Momentum

Momentum is the empirical regularity that asset prices which have risen over the past several months tend to keep rising for several more before reversing. It is the strongest and most robust failure of the efficient markets hypothesis documented in modern asset pricing — surviving every standard risk adjustment, every continent, every asset class, and (in trend-following form) more than two centuries of out-of-sample data.

Momentum strategies come in two canonical forms that share the same cognitive story but differ in construction:

• Cross-sectional momentum (Jegadeesh & Titman, 1993) ranks assets against each other and trades a long-short winners-minus-losers (WML) portfolio. Returns about $1\%$ per month with ${\beta }_{\text{WML}}\approx -0.08$ on the U.S. equity cross-section.
• Time-series momentum (Moskowitz, Ooi & Pedersen, 2012) trades each asset on the sign of its own past return, with positions sized to a constant volatility target and aggregated across asset classes. Every one of $58$ liquid futures contracts has positive Sharpe; the diversified portfolio Sharpe is above $1$.

Both strategies share the same failure mode: momentum crashes of $-50\%$ to $-75\%$ in a single month, driven by a sharply negative state-dependent beta in early-rebound regimes that the long-run average $-0.08$ hides entirely.

Decision rule. Use cross-sectional momentum when trading a single liquid universe (e.g., U.S. equities) and the goal is a market-neutral spread bet. Use time-series momentum when trading a diversified cross-asset book where the directional exposure of each contract carries useful information. In either case, condition on a panic-state indicator — neither static strategy is safe to run unhedged at scale.

The leading explanation is the two-stage over-reaction story: investors are slow to incorporate news (creating $1$–$12$ month continuation), then late buyers push prices past fundamentals (creating $13$–$36$ month reversal). Any complete behavioral model must produce both halves jointly. See BSV and Hong-Stein for the two leading candidates.

Intuition

Momentum is the embarrassment of efficient-markets finance. Standard asset pricing says past returns should not predict future returns: if they did, traders would exploit the predictability and arbitrage it away. And yet for nearly a century, in every market that has been tested, the simple rule “buy what has gone up, sell what has gone down” has earned a positive risk-adjusted return that no rational risk model has been able to absorb.

The two canonical forms differ only in what the past return is measured against. Cross-sectional momentum compares each stock to its peers — a stock is a “winner” if it outperformed the rest of the universe, regardless of whether the market itself rose or fell. Time-series momentum compares each asset to its own past — a contract is in an uptrend if its own price went up, regardless of what its peers did. The two are correlated but distinct (correlation around $0.6$ in equities, lower elsewhere): cross-sectional momentum is a spread bet on the gap between winners and losers within a universe, while time-series momentum is a directional bet on each individual asset’s trend.

The shared cognitive story (formalized in BSV and Hong-Stein) has two stages. First, investors are slow to absorb news — from conservatism, slow information spread, or limited attention — so prices drift in the direction of the news for several months instead of jumping right away. This drift is what the $1$–$12$ month momentum strategies capture. Second, late buyers pile in and push prices past what the news justified, generating the long-term reversal first documented by De Bondt & Thaler (1985) over months $13$–$36$. Continuation and reversal are two halves of one trajectory.

The dark side is that momentum’s safety margin is illusory. The average market beta of cross-sectional momentum is essentially zero, which makes it look like a free lunch by standard risk metrics. But in panic states the conditional beta becomes sharply negative — Daniel & Moskowitz (2016) estimate values around $-2$ — and a sharp rebound blows up both legs at once. The two largest U.S. crashes (August $1932$ and April $2009$) account for a substantial fraction of all the variance in the entire $1927$–$2013$ WML series. Anyone running unhedged momentum at scale must either size to survive a single-month $-75\%$ drawdown or use a dynamic strategy that cuts exposure in panic states.

Comparison

When to Use Which

Cross-sectional momentum is the right tool when the goal is a market-neutral spread on a single liquid universe. The construction is naturally hedged against the level of the market — buying winners and shorting losers in the same market cancels out direct market exposure — and the strategy plugs cleanly into the existing factor-investing toolkit, where it shows up as the Carhart UMD factor. Run cross-sectional momentum when the universe is large enough to support a meaningful decile spread (a few hundred names minimum), when shorting is cheap, and when the goal is alpha relative to a passive benchmark in that universe.

