Recency Bias in Portfolio Recommendations

Issue: Model Overweights Recent Performance; Recommends Assets That Performed Well Recently but Are Mean-Reverting

Frequency: Very Common

Symptoms

• Recommendations chase recent winners (high drawdown risk)
• Portfolio concentrated in recently-hot sectors
• Underrepresents assets in recent downturn (buying low missed)
• High portfolio turnover from chasing trends

Root Cause Recommendation models trained on historical return data develop implicit time-series momentum bias. Recent periods have higher data density; models weight them more. No explicit mean-reversion prior built in. Temporal data suggests “what went up will stay up” (false).

Example

Scenario: Robo-advisor portfolio construction
Tech sector: +40% in last 6 months
Model recommendation: Overweight tech (40% of portfolio)
Reality: Tech sector mean-reverting; subsequent 6 months: -15%
Client loss: Significant underperformance
Impact: Poor risk-adjusted returns; client dissatisfaction

Key Statistics

• Win rate (positive returns in subsequent period): 40-55% (barely >50%)
• Concentration in recent winners: 30-50% of portfolio often in top 3 performers
• Turnover: 50-100% quarterly (transaction costs erode returns)

Mitigation Strategies

1. Mean-Reversion Prior: Explicitly penalize recent strong performers
2. Longer Lookback: Use 3-5 year data, not 1-2 year
3. Fundamental Factors: Base recommendations on earnings, valuations, not price momentum
4. Turnover Penalties: Penalize high turnover in optimization

Metrics

• Recommendation accuracy in forward period (>50% = good)
• Portfolio turnover (lower is better)
• Concentration (lower is better)

Alerts

• Turnover >50% quarterly → Review model bias
• Win rate <55% → Model underperforming

References

• Momentum and Reversal in Asset Returns
• Behavioral Bias in Algorithmic Trading