value. Even though these strategies underperform buy-and-
hold at times, they have higher compound returns and Sharpe ratios. A disciplined investor
should not lose faith in these strategies when they are temporarily underperforming the market.
23 | P a g e
KS CS RTCS SPY
2001 3.49% 1.75% 4.45% -8.47%
2002 -1.08% 3.72% 16.30% -21.59%
2003 19.80% 9.16% -1.43% 28.19%
2004 5.51% 5.91% 0.61% 10.70%
2005 -1.76% 2.13% -0.22% 4.83%
2006 2.79% 7.44% 4.40% 15.85%
2007 0.87% 8.53% 2.85% 5.15%
2008 16.14% 18.96% 23.85% -36.69%
2009 -12.67% 40.32% 40.82% 26.36%
2010 -2.23% 2.21% 3.76% 15.06%
2011 3.37% 7.69% 7.99% 1.90%
2012 3.04% 15.47% 15.47% 15.99%
2013 37.54% 34.79% 34.79% 32.31%
2014 13.38% 14.64% 14.64% 13.47%
2015 1.82% 2.45% 2.45% 2.85%
One issue that we have to address while in implementing market-timing strategies is taxes.
Past work on return predictability often does not consider the effect of taxes if one were to
implement a market-timing strategy. In practice, most smart beta and tactical asset allocation
(TAA) products are able to create some alpha, but the outperformance is often eroded by taxes.
Our market-timing strategy suffers from the same problem, and we recommend this strategy
only be used by retirement accounts or foundations. We try to keep our transactions to a
minimum, and do not make small adjustments to our portfolio positions. We also consider
transactions costs, and assume we pay two cents per share to buy or sell SPY. We assume cash
earns daily interest at the three-month TBill rate minus 30 basis points. We also assume that
we pay interest on the shares borrowed at the Fed Funds rate plus 30 basis points.
24 | P a g e
There is a lot of noise in return forecasts. Aside from data issues, information that may impact
expected returns arrives at irregular frequencies. One needs to continuously monitor a large
number of factors that may or may not provide information about future returns. A forecasting
variable that has been proposed in the past may have worked for some specific time periods,
but not for other periods. It is inherently difficult to assess if that means the result was
spurious and the variable does not have any forecasting power, or the result was genuine but
the data just had a bad run and the variable may work again in the future.
Investors who wish to market-time must maintain strict discipline and keep emotions out of the
investment process. The game is to optimally find the right mix of indicators, appropriately
assess them, and trade immediately when an opportunity presents itself. A traditional
investment committee may meet on the third Thursday following the end of a quarter. Such a
structure would be much too slow to react to the much faster pace of market-timing. One
needs to continually track the market, and effectively execute on the tiny signals that
sometimes present themselves in a sea of noise. Few retail investors or even professionals
have the discipline to act continuously in an unbiased manner. A market-timing strategy
requires a contrarian spirit — selling in hot booms and buying in market downturns.
Furthermore, the strategy may not always work, and one must maintain complete faith and
continue to trade even if it is currently losing money. The uncertainty may partially explain why
so few institutions have tried to build this type of system.
There are still some miscellaneous difficulties for investors who want to carry out a market-
timing strategy. Many money managers have to worry about what their investors think, and
naturally place more focus on the near-term. If the market-timing strategy fails to work for a
period of time, these delegated managers may just abandon the strategy as they find it
increasingly burdensome to explain the poor results to investors. Another problem is that
investors may become very risk-averse at the exact time that the market-timing strategy is the
most valuable. The last quarter of 2008 and the first quarter of 2009 were great times to
increase equity market exposure, but many funds liquidated instead because they did not want
to be invested in risky assets in those periods. The high cost of information acquisition and
hiring staff to perform the necessary analysis may be yet another reason why market-timing is
not more common.
25 | P a g e
5. Economic Significance
The acid test for return predictability is if investors can make a profit through timing the market.
