Multi AssetSeasonalityandTrend FollowingStrategies
Multi AssetSeasonalityandTrend FollowingStrategies
Multi AssetSeasonalityandTrend FollowingStrategies
net/publication/291166794
CITATIONS READS
2 2,199
1 author:
Nick Baltas
Goldman Sachs
15 PUBLICATIONS 154 CITATIONS
SEE PROFILE
All content following this page was uploaded by Nick Baltas on 19 January 2016.
Forthcoming at Special Issue on Hedge Funds for Bankers Markets & Investors, 2016
Abstract
This paper investigates the seasonality patterns within various asset classes. We find
that a strategy that buys the assets with the largest same-calendar-month past average
returns (up to ten years) and sells the assets with the smallest same-calendar-month
past average returns, earns statistically and economically significant premia within
commodity and equity index universes. Capitalising these premia directly appears
practically difficult, due to the high strategy turnover and associated costs. We
therefore suggest a way to actively incorporate seasonality signals into a trend-
following strategy by switching off long and short positions, when the respective
seasonality signals argue otherwise. The seasonality-adjusted trend-following strategy
constitutes a significant improvement to the raw strategy across both commodities and
equity indices. The increased turnover can impact the performance pickup, but the
relatively low trading costs of liquid futures contracts as well as methodological
amendments that optimise position smoothing can render the improvement genuine.
1
The opinions and statements expressed in this paper are those of the author and are not necessarily the opinions of any
other person, including UBS AG and its affiliates. UBS AG and its affiliates accept no liability whatsoever for any
statements or opinions contained in this paper, or for the consequences which may result from any person relying on such
opinions or statements.
2
Nick Baltas is an Executive Director in the Quantitative Research Group of UBS Investment Bank and a visiting
lecturer at Queen Mary University of London and Imperial College Business School.
1
1. Introduction
Return cyclicality is relatively common within equity returns and the literature on
seasonality effects is rather rich. From the famous January effect to the Hallowe'en
effect and the lunar cycle, academic studies have documented a multitude of seasonal
patterns within equity markets and equity strategy returns. In explaining these
patterns, researchers have associated them either to various market-driven incentives,
like tax year-end activities and institutional window dressing, or to behavioural effects
due to seasonal risk appetite (e.g. investors acquiring a more gambling behaviour in
the beginning of the year). What is very common across all these patterns is that it is
typically rather expensive to exploit them, and as a consequence, whatever appears to
be significantly strong on paper might be impossible to capitalise in practice. Ilmanen
(2012, Chapters 25-26) provides a very comprehensive overview of the broad family
of seasonal effects and their attempted explanations.
More recently Keloharju, Linnainmaa and Nyberg (2015) extend the findings of
Heston and Sadka (2008, 2010) and explore the long-run seasonal patterns within
equity anomalies (they focus on 15 strategies based on value, momentum, gross
profitability etc.). Interestingly, even though their primary focus is on equity
anomalies, they also include a short section where they test for similar seasonality
patterns within commodities and international stock market indices.
Motivated by these recent findings, we investigate in detail the existence of any long-
run seasonal patterns across various non-equity asset classes, namely commodities,
government bonds, FX rates and country equity indices. By using a panel regression
framework we document strong statistical evidence of long-run (up to ten years due to
limited data availability) seasonal patterns within commodities and equity indices, but
nothing of particular significance within government bonds or FX rates 4. These
findings are in line with Keloharju et al. (2015).
Commodity markets are traditionally driven by supply and demand shocks between
producers and consumers and the seasonal variation of these shocks can potentially be
the reason of the seasonal return regularities. For instance, heating energy commodity
3
In particular, Heston and Sadka (2008) show that the seasonal patterns remain robust across different size or industry
bands, and are not concentrated on any particular calendar month of corporate event (like earnings announcement
months, dividend announcement months, ex-dividend months or fiscal year-end months).
4
FX rates exhibit strong seasonal patterns but at a higher frequency and mostly intra-daily; see for example Andersen and
Bollerslev (1997) and Ito and Hashimoto (2006).
2
prices are expected to predictably fluctuate between seasons (winter and summer);
indicatively, Todorova (2004) studies the seasonal patterns on crude oil and natural
gas futures. Similarly, agricultural commodity prices are expected to predictably
fluctuate around (before and after) the harvesting period; indicatively, Sørensen
(2002) studies the seasonal patterns on corn, soybean and wheat futures.
In order to capture the seasonality premia across commodity and equity index
markets, we construct long-short portfolios that buy assets with the largest same-
calendar-month average returns over the past ten years and short assets with the
lowest same-calendar-month average returns over the same period. The strategy
returns are statistically and economically strong, and fairly uncorrelated with other
investment strategies like value or momentum. However, in line with expectations,
capitalising on these premia can be very costly due to the high turnover that the
respective strategies exhibit. At times, the trading costs can even completely eliminate
the positive returns, hence casting doubts on the practical importance of our findings.
The paper is organised in three main parts. Section 2 contains the empirical analysis
for the identification of seasonality patterns within the various asset classes. Section 3
provides insight upon the existence of the seasonality premia within commodities and
equity indices by looking into risk-based or limits-to-arbitrage explanations. Section 4
presents the construction of a typical trend-following portfolio and explores the added
value from incorporating seasonality signals. Finally, Section 5 concludes.
