Behavioral Model Interactive
Recency Bias
Users overweight recent performance when predicting future outcomes. Exploit this by identifying regression to the mean opportunities.
🧠 The Cognitive Bias
What People Do
Overweight the most recent data points. A player's last game feels more predictive than it actually is.
- • "He just dropped 40, he's on fire!"
- • "He's cold, only 12 pts last game"
- • "This team can't lose, 5 game win streak"
What Actually Happens
Performance regresses to the mean. Hot and cold streaks are mostly random variation.
- • Variance in small samples is high
- • True ability changes slowly
- • Outliers are... outliers
Player Stats
10 35
10 45
5 60
📊 Projection Comparison
Public Projection
33.2
Heavy recency weighting
Optimal Projection
26.4
Regression-adjusted
Mispricing
+6.8 pts
Fade OVER
Weighting Schemes
Public (Recency Biased)
Season Avg 20%
Last 5 Games 40%
Last Game 40%
Optimal (Regression-Adjusted)
Season Avg 60%
Last 5 Games 30%
Last Game 10%
Simulated Regression to Mean
Hot streaks regress toward season average, not stay at elevated levels.
📋 Common Scenarios
Hot Streak
Recent: +35% vs avg
Public thinks: Will keep it up!
Reality: Likely regression
Edge: Fade the over
Cold Streak
Recent: -25% vs avg
Public thinks: Lost it, done
Reality: Likely bounce back
Edge: Take the over
Blowup Game
Recent: 50 pts (2x avg)
Public thinks: New level!
Reality: Outlier, regress
Edge: Strong fade
Injury Return
Recent: 15 pts (low)
Public thinks: Not same player
Reality: Ramping up
Edge: Buy the over
💰 Pricing Implications
Exploit Recency Bias
- → Shade lines toward public perception (let them bet wrong side)
- → Boost props for players on cold streaks (public under-bets)
- → Reduce props for hot players (public over-bets)
For Your Models
- → Use Bayesian shrinkage to weight season over recent
- → Calibrate optimal recency weights empirically
- → Identify situations where recent IS more predictive (injuries, role changes)
R Code Equivalent
# Regression-adjusted projection
project_with_regression <- function(season_avg, last5_avg, last1,
weights = c(0.6, 0.3, 0.1)) {
projection <- season_avg * weights[1] +
last5_avg * weights[2] +
last1 * weights[3]
return(projection)
}
# Compare public vs optimal
public_proj <- project_with_regression(22, 30, 42,
weights = c(0.2, 0.4, 0.4))
optimal_proj <- project_with_regression(22, 30, 42,
weights = c(0.6, 0.3, 0.1))
mispricing <- public_proj - optimal_proj
cat(sprintf("Public: %.1f, Optimal: %.1f, Edge: %.1f pts\n",
public_proj, optimal_proj, mispricing))✅ Key Takeaways
- • Humans overweight recent performance
- • Hot streaks regress, cold streaks bounce back
- • Season average is more predictive than last game
- • Fade extreme recent performance
- • Use Bayesian shrinkage for optimal weighting
- • Exception: genuine role/situation changes