We have shown it is indeed possible to construct profitable strategies based on market return
forecasts. But how well are we doing? Returns of 200-300 basis points above the market over
a long period of time are exceptional. We have shown that it was possible to gain more than
that during the fourteen year period from 2001 to 2015. The historical Sharpe ratio during our
test period is 0.21, while the long-term Sharpe ratio from 1926-2015 is around 0.4. Our market-
timing strategy produces a Sharpe ratio of 0.85, so it would seem investors can significantly
time the market. We need a more precise way of measuring our performance though.
We examine the maximum potential to time the market through estimating the theoretical
Sharpe ratio from a six-month forecast. Grinold and Kahn (2000), in their book “Active Portfolio
Management”, provide one way to evaluate that question. They provide a calculation for the
maximum possible information ratio, assuming investors know with certainty stock returns over
the next six months and they trade to maximize their wealth. In our case, the maximum
information ratio is 1.59. Since we benchmark against cash, this is also the theoretical
maximum Sharpe ratio.
Another way to compute the maximum possible Sharpe ratio of a strategy is to assume perfect
knowledge of future returns (private correspondence with Rick Anderson 3). Anderson uses
daily data on the Dow Jones Industrial Average from 1926 to 1996 and assumes investors have
perfect information. Each day he calculates the return in the next 252 days. He takes a long
position if the return is positive and a short position if the return is negative. He finds the
Sharpe ratio of such a strategy is 1.5. We repeat Anderson’s exercise for the CRSP value-
weighted index with a look-forward period of 130 days. For 2011-2014, the maximum Sharpe
ratio is 1.15.
Considering the maximum possible Sharpe ratios range from 1.15 to 1.59, our market-timing
strategy has a Sharpe ratio that’s about two-thirds of the theoretical maximum. Our strategy is
capturing a significant amount of the time-varying expected returns that can possibly be
captured even with perfect knowledge about future returns. Although we have not employed
26 | P a g e
3. Rick Anderson is the Author of Market Timing Models and is the Chief
Investment Officer of Hull Investments LLC.
Although our simulation worked well, it is important to recognize that two significant
downturns in 2002 and 2008 contributed to the outperformance of the market-timing
strategies over the buy-and-hold strategy. In fact, those two events were two of the three
largest cumulative negative returns in the last 100 years (the third being the Great Depression).
Our market-timing strategies are designed to outperform in periods of persistently low returns,
as we adjust our positions to changing conditions while the market continues to underperform.
The rate of outperformance in our simulation must be interpreted with caution as downturns of
the magnitude observed in 2002 and 2008 usually do not occur only six years apart.
Another caveat in interpreting our results is an inherent publication bias in academic finance.
The publication process favors positive results over negative results, so variables showing
predictive power are more likely to be published. The proposed predictors may work well
precisely because they have been published — many poor predictors may have been tested and
never made public. This data-dredging concern is difficult to address, as we do not observe
how many other variables have been tried before we found these twenty. Only time will tell if
our market-timing models truly outperform: New data is the best out-of-sample test.
6. Concluding Remarks
In this paper, we revisit return predictability. We examine twenty prominent return predictors
proposed in the literature, and combine them using correlation screening. We find that we can
forecast market returns six month into the future. A market-timing strategy taking positions in
SPY proportional to our model estimates outperforms buy-and-hold SPY returns. We illustrate
the economic significance of return predictability by simulating a market-timing strategy that
makes large economic profits. Furthermore, we discuss the execution details of our strategy,
emphasizing various implementation difficulties.
We have only addressed return predictability in a limited setting. By focusing on a set of return
predictors previously proposed, and a return forecast horizon of roughly six months, we have
shrunk the large universe of potential return-forecasting models to a much smaller one. An
interesting extension would be to examine return predictability at alternative forecast horizons,
especially at one-year and five-years for which many of these predictors were first proposed.