3
2. Exploring Seasonality Patterns
Testing for the existence of seasonality patterns in the cross-section of asset returns
implies testing whether lagged returns over a certain frequency (e.g. lagged annual
returns, lagged semi-annual returns etc.) can forecast future returns. Instead of
running univariate regressions of asset returns on their lagged values on an asset-by-
asset basis, we can increase the statistical power of the experiment by running panel
Fama and MacBeth (1973) regressions. This involves running, first, a cross-sectional
regression of the returns of all assets on their respective 𝑘𝑘-month lagged monthly
return at the end of each month 𝑡𝑡:
𝑟𝑟𝑖𝑖,𝑡𝑡 = 𝛼𝛼𝑘𝑘,𝑡𝑡 + 𝛽𝛽𝑘𝑘,𝑡𝑡 ⋅ 𝑟𝑟𝑖𝑖,𝑡𝑡−𝑘𝑘 + 𝜖𝜖𝑖𝑖,𝑡𝑡 , for 𝑖𝑖 = 1, … , 𝑁𝑁𝑘𝑘,𝑡𝑡 and each 𝑡𝑡 (1)
and subsequently average across time the regression coefficients 𝛼𝛼𝑘𝑘,𝑡𝑡 and 𝛽𝛽𝑘𝑘,𝑡𝑡 to
estimate the values of 𝛼𝛼𝑘𝑘 and 𝛽𝛽𝑘𝑘 :
1 1
𝛼𝛼𝑘𝑘 = 𝑇𝑇 ∑𝑇𝑇𝑡𝑡=1 𝛼𝛼𝑘𝑘,𝑡𝑡 and 𝛽𝛽𝑘𝑘 = 𝑇𝑇 ∑𝑇𝑇𝑡𝑡=1 𝛽𝛽𝑘𝑘,𝑡𝑡 . (2)
In equation (1) above, 𝑟𝑟𝑖𝑖,𝑡𝑡 denotes the return of asset 𝑖𝑖 in month 𝑡𝑡 and 𝑁𝑁𝑘𝑘,𝑡𝑡 denotes the
number of assets at the end of month 𝑡𝑡 that have data up to 𝑘𝑘 months ago.
Heston and Sadka (2008, 2010) and Keloharju et al. (2015) run the panel regression
for lags up to 240 months (20 years). Our significantly smaller multi-asset data
sample does not allow for strong statistical power for such long lags and we therefore
run the panel regression for lags up to 60 months 5 and separately for each asset class
as the seasonal patterns can vary significantly from one asset class to another. This
choice for maximum lag implies that our first cross-sectional regression will take
place five years after the first month of data in each asset class. As already explained
in the previous section, not all assets and asset classes start at the same point in time.
Hence, in an effort to have a reasonable balance between cross-sectional and time-
series data availability, we conduct the panel regression analysis starting in January
1975 for equity indices (all 19 assets available), in July 1991 for commodities (14 out
of the 18 assets available), in July 1994 for government bonds (13 out of the 16 assets
available) and in January 1991 for foreign exchange (all ten currencies available).
5
Moskowitz, Ooi and Pedersen (2012) as well as Baltas and Kosowski (2013, 2015), who run similar panel regressions
on a multi-asset space also use lags up to 60 months.
4
Table 1: Dataset
Figure 1 reports the slope coefficients 𝛽𝛽𝑘𝑘 for all lags (annual lags are coloured in
black) and asset classes and Figure 2 reports the respective t-statistics, which are
calculated using Newey and West (1987) standard errors with 12 lags in order to
adjust for potential serial-correlation and heteroskedasticity patterns in the error
terms. These two Figures should be studied side by side.
What we look for is a combination of two things. On one hand, we look for a large
numerical value for 𝛽𝛽𝑘𝑘 (in Figure 1). As Fama (1976, Chapter 9) explains, these slope
coefficients can be interpreted as the return of a long-short, zero-cost portfolio of all
the assets that are used in the panel regression in equation (1). On the other hand, we
look for a large t-statistic for 𝛽𝛽𝑘𝑘 (in Figure 2). Critical values of ±1.96 (5%
significance) are denoted with dashed lines in the figure. A more conservative
threshold of significance is that of ±1.64 (10% significance). Several interesting
patterns emerge from these Figures, which we discuss separately for each asset class.
5
Figure 1: Panel Regression Coefficients Figure 2: Panel Regression t-statistics
Notes: The charts report the Fama-MacBeth panel regression coefficients from Notes: The charts present the Fama-MacBeth t-statistics of the panel regression
regressing monthly returns on lagged returns, for lags ranging from 1 to 60 months. coefficients from regressing monthly returns on lagged returns, for lags ranging from 1
The sample period differs for each asset class and is mentioned in the chart titles. All to 60 months. The t-statistics are adjusted for serial-correlation and heteroskedasticity
data is obtained from Datastream. using the Newey and West (1987) correction with 12 lags. The dashed lines at -1.96
and +1.96 denote the 5% significance thresholds. All data is obtained from Datastream.
6
Equity Indices: We document strong serial-correlation effects over the course of the
first 12 months, which is a clear indication of trend-following (time-series
momentum) patterns. The return predictability becomes economically (judging by the
level of 𝛽𝛽𝑘𝑘 ) and statistically (judging by the level of the t-statistic of 𝛽𝛽𝑘𝑘 ) strong in
quarterly lags (3, 6 and 9 months). The first-year momentum patterns revert straight
after the course of the year in line with the findings of Moskowitz, Ooi and Pedersen
(2012). Regarding seasonality, which is the main focus of this paper, we document
strong, both statistically and economically, positive return predictability at a lag of 36
months and less significantly so at lags of 12 and 60 months.
Commodities: Possibly expected due to the nature of this asset class, commodities
appear to exhibit the strongest seasonal patterns across all asset classes. At annual lags
of 36, 48 and 60 months (even up to 120 months as it is shown in the Appendix A),
there is economically strong return predictability, as deduced by the levels of the
regression coefficients 𝛽𝛽𝑘𝑘 , with the 60-month lagged return also enjoying very strong
statistical significance with a t-statistic of 2.5. We also document very strong end-of-
first-year trend-following patterns, followed by strong reversals after the course of the
year, in line with the findings of Moskowitz et al. (2012).
Government Bonds: The fixed income market exhibits strong times-series return
predictability over the first 12 months, which constitutes evidence of the profitability
of trend-following strategies within fixed income markets over the recent decades; see
Baltas, Jessop, Jones and Zhang (2013). However, following the first year, there is
hardly any significant pattern worth documenting. No seasonal patterns emerge.
