Recency bias

Definition

Recency bias is overweighting recent observations when judging what is normal or likely — availability (Tversky-Kahneman 1973) applied to time series. Its market form is EXTRAPOLATION: Greenwood-Shleifer (2014) showed survey-measured return expectations track RECENT past returns and correlate NEGATIVELY with subsequent returns — the crowd expects most at tops and least at bottoms, measurably. It is the bias that turns regimes into surprises.

How it works / structure

• The mechanism: recent, vivid observations are easier to retrieve, so they dominate frequency judgments; in markets, the lookback that “feels representative” shrinks toward whatever just happened — vol estimates after calm years, return expectations after rallies, tail estimates after tail-free samples.
• The documented market forms: extrapolative expectations (Greenwood-Shleifer — surveys vs realized), return-chasing flows (sent-fund-flows — the same shape in money movement), post-crisis over-hedging and late-cycle under-hedging (insurance demand tracks recent pain, not forward risk), and strategy abandonment at the trough of its cycle (documented in factor-timing behavior — strategy-factor-investing value droughts).
• The model analogue (engine-relevant): short training windows, recency-weighted fits, and regime-blind hyperparameters encode the same bias in code; walk-forward evaluation across MULTIPLE regimes is the mechanical counter (lens-quantitative).
• Platform counters: regime-stratified base rates (statistics quoted per regime, with the current regime’s age displayed — regime-volatility), lookback-window audits (does the estimator’s window include a stress episode?), and expectation-vs-history dashboards (what the thesis assumes vs full-sample and regime-matched bases).

When it applies

Every estimate with a lookback window — vol, correlation, expected return, hit rate; strategy evaluation timing (abandonment/adoption decisions cluster at exactly the wrong points per the flow evidence); post-streak state (both hot and cold streaks reshape felt probabilities — bias-overconfidence interaction).

Risk profile & failure modes

• Calm-regime tail amnesia: sizing to recent realized vol after a quiet year is the canonical institutional form (risk-volatility-targeting shares the limit — its estimator IS a lookback).
• Post-crisis overcorrection: the mirror error — pricing the last crisis into everything for years after (documented in post-2008 hedging demand).
• Narrative freshness: the latest explanatory story crowds out base rates (qualitative-analysis folklore discipline).
• Un-fixable by awareness: knowing about recency bias does not remove it (documented for biases generally) — counters must be structural (windows, stratification), not aspirational.

Evidence & limits

Availability is foundational cognitive psychology; Greenwood-Shleifer (2014) is the peer-reviewed market measurement (six survey sources, consistent extrapolation); return-chasing flows corroborate at the aggregate. The bias is population-level with wide individual variance; platform counters are mechanical estimator-design rules rather than user warnings.

Falsifiable-thesis examples

Illustrations only, not signals:

• “Survey/positioning-implied expected returns (currently top-decile optimistic) will correlate negatively with the index’s next-3-year realized return (Greenwood-Shleifer pattern)” — falsified by the long-window pair.
• “Vol estimates using a 1-year lookback will under-forecast next-quarter realized vol more often after calm years than after volatile ones (window-bias check)” — falsified by the stratified forecast errors.

Cross-references

• The crowd expression: sent-fund-flows, bias-herding
• The state it blinds: regime-volatility, regime-seasonality (small samples amplify it)
• The self-reinforcing pair: bias-overconfidence
• The mechanical counters: lens-quantitative (walk-forward, regime stratification)