Momentum crashes: why momentum collapses rarely but violently, and what volatility scaling and crowding measures show
Momentum is one of the best-documented factor premiums and at the same time one of the most dangerous. Momentum crashes are rare, very deep collapses of the strategy that buys winners and sells losers short. They almost always occur in the same situation: after a bear market, at high volatility, when the market suddenly turns sharply. Here you will learn the mechanics behind the crashes, work through a volatility scaling rule and see how crowding hit quantitative strategies simultaneously in 2007.
What a momentum crash is
The classic momentum strategy sorts stocks by their return over the last 12 months, usually excluding the most recent month, buys the top decile and sells the bottom decile short: winners minus losers, WML for short. Jegadeesh and Titman (1993) documented the effect for US stocks, Asness et al. (2013) in many markets and asset classes; the basics are covered in momentum strategies and Fama-French factors. This article looks at the dark side.
Barroso and Santa-Clara (2015) sum up the dilemma: compared with market, value and size, momentum offered the highest Sharpe ratio, but also had the worst crashes. That makes it unattractive for anyone averse to negative skewness and high kurtosis. Daniel and Moskowitz (2016) describe the crashes as rare, persistent strings of negative returns despite high average returns in many asset classes. They are partly predictable: they occur in panic states, after market declines and at high volatility, and coincide with sharp market rebounds (see momentum crash).
The mechanics: time-varying betas and the short side after bear markets
Why precisely then? The answer lies in the betas. If you buy the winners of the last twelve months, you automatically buy stocks that have behaved like the market over that period. After a long rise, the winners are high-beta stocks; after a bear market, they are low-beta stocks, while the losers are the risky stocks that have fallen sharply. Grundy and Martin (2001) show that momentum therefore has time-varying factor loadings, depending on how the factors performed in the ranking period.
β_WML,t = β_winners,t − β_losers,t
r_WML,t ≈ α + β_WML,t × r_market,t + ε_t- β_winners, β_losers
- market beta of the stocks bought and of those sold short
- β_WML
- beta of the whole long-short strategy
On top of this comes an asymmetry. The losers after a crash are often highly indebted, distressed companies. Their shares behave like calls on the value of the firm: when the market recovers, they rise disproportionately. The short side of the strategy is then a short call. Daniel and Moskowitz (2016) interpret the low expected returns in panic states in exactly this way: as a high premium that the market then demands for the option-like payoffs of the past losers. For hedging, the good news from Grundy and Martin (2001) matters: adjusted for the dynamic factor loadings, momentum profits have been remarkably stable since 1926, and momentum based on stock-specific returns was more profitable than momentum based on total returns.
Volatility scaling: limiting risk in advance
Because the risk of momentum varies strongly and is predictable, you can scale the position accordingly. Barroso and Santa-Clara (2015) find that such risk management virtually eliminates the crashes and almost doubles the Sharpe ratio. Moreover, they argue, risk-managed momentum is an even bigger puzzle than the original. The rule is the same as for target volatility:
w_t = min(σ* / σ̂_t−1, w_max)
σ̂_t−1 = standard deviation of the strategy's daily returns up to the previous month × √252- σ*
- target volatility per year
- σ̂
- realized volatility, using only data up to the previous month
- w_max
- cap on leverage
Weight = min(12 / σ̂, cap) as in the example above. As soon as the cap no longer binds, the curve falls hyperbolically: if volatility triples, the position shrinks to a third.
Daniel and Moskowitz (2016) go one step further: their dynamic strategy weights according to forecasts of the mean and variance, i.e. roughly in proportion to μ̂ / σ̂². In panic states, expected variance rises and, at the same time, expected return falls. The position shrinks twice over. In their data, the strategy roughly doubled the alpha and Sharpe ratio of static momentum, could not be explained by other factors and held up across time periods, international equity markets and other asset classes. More generally, Moreira and Muir (2017) show that portfolios that take less risk when volatility is high achieved higher Sharpe ratios for the market and many factors.