Foreign Exchange: We fail to document any statistically or even economically
strong patterns across any subperiod or specific lag within the FX universe, at least
for the given dataset and the respective sample period. A statistically strong reversal
effect at a lag of 36 months constitutes, to our view, more of an artefact of data fitting
and less so of any practical use.
In summary, we document significant seasonality patterns within commodity indices
and –less significantly so– within equity indices. These results are in line with
Keloharju et al. (2015), who similarly document some significant seasonality patterns
in commodity futures and country equity indices, but do not present any analysis
within the fixed income and foreign exchange markets. Based on this analysis and for
the remainder of this paper, we solely focus on commodity and equity index markets.
For additional robustness and after acknowledging any data limitation issues, we
present in Appendix A the results of our panel regression framework for lags up to
120 months (ten years) for equity index and commodity markets (Figures A.1 and A.2
respectively). The evidence of long-term seasonality for commodities is
overwhelming with 𝛽𝛽𝑘𝑘 for all annual lags between three and ten years (excluding the
lag of nine years) being the largest across the respective years and statistically
significant for lags above five years. In contrast, the evidence for equity indices is not
so statistically strong; all annual lags between five and ten years are all positive, but
lack statistical significance. The strongest effect remains at a lag of three years as
already documented in Figures 1 and 2.
7
3. Capturing the Seasonality Premium
In order to evaluate the investment implications of the documented patterns within
commodities and equity indices, we next construct long-short portfolios between past
same-calendar-month winners and past same-calendar-month losers. The analysis is
conducted separately for each of the two asset classes of interest.
In order to construct these portfolios, we first rank all assets of the same asset class at
the end of each month based on their average past same-calendar-month return for a
lookback period of ten years (we require at least five years of data for an asset in
order to be included in the ranking). As an example, at the end of December 2000, we
are expected to form a portfolio that will be held over January 2001 and for that
reason we rank the available assets based on their average return over the past ten
Januaries (1991 to 2000). Based on this ranking we then split the universe of assets
into quintiles and we construct a long-short portfolio by taking a long position on the
top quintile and a short position on the bottom quintile, equally weighting the
respective constituents. Using the terminology of Heston and Sadka (2008, 2010), we
call this portfolio the "annual" strategy.
Before proceeding with the presentation of the results, we have to note that for the
construction of annual and non-annual strategies we switch to a futures dataset for
commodities. So far, for the purposes of the panel regression analysis in the previous
section we have used spot commodity price index data mainly due to the longer data
availability (compared, say, to a futures universe). Estimating regressions with lags up
to five or ten years requires a relatively long history and cross-section of assets, hence
our decision to use spot return data. More importantly, a regression captures serial
dependence of returns and does not focus on the absolute level of asset returns. 6
Other than the practical implications of physical versus cash settlement, commodity
spot return data typically differ significantly from futures returns due to the level of
the cost of carry in the calculation of futures prices. The nature of commodity markets
(ranging widely from energy, softs, metals etc.) suggests that commodities constitute
an asset class with relatively large (if not the largest across all asset classes) costs of
6
Given asset pricing principles (and skipping any further details at this stage), spot price returns and returns of the first-
to-expire (and therefore most liquid) futures contract should be largely correlated. However, large correlation does not
imply anything about the underlying levels of returns, which can be significantly different, if the cost of carry in the
calculation of the futures price is large. As an example, consider how different the spot and futures returns are in a market
of an asset that exhibits strong contango or backwardation behaviour for a long period.
8
carry, which includes storage costs and convenience yields. Our expectation (which is
confirmed in unreported results) is that spot returns and futures returns for
commodities differ largely in practice. For this reason, in order to construct
seasonality strategies that can approximate closely the returns that an investor would
experience, we use daily commodity front futures price data, collected from
Bloomberg. 7 Table B.1 in the Appendix compares the spot and futures dataset in
terms of data availability; clearly, the only downside from using futures data is that
the sample periods of the contracts are generally smaller compared to the spot data.
The construction of the long-short seasonality portfolios, as outlined above, uses one
fifth of assets on the long side and another fifth of assets on the short side, whilst
requiring at least five years of data for each asset in order to be included in the cross-
sectional rankings. Given these restrictions and the data availability, we form
commodity portfolios from January 1992 onwards, at which point the portfolio has
two assets long and two assets short; from August 1994 the portfolio contains three
assets long and three assets short, in line with Keloharju et al. (2015), who construct
similar seasonality strategies within commodity futures markets.
Given our spot equity index dataset, all 18 assets are available from January 1970 and
we can therefore construct long-short seasonality portfolios starting from January
1980 that would have three assets long and three assets short.
7 In particular, we use generic continuous-price series provided by Bloomberg, which are constructed in a way that we
always trade the most liquid contract (typically the "front" contract), and we calculate for each futures contract fully-
collateralised monthly returns in excess of the prevailing risk-free rate as shown in Baltas et al. (2013) and Baltas and
Kosowski (2015).
8
de Roon, Nijman and Veld (2000) and Moskowitz et al. (2012) document that equity index returns calculated using spot
price series or front futures series are highly correlated.
9
3.1. Performance Evaluation
We next present our main empirical findings regarding the performance of long-short
strategies that aim to capitalise on the documented seasonal patterns within
commodity and equity index markets. Figure 3 presents the cumulative returns for the
annual and the non-annual portfolios for both asset classes. In order to have the same
point of reference between the two asset classes, January 1992 has been set equal to
100 for both of them (this is denoted by the vertical dashed line in the equities chart).
