Concepts5 min read

Mean Reversion vs Trend Following

Mean reversion and trend following are the two foundational regimes a price series can exhibit, and every systematic strategy implicitly bets on one or the other. Mean reversion assumes prices oscillate around a stable equilibrium; trend following assumes prices exhibit serial correlation strong enough to ride. Understanding which regime your instrument is in — and over what timescale — is upstream of strategy selection itself.

The underlying statistical distinction

Both regimes can be framed through the autocorrelation of returns. A trending series has positive autocorrelation: today's return predicts tomorrow's in the same direction. A mean-reverting series has negative autocorrelation: today's return predicts a reversal.

ρ(k) = Cov(r_t, r_t+k) / Var(r_t)

A more rigorous test is the Hurst exponent H, which measures how variance scales with time horizon. For a random walk, variance scales linearly with time and H = 0.5. For trending series, variance scales faster and H > 0.5. For mean-reverting series, variance scales slower and H < 0.5.

E[|X(t+τ) - X(t)|²] ∝ τ^(2H)

The Ornstein-Uhlenbeck process formalizes mean reversion explicitly, with a parameter θ controlling speed of reversion toward a long-run mean μ:

dX_t = θ(μ - X_t)dt + σ dW_t

How to interpret the regime

For Hurst values, the practical thresholds are: H between 0.45 and 0.55 indicates near-random behavior — neither strategy class has structural edge. H below 0.4 indicates strong mean reversion, typical of stationary spreads, pairs residuals, and intraday equity returns. H above 0.6 indicates persistent trending, common in commodities, FX carry, and equity indices over multi-month horizons.

Autocorrelation at lag 1 follows a similar logic. Significantly negative ρ(1) on daily equity returns is well-documented and drives short-term reversal strategies. Significantly positive ρ(1) on weekly-to-monthly returns in trend-prone assets drives the classic time-series momentum literature.

Critically, the same instrument can be mean-reverting on one timescale and trending on another. SPX exhibits short-horizon reversal (1-5 day), medium-horizon momentum (3-12 month), and long-horizon reversal (3-5 year). A strategy that conflates these timescales will trade against itself.

Mean reversion strategies typically show high win rate, low average win size, and fat left tails. Trend following shows low win rate, large average win size, and fat right tails. The P&L distributions are near-mirror images, which is why combining them at the portfolio level often improves risk-adjusted returns more than either alone.

What this framing does not capture

Regime classification via Hurst or autocorrelation assumes stationarity of the regime itself. In practice, regimes shift — a mean-reverting pair can break and trend indefinitely once the cointegrating relationship fails. No backward-looking statistic tells you when this transition occurs.

Neither metric captures execution cost asymmetry. Mean reversion strategies trade against momentum, which means they tend to provide liquidity and earn the spread, but they accumulate losses precisely when liquidity disappears. Trend following crosses the spread to chase, paying cost on every entry, but its P&L tail is positively skewed when volatility spikes.

The framing also obscures the source of the edge. A trend-following P&L can come from genuine serial correlation, from a volatility risk premium, or from a behavioral underreaction effect — these have very different decay properties. Mean reversion can come from inventory effects, microstructure noise, or fundamental anchoring, again with different decay properties. Two strategies with identical autocorrelation profiles can have entirely different out-of-sample longevity.

Hurst exponent estimates are noisy on short windows and biased on series with structural breaks. A single point estimate of H = 0.42 does not establish mean reversion — confidence intervals on rolling Hurst are typically wide enough to span both regimes.

How Kestrel Signal surfaces this

Kestrel Signal computes rolling Hurst, lag-k autocorrelation, and variance-ratio tests on every loaded series, and exposes them as features available to strategy logic. Regime statistics are also reported alongside backtest results, so you can see whether the in-sample period was structurally mean-reverting or trending. This makes it explicit when a strategy's reported performance is regime-dependent rather than robust across both states.