Crowding and the quant meltdown of August 2007
Crashes arise not only from betas but also from positions. When many investors hold the same factor portfolios, one investor's selling becomes everyone's loss. This is called crowding. The clearest example is the quant meltdown of August 2007. Khandani and Lo (2011) simulate long/short portfolios based on five valuation factors and find that the crisis began as early as July and lasted until the end of 2007. Using transaction data, they identify two waves of unwinding: on August 1 from 10:45 to 11:30 and on August 6 from the open until 13:00. Both began in financials, on the long side of value by book-to-price and on the short side of earnings momentum. A simulated market-making strategy lost money in the week from August 6, but was profitable before and after. This points to market-wide deleveraging and the sudden withdrawal of market-making capital from August 8. The article pairs trading describes the sequence of events from the perspective of market-neutral funds.
How do you measure crowding before it blows up? Lou and Polk (2022) propose comomentum: the correlation of the excess returns of the stocks that a typical momentum strategy bets on. When it is low, arbitrageurs are correcting an underreaction and momentum has a stabilizing effect. When it is high, the returns of the momentum stocks subsequently reverse sharply. This is a sign that crowded trading has pushed prices away from fundamentals. The idea behind it: when many investors buy and sell the same stocks at the same time, these stocks move together beyond what common factors explain.
Comomentum_t ≈ average pairwise correlation of the residuals ε_i,τ (from a factor model) of the winner or loser stocks in the window before t- ε_i,τ
- a stock's excess return after removing the factors, measured at high frequency
- Window
- short period before the measurement date
Another channel is liquidity. Pástor and Stambaugh (2003) find that a factor for market-wide liquidity risk explains half of momentum profits from 1966 to 1999. Momentum therefore earns money partly because it loses in liquidity crises, precisely when forced selling moves prices. More on this in the article liquidity risk.
Advanced: why volatility scaling works
Scaling by volatility pays off when volatility is persistent but expected return does not rise to the same extent. Then the ratio of expected return to variance is high in calm phases and low in turbulent ones, and constant risk allocates capital better over time. This applies to momentum in particular: its expected return even falls in panic states (Daniel and Moskowitz 2016), while volatility rises. Formally, a mean-variance investor chooses w_t = μ_t / (γ σ_t²) in each period; with constant μ, this becomes scaling by 1/σ², and with μ proportional to σ, scaling by 1/σ.
- No look-ahead: σ̂ may only use data up to the rebalancing date. If you use the volatility of the month itself, you measure the crash instead of anticipating it.
- Estimators: daily returns over a few months react faster than monthly returns; GARCH models or exponentially weighted estimators are alternatives. More responsiveness means more turnover.
- Leverage and costs: in calm phases, the rule calls for leverage above 1. Financing, borrowing fees on the short side and trading costs eat up part of the advantage; a cap such as w_max = 2 is common.
- Residual momentum: according to Grundy and Martin (2001), it is better to sort by stock-specific returns. This removes part of the time-varying betas before they turn into a crash.
- Beta hedging: instead of cutting the entire risk, you can estimate the conditional beta and neutralize it with index futures (see overlay management). This helps against the beta component, not against the option-like asymmetry.
What research shows
- Crashes are predictable: Daniel and Moskowitz (2016) locate them in panic states after market declines at high volatility; a dynamic strategy roughly doubled alpha and the Sharpe ratio.
- Managing risk helps: Barroso and Santa-Clara (2015) almost eliminate the crashes with risk management and nearly double the Sharpe ratio; Moreira and Muir (2017) show the effect for other factors too.
- Dynamic betas: Grundy and Martin (2001) explain a large part of momentum risk with time-varying factor loadings; factor models explain 95% of the variation of the winner and loser portfolios, but not their average returns.
- Crowding: Lou and Polk (2022) measure arbitrage activity with comomentum; Khandani and Lo (2011) document the deleveraging in August 2007.
- Liquidity: Pástor and Stambaugh (2003) explain half of momentum profits with a liquidity risk factor.
Limitations and common mistakes
- Backtests without costs: momentum has high turnover, the short side incurs borrowing fees, and scaling adds leverage. Gross figures overstate the result.
- Trying out parameters: target volatility, window and cap invite optimization with hindsight (see backtest overfitting).
- Crowding of the scaling itself: if many investors scale by volatility, they all sell at the same time as soon as volatility rises. The rule can amplify the next sell-off.
- Long-only is something else: a momentum ETF without a short side carries the full market risk, but not the short call on the losers; its crashes look different.