Finally, in order to allow for visual comparison between the two strategies (annual
and non-annual), they have both been dynamically adjusted at end of every month in
order to target a prescribed level of volatility equal to 7%. 9
Figure 3: Seasonality Patterns in Commodities and Equity Indices (Top vs. Bottom Quintile with volatility target at 7%)
Commodities Equity Indices
250 250
Annual Annual
nonAnnual nonAnnual
200 200
150 150
100 100
50 50
0 0
Notes: The figure presents the cumulative returns of long-short seasonality portfolios (top versus bottom quintile) within commodity and equity index markets. The annual strategy
ranks all available assets by the average same-month return over the past ten years (minimum five years of data are required for each asset). The non-annual strategy ranks the
available assets by their past ten year average return (minimum five years of data are again required) after excluding for the same-month returns. All strategies in this figure employ
a 7% volatility target. The rebalancing takes place at the end of each month and the assets of all top and bottom quintiles are equally weighted. The sample period is January 1992
to February 2015 for commodities and January 1980 to February 2015 for equity indices. The commodity dataset is obtained from Bloomberg, whereas the equity index dataset is
obtained from Datastream.
Focusing first on the annual strategy, it is evident that assets with the largest same-
calendar-month past returns outperform their peers with the lowest same-calendar-
month past returns, both within commodities and equity indices. Setting the annual
strategies against their non-annual counterparties, the effects become stronger. The
annual strategies outperform their non-annual counterparties especially during the
most recent decade and also during the 80's for the equities universe. The relative
benefit (annual versus non-annual) was significantly smaller or even non-existent
during the 90's for both asset classes. Table 2 reports various performance statistics
for these strategies.
All annual seasonality strategies (within both asset classes and across sample periods)
generate strong and statistically significant positive returns. Focusing on the common
sample period (January 1992 to February 2015), the annualised arithmetic returns are
close to 4% for commodities and 3% for equity indices (both statistically significant
at 5%), hence resulting to Sharpe ratios of 0.49 and 0.45. On the contrary, the non-
9
In order to implement a constant-volatility (𝐶𝐶𝐶𝐶) overlay on a portfolio with monthly returns denoted by 𝑟𝑟𝑡𝑡 in month 𝑡𝑡,
we scale the exposure to it (by effectively employing dynamic leverage) based on an estimate of its running volatility at
the end of the previous month 𝜎𝜎𝑡𝑡−1 ; we use a rolling window of 100 days for our analysis. The returns of the 𝐶𝐶𝐶𝐶 strategy
are given by:
7%
𝑟𝑟𝑡𝑡𝐶𝐶𝐶𝐶 = ⋅ 𝑟𝑟
𝜎𝜎𝑡𝑡−1 𝑡𝑡
10
annual seasonality strategies fail to deliver any positive and statistically significant
returns. As a consequence, the spread return between the two strategies, annual minus
non-annual (presented in the last two rows of the table) is always positive and it is
statistically significant for commodities (at 10%) and for equity indices, but only over
the longer sample period (at 1%).
11
constructed using futures contracts; for further details see the aforementioned paper as
well as Asness, Moskowitz and Pedersen (2013) and. Briefly, we use the following
three factors for each asset class:
Panel A of Table 3 presents the results of the following monthly regression over the
period January 1992 to February 2015 for commodities and January 1995 to February
2015 for equity indices:
The evidence shows that the returns from the annual seasonality long-short strategy
cannot be explained through exposure to existing return factors. All regression
coefficients for both asset classes are small in magnitude and statistically
insignificant. This finding can also be supported by the very low correlations between
the seasonality strategies and the rest of the factors in Panel B of Table 3. Very
interestingly, the seasonality strategies do not exhibit any beta with the underlying
market (they are effectively market neutral strategies). The only statistically strong (at
10% level) parameter in Panel A of Table 3 is the alpha of the seasonality strategies.
10
See Baltas and Kosowski (2015).
12
Table 3: Factor Decomposition of Seasonality Strategies
Seasonality 1.00
MKT 0.04 1.00
VAL -0.06 -0.23 1.00
MOM 0.13 0.11 -0.47 1.00
Equity Indices
Seasonality 1.00
MKT 0.02 1.00
VAL -0.01 0.46 1.00
MOM -0.07 -0.10 -0.36 1.00
Notes: The figure presents in Panel A the results of time-series regressions of annual seasonality strategies on a market
factor (equally-weighted), a value factor and a momentum factor. The t-statistics are calculated using White (1980)
heteroskedasticity robust standard errors. Statistical significance at 1%, 5% and 10% levels is denoted by ***, ** and *
respectively. Panel B reports the correlation matrix of the factors for each asset class. The commodity dataset is
obtained from Bloomberg, whereas the equity index dataset is obtained from Datastream.
For that purpose, Table 4 reports the average annualised arithmetic return and the
respective Sharpe ratio before and after transaction costs for the annual seasonality
strategy of each asset class, as well as for the list of factors that were used in the
factor analysis of the previous section (market, value and momentum). For
comparison purposes, we also include a trend-following (TF) strategy, which is
formed by long and short positions on assets with positive or negative 12-month
excess returns respectively on an inverse-volatility weight basis. Further details on the
construction of trend-following strategies follow in the next section.
Panel A of Table 4 reports statistics for raw strategies, whereas Panel B reports
statistics for their respective constant-volatility counterparts that employ dynamic
leverage in order to target ex-ante a volatility of 7%. Employing dynamic leverage
and targeting a certain level of volatility can scale a strategy to the risk-target of a
portfolio manager, but this requires additional turnover and therefore transaction
costs. This is why we report results both for the raw strategies and their constant-
13
volatility versions. Our aim is to evaluate the after-costs performance of the strategies
when reasonable practical levels of target volatility are employed.
In order to estimate the realised costs, we use a very simplistic rule (our after-costs
illustrative results should therefore be treated with the required caution), suggested by
Frazzini, Israel and Moskowitz (2012), based on which the realised costs are the
product of portfolio turnover and an average level of market impact:
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 = 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 × 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼 (4)
Frazzini et al. (2012) report an average market impact cost for conventional equity
styles (value, momentum, short-term reversals) around 15-25 basis points. We decide
to use a reasonable, yet conservative level of 25 basis points (trading liquid
commodity futures or equity indices using ETFs is typically substantially cheaper).
