Basic Prerequisites: Probability & Measure Theory (skewness, kurtosis, tail risk) and Econometrics & Time Series (conditional moments, regime models).
1. Intuition & Practical Objective
Momentum’s average return is attractive, but its payoff is violently negatively skewed. Daniel & Moskowitz (2016) find the US WML monthly log-return skewness is (1927–2013): the strategy quietly earns most of the time, then loses a lot in a few clustered months. This page is about those months - what causes them, why they are partly forecastable, and how practice copes with crashes, crowding, and the 2009 reversal.
The two canonical crashes (all verified corpus numbers):
| Episode | Market context | What happened to WML |
|---|---|---|
| July–Aug 1932 | market rose +82% in two months | losers vs winners ; WML , |
| 2009 reversal (Mar/Apr/Aug) | market rose +26% over the rebound | losers vs winners ; WML (Mar), (Apr), (Aug) |
The crash is a “loser crash-up,” not a market crash. Momentum does not lose because the market fell - it loses when the market rebounds sharply after a decline. Fourteen of the 15 worst momentum months follow a negative two-year market return and occur in months the market rose contemporaneously.
2. Mathematical Ground Truth & Derivations
The mechanism: time-varying beta of return-sorted portfolios. Past-return-sorted portfolios have strongly time-varying exposure to the market (Kothari & Shanken 1992; Grundy & Martin 2001). After a large market decline, the firms that fell with the market were mostly high-beta, and those that held up were low-beta. So following a decline the momentum portfolio is long low-beta (winners) and short high-beta (losers):
Daniel & Moskowitz estimate loser-decile betas can rise above 3 (winners fall below 0.5). The WML monthly return in state with market return is
When the market rebounds, and the negative beta spread makes - the crash. In bear markets the up-beta is more than double the down-beta ( vs , -stat 4.5), so the payoff is asymmetric and option-like: the short losers behave like a written call on the market - little gain when it falls, large loss when it rises.
Forecastability. Crashes cluster in a definable “panic state”: (a) market down over the trailing ~2 years, (b) high ex-ante volatility, (c) a contemporaneous sharp rebound. This state can be detected in real time from trailing market returns and VIX-type volatility - which is exactly what makes the dynamic/volatility-managed strategies of 06 possible.
2009 as a special case. The 2009 crash was not a fast single-month event; it was the momentum counterpart of the “short-vol/quality” crowding unwind. Past losers were the levered financials (Citigroup, Bank of America, Ford, GM) whose equity was, in the Merton (1974) sense, an out-of-the-money call on firm value; past winners were defensive/quality names. When the market ripped up, the option-like losers crashed up far faster than winners. The AQR view (Asness et al. 2014): momentum did not suffer during the crash - it suffered in the sharp reversal from late March / early April 2009.
3. Computational Implementation - simulating a momentum crash
Standard library only. Constructs a panic state: after a market decline, loser beta spikes (1.8) and winner beta collapses (0.5); a violent rebound then crushes WML. Verifies the crash is (a) in the rebound month, (b) negatively skewed, (c) preceded by a market decline.
The worst WML month is , occurring exactly in the rebound month, after a two-year market decline, driven by loser beta (1.80) far above winner beta (0.50). The monthly skewness is - the crash signature.
4. Failure Modes & First-Principles Breakdowns
- Momentum crashes (negative skewness). The dominant risk. Crashes are clustered, slow, and partly forecastable - not Poisson jumps. They are driven by the option-like short-loser leg in panic states. Do not hedge by buying S&P variance swaps in bear markets: Daniel & Moskowitz show variance hedging does not restore momentum profitability in bear states.
- The 2009 crowding reversal. When everyone runs the same momentum/factor trade, the unwind becomes synchronized: the reversal in early 2009 (crowded short losers + crowded long quality) produced momentum’s worst modern drawdown. Crowding means momentum’s own popularity is a risk input.
- Whipsaw / turning-point drawdown. Distinct from the fast crash: 12–18 months of buying tops and shorting bottoms in a mean-reverting regime - slow capital bleed, harder to detect on a Sharpe basis.
- Hedged-momentum illusion. Grundy & Martin’s market/size-hedged momentum looked excellent, but their hedge used ex-post/forward-looking betas; an implementable ex-ante hedge does not avoid the crashes (Daniel & Moskowitz 2016). Any backtest claim that “hedging fixes momentum” must be checked for look-ahead bias.
- Transaction costs & short-leg capacity. The loser leg is high-turnover and often illiquid/small; shorts are expensive and can be impossible in a squeeze. Live costs (AQR, 1998–2013) are manageable but do not vanish.
- Multiple testing. With the historical record mined at every horizon and market, the published momentum Sharpe overstates tradability - deflate it (see Backtesting Hygiene).
5. References
- Daniel & Moskowitz (2016), Momentum Crashes, J. Financial Economics 122(2)
- Grundy & Martin (2001)
- Barroso & Santa-Clara (2015), Momentum Has Its Moments
- Asness, Frazzini, Israel & Moskowitz (2014), Fact, Fiction, and Momentum Investing
6. Connected Graph Bridges
- Base: Probability & Measure Theory (skewness/kurtosis) · Econometrics & Time Series
- Remedies: 06 · Advanced Extensions (vol scaling, dynamic weighting) · 04 · Value–Momentum
- Risk: ES) · Extreme Value Theory & Fat Tails
- Hygiene: Backtesting Hygiene & Deflated Sharpe