- Rare events: the big crashes can be counted on a few episodes. There is hardly any statistical certainty about their frequency.
- Not investment advice: this article explains mechanics and research; it does not recommend any strategy.
How Portfolio Lab calculates
Under Factors, the factor statistics show the Fama and French momentum factor for a chosen model or across regions: average and geometric premium, volatility, Sharpe ratio, worst month and worst year, maximum drawdown, the longest period below the high, skewness and excess kurtosis, and the rolling premium over 3, 5, 10 or 20 years. This lets you see the crashes directly in the factor returns. The factor regression estimates how strongly a portfolio loads on momentum.
The app does not calculate a long-short strategy with volatility scaling. The closest is the “Target volatility” model in Tactical: it scales a long-only portfolio so that the realized volatility from monthly returns matches the target. The defaults are 10% with a 6-month lookback, selectable from 2 to 30% and 3 to 36 months, with an invested share of 50 to 200% (default 100%). A suitable building block is, for example, the “Equities world momentum” asset class (@FAKTOR_MOMENTUM) from Asset classes. More in the factors methodology.
Frequently asked questions
What is a momentum crash?
A rare, very deep loss of the strategy that buys winners and sells losers short. It typically occurs after bear markets at high volatility, when the market suddenly rises sharply.
Why does momentum lose when the market recovers?
After a bear market, the losers are high-beta stocks. The strategy is then short these stocks and has a negative overall beta; a sharp rebound hits it directly.
What does volatility scaling do for momentum?
It reduces the position when realized volatility rises, usually before the crash. In research, this almost eliminated the crashes and nearly doubled the Sharpe ratio.
What was the 2007 quant meltdown?
Losses in quantitative long/short equity strategies in August 2007, especially in the week from August 6, triggered by market-wide deleveraging and the withdrawal of market-making capital.
What is factor crowding?
Many investors hold the same factor positions. If one has to sell, they all lose; measures such as comomentum try to read this from the co-movement of the stocks.
Does a momentum ETF have the same crashes?
Not quite. Without a short side, it carries market risk but not the short call on the losers. It falls with the market and can lag during a rotation into former losers.
Literature
- Asness, Clifford S.; Moskowitz, Tobias J.; Pedersen, Lasse Heje (2013): Value and Momentum Everywhere. The Journal of Finance 68 (3), pp. 929–985. doi:10.1111/jofi.12021
- Barroso, Pedro; Santa-Clara, Pedro (2015): Momentum has its moments. Journal of Financial Economics 116 (1), pp. 111–120. doi:10.1016/j.jfineco.2014.11.010
- Daniel, Kent; Moskowitz, Tobias J. (2016): Momentum crashes. Journal of Financial Economics 122 (2), pp. 221–247. doi:10.1016/j.jfineco.2015.12.002
- Grundy, Bruce D.; Martin, J. Spencer (2001): Understanding the Nature of the Risks and the Source of the Rewards to Momentum Investing. The Review of Financial Studies 14 (1), pp. 29–78. doi:10.1093/rfs/14.1.29
- Jegadeesh, Narasimhan; Titman, Sheridan (1993): Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency. The Journal of Finance 48 (1), pp. 65–91. doi:10.1111/j.1540-6261.1993.tb04702.x
- Khandani, Amir E.; Lo, Andrew W. (2011): What happened to the quants in August 2007? Evidence from factors and transactions data. Journal of Financial Markets 14 (1), pp. 1–46. doi:10.1016/j.finmar.2010.07.005
- Lou, Dong; Polk, Christopher (2022): Comomentum: Inferring Arbitrage Activity from Return Correlations. The Review of Financial Studies 35 (7), pp. 3272–3302. doi:10.1093/rfs/hhab117
- Moreira, Alan; Muir, Tyler (2017): Volatility-Managed Portfolios. The Journal of Finance 72 (4), pp. 1611–1644. doi:10.1111/jofi.12513
- Pástor, Ľuboš; Stambaugh, Robert F. (2003): Liquidity Risk and Expected Stock Returns. Journal of Political Economy 111 (3), pp. 642–685. doi:10.1086/374184
This article explains a method; it is not investment or tax advice. Worked examples are simplified and rounded. All terms in the glossary.