Table 4: The Effect of Turnover and Transaction Costs (assuming Market Impact costs of 25 basis points)
Before costs:
Arithmetic Mean (%) 4.45 4.69 15.09*** 6.45*** 15.36** 7.58** 3.85 4.27 8.06*** 4.80**
Volatility (%) 14.46 25.35 23.59 9.11 31.18 15.59 12.98 15.06 12.30 11.23
Sharpe Ratio 0.31 0.19 0.64 0.71 0.49 0.49 0.30 0.28 0.66 0.43
Costs:
Monthly Turnover (%) 6.48 39.36 41.40 33.04 160.05 4.59 16.19 46.32 22.18 164.35
Realised Costs (%) 0.19 1.18 1.24 0.99 4.80 0.14 0.49 1.39 0.67 4.93
After costs:
Arithmetic Mean (%) 4.25 3.51 13.85 5.46 10.56 7.45 3.36 2.88 7.40 -0.14
Sharpe Ratio 0.29 0.14 0.59 0.60 0.34 0.48 0.26 0.19 0.60 -0.01
Before costs:
Arithmetic Mean (%) 2.66* 2.14 5.22*** 6.88*** 3.97** 4.43** 1.56 2.50* 6.23*** 2.93**
Volatility (%) 7.65 8.23 7.70 7.56 8.11 7.90 5.70 6.20 7.74 6.49
Sharpe Ratio 0.35 0.26 0.68 0.91 0.49 0.56 0.27 0.40 0.81 0.45
Costs:
Monthly Turnover (%) 9.39 42.01 43.42 35.89 161.29 10.58 20.91 50.59 27.37 166.74
Realised Costs (%) 0.28 1.26 1.30 1.08 4.84 0.32 0.63 1.52 0.82 5.00
After costs:
Arithmetic Mean (%) 2.38 0.88 3.92 5.80 -0.87 4.11 0.96 0.98 5.41 -2.07
Sharpe Ratio 0.31 0.11 0.51 0.77 -0.11 0.52 0.16 0.16 0.70 -0.32
Notes: The figure presents the average annualised arithmetic return, volatility and Sharpe ratio for a market (MKT) factor (equally-weighted), a value (VAL) factor, a momentum
(MOM) factor, a trend-following (TF) factor and an annual seasonality long-short (top versus bottom quintile) strategy before and after trading costs within commodities and equity
indices. The realised costs are estimated as the product of the annualised turnover and an assumed level of average market impact equal to 25bps. Panel A presents the statistics for raw
strategies and Panel B presents the statistics for constant-volatility (at 7%) strategies. The sample period is January 1992 to February 2015; the only exception is the equity value
strategy, which starts in January 1995. All the strategies have been constructed using futures contracts except for the annual seasonality strategy within equities which has been
constructed using MSCI country total return indices. The statistical significance of the average arithmetic returns before costs are tested using White (1980) heteroskedasticity robust
standard errors; statistical significance at 1%, 5% and 10% levels is denoted by ***, ** and * respectively. The commodity dataset is obtained from Bloomberg, whereas the equity
index dataset is obtained from Datastream.
14
It is fairly obvious from both Panels of Table 4 that the annual seasonality strategy
exhibits at least one order of magnitude larger turnover than any other strategy that we
study. This was –to some extent– expected, as the seasonality strategy buys every
month the same-calendar-month past winners and sells the same-calendar-month past
losers and as a consequence almost the entire portfolio has to be liquidated and re-
constructed (monthly turnover estimates for the long-short portfolio exceed 160%).
The substantially larger turnover of the seasonality strategy turns out to be detrimental
as the after-costs Sharpe ratios turn negative in most cases. On the contrary, all other
premia remain relatively strong even after incorporating trading costs. To obtain a
visual perspective on the before-costs and after-costs Sharpe ratios, both for the raw
and the constant-volatility strategies, we present our estimates in Figure 4.
0.6
Sharpe Ratio
0.4
0.2
0.0
-0.2
-0.4
MKT VAL MOM TF Annual MKT VAL MOM TF Annual
Commodities Equities
0.6
Sharpe Ratio
0.4
0.2
0.0
-0.2
-0.4
MKT VAL MOM TF Annual MKT VAL MOM TF Annual
Commodities Equities
Notes: The figure presents the before-costs and after-costs Sharpe rations for a market (MKT) factor (equally-weighted), a
value (VAL) factor, a momentum (MOM) factor, a trend-following (TF) factor and an annual seasonality long-short (top
versus bottom quintile) strategy within commodities and equity indices. The realised costs are estimated as the product of
the annualised turnover and an assumed level of average market impact equal to 25bps. The top chart presents the results
for raw strategies and the bottom chart presents the results for constant-volatility (at 7%) strategies. The sample period is
January 1992 to February 2015; the only exception is the equity value strategy, which starts in January 1995. All the
strategies have been constructed using futures contracts (obtained from Bloomberg) except for the annual seasonality
strategy within equities which has been constructed using MSCI country total return indices (obtained from Datastream).
15
The bottom-line of this analysis is that the seasonality premium within commodity
and equity index markets might appear positive and statistically significant on paper,
but in practice it might be completely eliminated by trading costs due to the very large
associated turnover. This constitutes one plausible explanation as to why the
seasonality premium has not been arbitraged away.
Following from these results, it is obvious that seasonality signals cannot be traded
directly by an investor. However, as the next section of the paper illustrates, these
signals bear important information that can be incorporated in the construction of
trend-following strategies.