Time-series momentum is the right tool when the goal is a diversified cross-asset trend book. The strategy works on every contract individually, so it can be run across equities, bonds, currencies, commodities, and credit at the same time, with the diversification across asset classes doing most of the work. The volatility scaling is essential here: without it, the wildest contract dominates the portfolio variance and the cross-asset Sharpe collapses. Run time-series momentum on a diversified futures book of $30$+ contracts, with per-contract volatility targets around $40\%$ annualized so the diversified portfolio sits at a sensible $10$–$15\%$ realized volatility.

In either case, the long-run Sharpe ratio is a misleading guide to risk. Both strategies carry sharp tail risk that is invisible in the long-run averages, and both have produced single-month losses of $-30\%$ or worse in identifiable panic states. The Daniel-Moskowitz dynamic strategy roughly doubles the cross-sectional Sharpe by cutting exposure in late-bear-market regimes, and an analogous regime filter is standard practice in time-series CTA programs. Always condition — neither strategy is safe to run unhedged at full size through a panic state.

If only one momentum book can be run, the right answer depends on the firm’s other exposures. A long-only equity manager benefits more from adding a market-neutral cross-sectional momentum overlay, because it diversifies the existing directional exposure. A multi-asset macro book benefits more from time-series momentum, because the directional trend information adds to the existing macro signal stack. Many large quant shops run both, sized down individually because the two strategies are correlated.

Explanations for Momentum

The empirical fact of momentum has been replicated so thoroughly that the live debate is no longer about whether it exists but about why. Three families of explanation compete:

Behavioral under-/over-reaction. The leading account, due to Barberis, Shleifer & Vishny (1998) and Hong & Stein (1999). BSV builds a single-investor model whose agent oscillates between believing earnings follow a mean-reverting regime and a trending regime; both regimes generate short-run continuation followed by long-run reversal. Hong-Stein builds a model with two kinds of agents — “newswatchers” who under-react to private information and “momentum traders” who chase past returns — and shows their interaction generates the same joint pattern. Both models predict $1$–$12$ month continuation and $13$–$36$ month reversal jointly.

Risk-based. A succession of authors have tried to rationalize momentum as compensation for some omitted risk factor — conditional CAPM, consumption beta, downside risk, liquidity risk, default risk. None has succeeded in absorbing the full WML alpha across all the samples and asset classes where momentum has been documented. The failure of risk-based explanations across two centuries and dozens of markets is the strongest single argument against the rational paradigm in asset pricing.

Structural / flow-based. Lempérière et al. ($2014$) and others observe that trend-following behavior is itself amplified by flows: because so much of professional asset management runs trend strategies, flows from trend-following funds push every nascent move further along, creating positive feedback at exactly the horizons where momentum is empirically strongest. This story is complementary to behavioral under-reaction rather than a substitute: it explains why momentum has persisted in the post-publication era even as more capital has chased the signal.

The three accounts are not mutually exclusive. The most defensible synthesis is that initial momentum is generated by genuine cognitive under-reaction to information (BSV / Hong-Stein), is amplified by trend-following flows in the middle of the trajectory, and is eventually reversed when the overshoot becomes large enough that fundamental traders push back. Momentum crashes are the visible scar of the synchronized unwind when the flow dynamics flip in a panic-state rebound.

See Also

• 8.8.1. Cross-Sectional Momentum — child: rank-based long-short on a single universe.
• 8.8.2. Time-Series Momentum — child: trend-following on each asset against its own past.
• 8.8.3. Momentum Crashes — child: the state-dependent-beta failure mode shared by both.
• 8.9. Reaction Models — parent of the leading behavioral explanations.
• 8.9.1. Barberis-Shleifer-Vishny Model — joint continuation/reversal from regime confusion.
• 8.9.2. Hong-Stein Model — joint continuation/reversal from heterogeneous agents.
• 8.3. Challenges to Rational Pricing — momentum is the flagship anomaly listed there.
• 8. Behavioral Finance — top-level hub.