Mean reversion

Definition

Mean-reversion strategies buy what has fallen unusually far and sell what has risen unusually far relative to a reference (its own average, a band, a paired instrument), on the premise that short-horizon price extremes partially reverse. It is momentum’s horizon-complement: reversal dominates at very short horizons (days to weeks) and very long ones (3-5 years); momentum owns the middle.

How it works / structure

• Reference and trigger: deviation from a moving average (dist_sma_pct), band touch (indicator-bollinger-bands), oscillator extreme (rsi_14), intraday VWAP distance (vwap_dist_pct), or a cross-sectional loser rank.
• Parameters (engine-executable): reference window, entry z-score/threshold, exit at mean vs opposite band, maximum holding time (mgmt-time-based-exit — reversion theses are time-bounded by construction), stop policy (see failure modes), regime filter (trend_state — fading a trend is the failure case).
• Portfolio form: many small independent reversion positions (the effect is statistical, not per-name reliable); strategy-pairs-trading is the hedged two-name version.

When it applies

Range-bound regimes and liquid instruments with mean-reverting microstructure; short holding windows; cross-sectional portfolios over single names. Explicitly NOT after information events — post-earnings moves underreact on average (event-earnings drift evidence), and “it fell a lot” is not a reversion thesis when the fall had a reason (qualitative-analysis falsifier discipline).

Risk profile & failure modes

• Catching the falling knife: the biggest losses come from averaging into a repricing trend — the strategy’s premise (temporary dislocation) is exactly wrong on real news; event filters are structural, not optional.
• Stop paradox: tight stops fight the entry logic (buying weakness), so risk is usually bounded by size and time rather than price stops — which makes sizing discipline (risk-fixed-fractional) carry the whole risk budget.
• Regime flips: reversion parameters tuned in ranges lose persistently when a trend regime starts (regime-volatility).
• Cost intensity: short horizons mean high turnover; edges are small per trade and friction-fragile (ms-slippage-friction).

Evidence & limits

Jegadeesh (1990) and Lehmann (1990) documented significant short-horizon (week-to-month) return reversal in US equities; De Bondt-Thaler (1985) documented multi-year loser-portfolio outperformance (long-horizon overreaction). Short-term reversal is partly compensation for liquidity provision and is heavily eroded by costs; net-of-friction profitability of simple public versions is questionable today and any specific parameterization is unproven until replayed. Band/oscillator folklore thresholds (“RSI 30 means bounce”) are labeled folklore.

Falsifiable-thesis examples

Illustrations only, not signals:

• “X, closing 2.5 standard deviations below its 20-day mean without a news event, will close back above the mean within 10 sessions” — falsified if it does not.
• “A daily cross-sectional reversal portfolio on universe U (buy bottom decile 5-day returns) will be net profitable after modeled friction this quarter in replay” — falsified by replay P&L.

Cross-references

• Horizon complement: strategy-momentum; hedged form: strategy-pairs-trading
• Triggers: indicator-bollinger-bands, indicator-rsi, indicator-vwap, indicator-sma
• Risk containment: mgmt-time-based-exit, risk-fixed-fractional, regime-volatility
• Event exclusions: event-earnings, lens-event-catalyst