More formally, let 𝑁𝑁𝑡𝑡 denote the number of available assets at time 𝑡𝑡. A trend-
following strategy involves taking a long or short position on each asset 𝑖𝑖, based on
the sign of the past excess return over a prescribed lookback period that is typically
equal to 12 months. 11 The weighting scheme that is typically employed (especially if
assets of the same asset class comprise the portfolio) is the inverse-volatility
weighting scheme. 12,13 The return of the strategy is given by:
𝑁𝑁𝑡𝑡 −1
�𝜎𝜎𝑡𝑡𝑖𝑖 �
𝑇𝑇𝑇𝑇
𝑟𝑟𝑡𝑡,𝑡𝑡+1 = 𝑖𝑖
� 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠�𝑟𝑟𝑡𝑡−12,𝑡𝑡 � ∙ 𝑖𝑖
∙ 𝑟𝑟𝑡𝑡,𝑡𝑡+1 (5)
𝑗𝑗 −1
𝑖𝑖=1 ∑𝑁𝑁
𝑗𝑗=1�𝜎𝜎𝑡𝑡 �
𝑡𝑡
For further details on the construction and dynamics of trend-following strategies see
Moskowitz, Ooi and Pedersen (2012), Hurst, Ooi and Pedersen (2012, 2013), Baltas,
Jessop, Jones and Zhang (2013) and Baltas and Kosowski (2013, 2015).
11
Moskowitz et al. (2012), Baltas and Kosowski (2013) and Baltas, Jessop, Jones and Zhang (2013) find that a 12-month
horizon generates the largest Sharpe ratio for trend-following strategies within each asset class.
12
It can be shown that the inverse-volatility weighting scheme can split portfolio volatility equally across portfolio
constituents as long as all pairwise correlations are equal.
13
As highlighted in Baltas, Jessop, Jones and Zhang. (2013, 2014) and Baltas (2015), working with assets across multiple
asset classes can qualify risk-parity (equal risk contribution) as a more appropriate weighting scheme, especially in
periods of extreme asset co-movement. For the purposes of this paper, this point is not relevant as we only construct
trend-following strategies using assets of the same asset class.
14
A similar technique has been employed by Barroso and Santa-Clara (2014) and Daniel and Moskowitz (2014), who
focus on cross-sectional winners-minus-losers momentum strategies. See also Hallerbach (2012, 2014).
16
The dynamic leverage equals the ratio 𝜎𝜎𝑇𝑇𝑇𝑇𝑇𝑇 /𝜎𝜎𝑡𝑡𝑇𝑇𝑇𝑇 . As an example, if 𝜎𝜎𝑇𝑇𝑇𝑇𝑇𝑇 = 10% and
at the end of some month the running volatility of the unlevered strategy is 5%, then
all positions for the forthcoming month should be doubled (a 2x leverage ratio).
One way to incorporate the seasonality signals into the construction of trend-
following portfolios is to avoid taking a certain long or short position, if there exists a
contradicting seasonality signal. In particular:
- Switch off a long position if the asset belongs in the bottom seasonality basket
(i.e. it has one of the lowest same-calendar-month average past returns).
- Switch off a short position if the asset belongs in the top seasonality basket
(i.e. it has one of the largest same-calendar-month average past returns).
As we have done throughout this paper, we measure the seasonality signals over the
course of the past ten years and require at least five years for an asset to be included in
the cross-sectional ranking. Given that the trend-following signal uses information
from the past 12 months, we exclude the past year's same-month return from the
seasonality calculation in order to safeguard against any correlation between the two
signals. Figure 5 illustrates the trend-following and seasonality signals generation.
Using these signal conventions, Figure 6 presents the cumulative returns of a simple
trend-following strategy and its seasonality-adjusted variant within commodities and
equity indices. All strategies in Figure 6 incorporate a 7% volatility target.
Figure 5: Trend-following and Seasonality Signals
Notes: The figure describes the signal generation for the seasonality-adjusted trend-following strategy.
500 500
400 400
300 300
200 200
100 100
0 0
Trend-Following Trend-Following with Seasonality overlay Trend-Following Trend-Following with Seasonality overlay
Notes: The figure presents the cumulative returns of a trend-following strategy and its variant that incorporates seasonality information within commodity and equity index
markets. All strategies in this figure employ a 7% volatility target. The sample period is January 1992 to February 2015. The commodity dataset is obtained from Bloomberg,
whereas the equity index dataset is obtained from Datastream.
17
What is fairly obvious from these plots is that the seasonality overlay improves the
performance of the trend-following strategies before any transaction costs are taken
into account. Table 5 presents performance statistics before accounting for any
transaction costs, as well as after accounting for different levels of costs.