Such an exercise will readily illustrate the importance of combining information in different
return predictors. Another interesting extension is to examine alternate methods of combining
forecasting variables. We have used correlation screening. Other potential methods including
27 | P a g e
Throughout this paper, we have included what we consider the most prominent return
predictors. With the twenty variables we do use, we uncover economically significant
outperformance compared to the buy-and-hold market. Clearly, there may be other variables
with strong predictive power that we have not covered. In fact, we have discovered some
proprietary predictors for the equity risk premium that perform well in-sample and out-of-
sample. The forecasting performance is even better if we include our proprietary predictors.
If an investor has the ability to reliably forecast market excess returns, then having a constant
exposure to different asset classes surely is suboptimal. The investor should increase his
exposure to equities when its expected returns are high, and decrease his exposure when the
expected returns are low. Such practice has been termed tactical asset allocation (TAA). TAA
has become pervasive in industry practices, and the academic community is growing
increasingly more interested (see Campbell and Viceira 2002). Moreover, retail investors
already engage heavily in TAA — they increase their exposure after a bull run, and sell stocks in
the middle of a financial crisis. However, this is exactly the opposite of what investors should
do. Daniel and Moskowitz (2013) have shown that a trend-following strategy could suffer large
infrequent losses.
The current investor behavior is not only harmful to investors themselves, but also subjects the
market as a whole to more risk than necessary. Trend-following behavior is destabilizing to the
market, since investors’ buys and sells push the market further towards its extremes. When
investors buy when the market is up, the price is bid up further and more investors buy. This
causes large swings in price. On the other hand, contrarian behavior may be stabilizing in the
sense that a contrarian investor would dampen swings in price since he sells when prices are
high and buys when prices are low. TAA with market-timing is in the spirit of a contrarian
strategy. When recent returns have been high, expected returns are likely to be low and a
market-timing strategy would sell instead of buy. If a market-timing product that automatically
buys when the market is low and expected returns are high were widely available for investors,
its effect would be a stabilizing one for the market. All else equal, market volatility would be
lower, and price swings would not be as large.
As our understanding of return predictability changes, so will the stigma associated with
market-timing strategies. Anybody who claimed to implement a “market-timing” strategy in
the past 30 years would have been considered irresponsible; as such a strategy was thought to
underperform the buy-and-hold strategy. In the upcoming 30 years, it is likely that it will be
considered irresponsible to not engage in informed market-timing. Investors should change
28 | P a g e
29 | P a g e
Bakshi, G., Panayotov, G., and Skoulakis, G. (2011). “The Baltic Dry Index as a Predictor of Global
Stock Returns, Commodity Returns, and Global Economic Activity”, AFA 2012 Chicago Meeting
Paper.
Berk, J. B. and Green, R. C. (2004). “Mutual Fund Flows and Performance in Rational Markets”,
Journal of Political Economy, 112, 1269-1295.
Black, A. J., Klinkowska, O., McMillan, D. G., and McMillan, F. J. (2014). “Forecasting Stock
Returns: Do Commodity Prices Help?”, Available at SSRN 2467095.
Bollerslev, T., Tauchen, G., and Zhou, H. (2009). “Expected Stock Returns and Variance Risk
Premia”, Review of Financial Studies, 22, 4463-4492.
Bouman, S. and Jacobsen, B. (2002). “The Halloween Indicator, ‘Sell in May and Go Away’:
Another Puzzle”, American Economic Review, 92, 1618-1635.
Campbell, J. Y. and Shiller, R. J. (1988a). “The Dividend-Price Ratio and Expectations of Future
Dividends and Discount Factors”, Review of Financial Studies, 1, 195-228.
Campbell, J. Y. and Shiller, R. J. (1988b). “Stock Prices, Earnings, and Expected Dividends”,
Journal of Finance, 43, 661-676.
Campbell, J. Y. and Viceira, L. M. (2002). Strategic Asset Allocation: Portfolio Choice for Long-
Term Investors, Oxford University Press.
Campbell, J. Y. and Vuolteenaho, T. (2004). “Inflation Illusion and Stock Prices”, No. w10263.
National Bureau of Economic Research.
Campbell, J. Y. and Thompson, S. B. (2008). “Predicting Excess Stock Returns Out of Sample: Can
Anything Beat the Historical Average?”, Review of Financial Studies, 21, 2008.
