Breaking Bad Trends

AbstractWe document and quantify the negative impact of trend breaks (i.e., turning points in the trajectory of asset prices) on the performance of standard monthly trend-following strategies across several assets and asset classes. In the years of the US economy’s expansion following the global financial crisis of 2008, we find an increase in the frequency of trend breaks, which helps explain the lower performance of these trend strategies during this period. We illustrate how to repair such strategies using a dynamic trend-following approach that exploits the return-forecasting properties of the two types of trend breaks: market corrections and rebounds.Keywords: Asset pricingbehavioral financemarket timingmean reversionmomentum speedtime-series momentumtrend followingturning pointsvolatility timingPL Credits: 2.0: AcknowledgmentsWe appreciate the comments of the Co-Editor, Daniel Giamouridis, and two anonymous referees. We thank Jaynee Dudley and Kay Jaitly, CFA, for editorial assistance. We also thank Ashish Garg, who was a collaborator on an early version of this paper. Harvey reports financial support was provided by Research Affiliates. The initial version of the paper was written when Goulding and Mazzoleni were employees of Research Affiliates. The views expressed in this article are those of the authors and do not necessarily reflect the positions of STRS Ohio.Disclosure statementNo potential conflict of interest was reported by the author(s).Notes1 The literature documents that asset returns measured over the recent past are positively correlated with future returns (Jegadeesh and Titman Citation1993, Citation2001; Asness Citation1994; Conrad and Kaul Citation1998; Lee and Swaminathan Citation2000; Gutierrez and Kelley, Citation2008). This phenomenon is stable across assets and countries (Rouwenhorst Citation1998; Griffin, Ji, and Martin Citation2003; Israel and Moskowitz Citation2013; Asness, Moskowitz, and Pedersen Citation2013). Studies of the merits of trend following and time-series momentum investing, in particular, include the following: Cutler, Poterba, and Summers (Citation1991), Silber (Citation1994), Fung and Hsieh (Citation1997, Citation2001), Erb and Harvey (Citation2006), Moskowitz, Ooi, and Pedersen (Citation2012), Menkhoff et al. (Citation2012), Baltas and Kosowski (Citation2013), Hurst, Ooi, and Pedersen (Citation2013), Baltas and Kosowski (Citation2015), Levine and Pedersen (Citation2016), Georgopoulou and Wang (Citation2017), Hurst, Ooi, and Pedersen (Citation2017), Ehsani and Linnainmaa (Citation2022), Goulding, Harvey, and Mazzoleni (Citation2023), Gupta and Kelly (Citation2019), and Babu et al. (Citation2020a, 2019b).2 See Goulding, Harvey, and Mazzoleni (Citation2023).3 The cross-sectional momentum literature has explored themes related to market cycles and turning points (Cooper, Gutierrez, and Hameed Citation2004; Daniel, Jagannathan, and Kim, Citation2012; Daniel and Moskowitz, Citation2016). Cooper, Gutierrez, and Hameed use a slow trailing 3-year return to define two market states: “up” and “down.” Daniel, Jagannathan, and Kim (Citation2012) use a two-state hidden Markov model of unobserved “turbulent” and “calm” states to predict crashes in cross-sectional momentum portfolios. Daniel and Moskowitz (Citation2016) study cross-sectional momentum crashes and recoveries and propose a dynamic cross-sectional weighting strategy. Goulding, Harvey, and Mazzoleni (Citation2023) use the intersection of slow and fast trailing return signals to characterize four market states and to define trend turning points.4 A 12-month lookback window is the standard window length for time-series momentum analyzed in the literature (Moskowitz, Ooi, and Pedersen, Citation2012 and Huang et al., Citation2020), among others. Some studies consider shorter lookback windows such as 1, 2, or 3 months or consider averages of strategies with 12-month and shorter lookback windows (e.g., Babu et al. Citation2020a). We look at such alternatives in a later section.5 Goulding, Harvey, and Mazzoleni (Citation2023) focus exclusively on equity indices and do not explore the role of turning-point frequency.6 Performance of the Société Générale (SG) Trend Index, an equally weighted index of major trend-focused funds, launched at the beginning of 2000, experienced an annualized Sharpe ratio of approximately 0.41 over its first decade (2000–2009). In its second decade (2010–2019), its annualized Sharpe ratio fell by nearly half (0.21) and the index experienced its worst drawdown, losing more than 20% over the 45-month period April 2015 to January 2019. Likewise, the annualized Sharpe ratio of a hypothetical multi-asset portfolio of 12-month trend-following strategies with monthly rebalancing decreased substantially in the post-GFC expansion period—see static multi-asset trend-following performance in later sections.7 The 11-year post-GFC expansion period ended with the brief COVID recession event of 2020, which experienced a lower number of turning points but of higher severity. The two years of subsequent expansion (2021–2022) exhibit some reversion of turning-points frequency to pre-GFC levels. In Appendix E, we report performance statistics of several trend-following strategies over various sample periods: Full sample 1990–2022; through GFC 1990–2008; post-GFC 2009–2022; and post-GFC expansion 2009–2019.8 This approach is distinct from moving average crossovers, which Levine and Pedersen (Citation2016) show are essentially equivalent to static blends of time-series momentum strategies. Hurst, Ooi, and Pedersen (Citation2013) show that the returns of trend-following strategies such as Managed Futures funds and CTAs can be explained by static blends of time-series momentum strategies.9 See Han et al. (Citation2021) for a survey of the literature on technical analysis including trend following.10 We define the asset’s trailing 12-month return as the arithmetic average of the preceding 12 months of monthly returns in excess of cash, which is the implied rate of market borrowing for institutions.11 Volatility scaling may have a distinct effect from time-series momentum (e.g., Kim, Tse, and Wald Citation2016; Moreira and Muir Citation2017; Harvey et al., Citation2018; Goulding, Harvey, and Mazzoleni Citation2023) and we seek to avoid intermixing the two mechanisms.12 Given our definition, observing a turning point does not necessarily reflect an actual trend break. In particular, in noisy periods, some turning points can be false alarms of a true turn. In later sections, we will refine our definition of turning points to distinguish between turning points from up to down (corrections) and from down to up (rebounds). For now, our classification is sufficient to illustrate our key finding.13 Performance is gross of costs to roll contracts or of any transaction costs. In Appendix E, we report the turnover of various trend strategies including static 12-month trend and a dynamic trend strategy developed in a later section. The dynamic trend strategy incurs more turnover than the static trend strategy. Nevertheless, for average transaction costs below 29 basis points—a comfortable upper bound for these assets—the dynamic strategy remains more profitable than the static strategy.14 In our sample, three asset-years have zero turning points, seven asset-years have 9 or 10 turning points, and no asset-years have 11 or 12 turning points.15 Equally weighted averages reflect more volatility from riskier assets such as commodities and equities. Our results are similar throughout our analyses if we weight each asset by its full-sample inverse volatility or by its trailing inception-to-date inverse volatility in order to normalize each asset’s underlying risk contribution to the multi-asset portfolio.16 In the estimation of the trend line, we exclude 2020 as an outlier due to COVID-19. Our turning-point measure captures frequency but not severity of trend breaks. In this outlier year, we find a lower-than-average number of turning points but a larger magnitude of market shift. If this year were included, the trend line would still exhibit a strong negative relationship (R2 = 0.56 and slope −0.18).17 The average increases to 4.80 asset turning points per year relative to 4.54 in other years. We can see in Figure 2 that the number of turning points in years 2020–2022 reverted to near typical levels, with 2020 being an outlier in terms of its relationship with performance. In 2020, investors in static 12-month trend would have experienced a relatively high loss despite a lower-than-average number of turning points.18 As with our definition of a turning point, noisy periods can create temporary and unintuitive correction or rebound classifications, which could be refined by other definitions beyond the scope of this study.19 In Appendix D, we show results of the dynamic approach using 1- and 12-month lookback horizons for fast and slow signals, respectively, and obtain consistent conclusions. Use of the 2-month signal highlights that each asset may respond differently to the market phases defined by different choices of slow and fast momentum strategies. For example, the disagreement between 3-month and 12-month trend directions might yield better informative states for bonds while the disagreement between 1-month and 12-month trend directions might be more informative for equities. Likewise, the diversification properties across different assets may also vary with the choice of slow and fast signals.20 In Appendix E, we report additional portfolio statistics over various subsamples for each strategy shown in Figure 4, but without normalized volatilities.Additional informationNotes on contributorsChristian L. GouldingChristian L. Goulding is Assistant Professor of Finance at Harbert College of Business, Auburn University, Auburn, Alabama.Campbell R. HarveyCampbell R. Harvey is Professor of Finance at Fuqua School of Business, Duke University, Durham, North Carolina, and Research Associate at the National Bureau of Economic Research, Cambridge, Massachusetts.Michele G. MazzoleniMichele G. Mazzoleni is Director of Asset Allocation and Strategy at STRS Ohio, Columbus.