TF Enhanced TF TF Enhanced TF
Before costs:
Arithmetic Mean (%) 6.45*** 7.52*** 8.06*** 8.72***
Volatility (%) 9.11 9.32 12.30 12.32
Sharpe Ratio 0.71 0.81 0.66 0.71
Monthly Turnover (%) 33.04 55.79 22.18 44.90
Wilcoxon Paired Test p-values:
Two-Sided 15.32% 10.12%
One-Sided 7.66% 5.08%
After Costs:
Sharpe Ratio (for 5 bps) 0.69 0.77 0.64 0.69
Sharpe Ratio (for 10 bps) 0.66 0.74 0.63 0.66
Sharpe Ratio (for 15 bps) 0.64 0.70 0.62 0.64
Sharpe Ratio (for 20 bps) 0.62 0.66 0.61 0.62
Sharpe Ratio (for 25 bps) 0.60 0.63 0.60 0.60
TF Enhanced TF TF Enhanced TF
Before costs:
Arithmetic Mean (%) 6.88*** 7.27*** 6.23*** 6.65***
Volatility (%) 7.56 7.42 7.74 7.80
Sharpe Ratio 0.91 0.98 0.81 0.85
Monthly Turnover (%) 35.89 58.01 27.37 49.90
After Costs:
Sharpe Ratio (for 5 bps) 0.88 0.93 0.78 0.81
Sharpe Ratio (for 10 bps) 0.85 0.89 0.76 0.78
Sharpe Ratio (for 15 bps) 0.82 0.84 0.74 0.74
Sharpe Ratio (for 20 bps) 0.80 0.79 0.72 0.70
Sharpe Ratio (for 25 bps) 0.77 0.75 0.70 0.66
Notes: The figure presents the average annualised arithmetic return, volatility and Sharpe ratio for a trend-following
(TF) strategy and its variant that incorporates information from seasonality signals (Enhanced TF) within commodities
and equity indices. After-costs Sharpe ratio estimates are also presented for different levels of average market impact (5,
10, 15, 20 and 25 basis points). The realised costs are estimated as the product of the annualised turnover and the
assumed level of average market impact. Panel A presents the statistics for raw strategies and Panel B presents the
statistics for constant-volatility (at 7%) strategies. Panel A also reports the p-values from two-sided and one-sided
paired Wilcoxon (1945) signed-rank tests for the difference in the monthly average arithmetic returns between the
Enhanced TF strategy and the simple trend-following strategy. The one-sided test has been designed in a way to test
whether the returns of the Enhanced TF strategy are greater than the returns of the simple TF strategy. The sample
period is January 1992 to February 2015. The strategies have been constructed using futures contracts obtained from
Bloomberg. The statistical significance of the average arithmetic returns before costs are tested using White (1980)
heteroskedasticity robust standard errors; statistical significance at 1%, 5% and 10% levels is denoted by ***, ** and *
respectively.
18
The results show that before costs the improvement for the trend-following strategy is
indeed genuine. The Sharpe ratio of the strategy increases from 0.71 to 0.81 for
commodities and from 0.66 to 0.71 for equity indices; the respective changes for
constant-volatility versions of the strategies are 0.91 to 0.98 and 0.81 to 0.85
respectively. Importantly enough, one-sided paired Wilcoxon (1945) signed-rank tests
confirm that the enhanced trend-following strategies generate significantly larger
average returns than the plain trend-following strategies; the p-values are 7.7% for
commodities and 5.1% for equity indices.
This performance pickup does not come at no cost. The turnover of the trend-
following strategies almost doubles after incorporating the seasonality signals. This
results in substantial degradation of the risk-adjusted performance after incorporating
trading costs. Assuming various levels of average market impact, between 5 and 25
basis points, we find that the break-event level of market impact is around 25 basis
points for the raw trend-following strategies (Panel A in Table 5), whereas for the
constant-volatility versions of the strategies (Panel B in Table 5), which anyway
exhibit even larger turnover, the break-even level is around 15 basis points. On a
positive note, these cost estimates might be relatively conservative in practice.
Typically, futures contracts are very liquid and tend to have low trading costs. To put
things in perspective, Hurst et al. (2012), who also study trend-following strategies,
report one-way transaction costs of 1 bp for government bonds, 3 bps for FX rates, 6
bps for equity indices and 10 bps for commodities for the period 2003 to 2012.
Finally, it's worth noting than in our implementation of the trend-following strategy,
we applied no cost optimisation or position smoothing techniques that in practice
would be employed to reduce rebalancing costs. As an example, instead of switching
off the inconsistent long or short positions of a trend-following strategy as instructed
by the respective seasonality signals, one suggestion would be to just halve the
positions. We empirically tested this and the turnover of the enhanced strategies falls
significantly with the performance pickup remaining strong. Baltas and Kosowski
(2015) discuss various other ways to reduce portfolio turnover and therefore improve
the after-costs performance of the strategy.
5. Concluding Remarks
The purpose of this paper has been to investigate the existence of seasonality patterns
in the cross-section of asset returns within different asset classes (commodities, FX
rates, government bonds and equity indices). Using panel regression framework, we
identify strong seasonality patterns within commodities and equity indices and for that
reason the rest of the analysis focuses mainly on these two asset classes.
The seasonality premium, which can be captured by a long-short portfolio that buys /
sells the assets with the largest / smallest same-calendar-month past average returns
(up to ten years), is statistically and economically significant for both asset classes
and does not appear to proxy for some known systematic source of risk. Instead, our
interpretation for its persistent behaviour over time draws from a limits-to-arbitrage
perspective. The seasonality strategies exhibit very large turnover and some
reasonable trading costs are enough to completely eliminate the premium, especially
when the strategies are run with a realistic volatility target overlay.
19
An attempt to make practical use of the seasonality signals is to actively use them to
switch off inconsistent long and short positions in a trend-following framework. Our
back-tests have shown statistically significant improvement of the plain trend-
following strategy within both commodities and equity indices. The increased
turnover, caused by the incorporation of the seasonal signals, can strongly impact the
performance pickup. However, given that these strategies are mainly constructed
using liquid front futures contracts and sophisticated position smoothing techniques,
the respective trading costs can be relatively managed to remain at low levels.
Figure 7: Risk & Return for Dynamic Leveraged Commodity and Equity Index Strategies (with volatility target at 7%)
Commodities Equity Indices
10.0 10.0
Trend
Following w. Trend
Annualised Geometric Returns (%)
Annualised Geometric Returns (%)
20
APPENDIX A - Panel Regressions with lags up to 120 months
Figure A.1: Panel Regression for Equity Indices – Lags up to 120 months
Regression Coefficients
10
8
6
4
2
0
-2
-4
-6
-8
-10
1 3 6 9 12 18 24 36 48 60 72 84 96 108 120
t-statistics
6
-2
-4
1 3 6 9 12 18 24 36 48 60 72 84 96 108 120
Notes: The charts report the Fama-MacBeth panel regression coefficients and respective t-statistics from regressing monthly returns of 18 equity index spot price indices on their respective lagged
returns, for lags ranging from 1 to 120 months. The t-statistics are adjusted for serial-correlation and heteroskedasticity using the Newey and West (1987) correction with 12 lags. The dashed lines at
-1.96 and +1.96 denote the 5% significance thresholds. The sample period is from January 1980 to December 2014. The dataset is obtained from Datastream.