Carhart, M. M. (1997). “On Persistence in Mutual Fund Performance”, Journal of Finance, 52,
57-82.
30 | P a g e
Center for Retirement Investing. (2013, September 11). Burton Malkiel: Market Timing Is
Dangerous [www.youtube.com/watch?v=unFn-ddtqiM]. Retrieved from http://
www.youtube.com/watch?v=unFn-ddtqiM
Cochrane, J. H. (2008). “The Dog that Did Not Bark: A Defense of Return Predictability”, Review
of Financial Studies, 21, 1533-1575.
Driessen, J., Maenhout, P. J., and Vilkov, G. (2013). “Option-Implied Correlations and the Price
of Correlation Risk”, Available at SSRN 2359380.
Efron, B., Hastie, T., Johnstone, I., and Tibshirani, R. (2004). “Least Angle Regression”, Annals of
Statistics, 32, 407-499.
Epstein, R. A. (1995). The Theory of Gambling and Statistical Logic, Revised Edition, Academic
Press.
Fama, E. F. and French K. R. (1988). “Dividend Yields and Expected Stock Returns”, Journal of
Financial Economics, 22, 3-25.
Fama, E. F. and French, K. R. (1989). “Business Conditions and Expected Returns on Stocks and
Bonds”, Journal of Financial Economics, 25, 23-49.
Friedman, J., Hastie, T., and Tibshirani, R. (2010). “Regularization Paths for Generalized Linear
Models via Coordinate Descent”, Journal of Statistical Software, 33, 1-22.
Goyal, A. and Welch, I. (2008). “A Comprehensive Look at the Empirical Performance of Equity
Premium Predictions”, Review of Financial Studies, 21, 1455-1508.
Graham, B. and Dodd, D. L. (1934). Security Analysis: Principles and Technique, McGraw-Hill,
1934.
31 | P a g e
Jones, C. S. and Tuzel, S. (2012). “New Orders and Asset Prices”, Review of Financial Studies, 26,
115-157.
Lettau, M. and Ludvigson, S. (2001). “Consumption, Aggregate Wealth, and Expected Stock
Returns”, Journal of Finance, 56, 815-849.
Merton, R. C. (1980). “On Estimating the Expected Return on the Market: An Exploratory
Investigation”, Journal of Financial Economics, 8, 323-261.
Moskowitz, T. J., Ooi, Y. H., and Pedersen, L. H. (2012). “Time Series Momentum”, Journal of
Financial Economics, 104, 228-250.
Neely, C. J., Rapach, D. E., Tu, J., and Zhou, G. (2014). “Forecasting the Equity Risk Premium: The
Role of Technical Indicators”, Management Science, 60, 1772-1791.
Pastor, L. and Stambaugh, R. F. (2009). “Predictive Systems: Living with Imperfect Predictors”,
Journal of Finance, 64, 1583-1628
Rapach, D., Ringgenberg, M., and Zhou, G. (2015). “Short Interest and Aggregate Stock Returns”.
Rapach, D. E., Strauss, J. K., and Zhou, G. (2010). “Out-of-Sample Equity Premium Prediction:
Combination Forecasts and Links to the Real Economy”. Review of Financial Studies, 23, 821-
862.
Warther, V. A. (1995). “Aggregate Mutual Fund Flows and Security Returns”, Journal of
Financial Economics, 39, 209-235.
Yang, D. and Zhang, Q. (2000). “Drift-Independent Volatility Estimation Based on High, Low,
Open, and Close Prices”, Journal of Business, 73, 477-492.
Ziemba, R. E. and Ziemba, W. T. (2013). Investing in the Modern Age, World Scientific Books.
32 | P a g e
33 | P a g e
34 | P a g e
Figure B1. Rounded Positions of Kitchen Sink, Correlation Screening, and Real-Time
Correlation Screening Models. The top figure plots the rounded positions for the kitchen sink
35 | P a g e
36 | P a g e