21
Figure A.2: Panel Regression for Commodities – Lags up to 120 months
Regression Coefficients
10
8
6
4
2
0
-2
-4
-6
-8
-10
1 3 6 9 12 18 24 36 48 60 72 84 96 108 120
t-statistics
6
-2
-4
-6
1 3 6 9 12 18 24 36 48 60 72 84 96 108 120
Notes: The charts report the Fama-MacBeth panel regression coefficients and respective t-statistics from regressing monthly returns of 19 commodity spot price indices on their respective lagged
returns, for lags ranging from 1 to 120 months. The t-statistics are adjusted for serial-correlation and heteroskedasticity using the Newey and West (1987) correction with 12 lags. The dashed lines at
-1.96 and +1.96 denote the 5% significance thresholds. The sample period is from July 1991 to December 2014. The dataset is obtained from Datastream.
22
APPENDIX B – Commodities dataset for strategy construction
Figure B.1: Commodity Datasets
Spot Futures
References
Andersen, T. G., & Bollerslev, T. (1997). Intraday Periodicity and Volatility Persistence in
Financial Markets. Journal of Empirical Finance, 4(2), 115-158.
Asness, C. S., Moskowitz, T. J., & Pedersen, L. H. (2013). Value and Momentum
everywhere. Journal of Finance, 68(3), 929-985.
Baltas, N. (2015). Trend-Following, Risk-Parity and the influence of Correlations, Risk-based
and Factor Investing (edited by Emmanuel Jurczenko), Elsevier and ISTE Press, 65-95.
Baltas, N., Jessop, D., & Jones, C. (2014). Harvesting Cross-Asset Value. UBS Investment
Research.
Baltas, N., Jessop, D., Jones, C., & Zhang, H. (2013). Trend-Following meets Risk-Parity.
UBS Investment Research.
Baltas, N., Jessop, D., Jones, C., & Zhang, H. (2014). Correlation, De-correlation and Risk-
Parity. UBS Investment Research.
Baltas, N., & Kosowski, R. (2013). Momentum Strategies in Futures Markets and Trend-
Following Funds. Available at SSRN: http://ssrn.com/abstract=1968996
Baltas, N., & Kosowski, R. (2015). Demystifying Time-Series Momentum Strategies:
Volatility Estimators, Trading Rules and Pairwise Correlations. Available at SSRN:
http://ssrn.com/abstract=2140091.
23
Barroso, P., & Santa-Clara, P. (2014). Momentum has its Moments. Journal of Financial
Economics, 116(1), 111-120.
Daniel, K. D., Moskowitz, T. J. (2014). Momentum Crashes. Available at SSRN:
http://ssrn.com/abstract=2486272.
de Roon, F. A., Nijman, T. E., & Veld, C. (2000). Hedging Pressure Effects in Futures
Markets. Journal of Finance, 55(3), 1437-1456.
Fama, E. F. (1976). Foundations of Finance. Basic Books.
Fama, E. F., & MacBeth, J. D. (1973). Risk, return, and equilibrium: Empirical tests. Journal
of Political Economy, 81(3), 607-636.
Frazzini, A., Israel, R. and Moskowitz, T. J. (2012). Trading Costs of Asset Pricing
Anomalies, Available at SSRN: http://ssrn.com/abstract=2294498
Hallerbach, W. G. (2012). A proof of the Optimality of Volatility Weighting over time.
Journal of Investment Strategies, 1(4), 87-99
Hallerbach, W. G. (2014). On the Expected Performance of Market Timing Strategies.
Journal of Portfolio Management, 40(4), 42-51.
Heston, S. L., & Sadka, R. (2008). Seasonality in the Cross-Section of Stock Returns. Journal
of Financial Economics, 87(2), 418-445.
Heston, S. L., & Sadka, R. (2010). Seasonality in the Cross-Section of Stock Returns: The
International Evidence. Journal of Financial and Quantitative Analysis, 45(5), 1133-1160.
Hurst, B., Ooi, Y. H., & Pedersen, L. H. (2012). A Century of Evidence on Trend-Following
Investing. AQR Capital Management.
Hurst, B., Ooi, Y. H., & Pedersen, L. H. (2013). Demystifying Managed Futures. Journal of
Investment Management, 11(3), 42-58.
Ilmanen, A. (2012). Expected Returns: An Investor's Guide to Harvesting Market Rewards.
John Wiley & Sons.
Ito, T., & Hashimoto, Y. (2006). Intraday Seasonality in activities of the Foreign Exchange
Markets: Evidence from the Electronic Broking System. Journal of the Japanese and
International Economies, 20(4), 637-664.
Keloharju, M., Linnainmaa, J. T., & Nyberg, P. M. (2015). Common Factors in Return
Seasonalities. Available at SSRN: http://ssrn.com/abstract=2224246.
Moskowitz, T. J., Ooi, Y. H., & Pedersen, L. H. (2012). Time Series Momentum. Journal of
Financial Economics, 104(2), 228-250.
Newey, W. K., & West, K. D. (1987). A Simple, Positive Semi-definite, Heteroskedasticity
and Autocorrelation Consistent Covariance Matrix. Econometrica, 55(3), 703-08.
Shleifer, A., & Vishny, R. W. (1997). The Limits of Arbitrage. Journal of Finance, 52(1), 35-
55.
Sørensen, C. (2002). Modeling Seasonality in Agricultural Commodity Futures. Journal of
Futures Markets, 22(5), 393-426.
Todorova, M. I. (2004). Modeling Energy Commodity Futures: is Seasonality part of it?
Journal of Alternative Investments, 7(2), 10-32.
White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a direct
test for heteroskedasticity. Econometrica, 48(4), 817-838.
Wilcoxon, F. (1945). Individual Comparisons by Ranking Methods. Biometrics Bulletin, 1(6),
80-83.
24