Overconfidence and Calibration

## The Overconfidence Problem Overconfidence is perhaps the most pervasive cognitive bias in betting: the tendency to be more certain about the accuracy of your judgments than the evidence warrants. Research across populations consistently shows that humans are poorly calibrated: when people say they are "90% confident" in a prediction, they are correct less than 75% of the time. In betting, overconfidence manifests as: - Underestimating the probability of upset results - Staking more than Kelly recommends for your actual edge - Viewing losing runs as anomalies rather than expected outcomes ## Measuring Your Calibration A calibration test: over your last 200 bets, group selections by your confidence level: | Stated confidence | Actual win rate | |---|---| | 60% | X% | | 65% | Y% | | 70% | Z% | If your stated confidence consistently exceeds your actual win rate, you are overconfident. If it matches (or understates), you are well-calibrated. ## Correcting for Overconfidence **Shrink your probability estimates toward 50%:** If your model says 70% win probability, consider using 63% in Kelly calculations (applying a 30% shrinkage factor toward 50%). This corrects for systematic overconfidence. **Compare to closing line:** If the market consistently closes at a lower implied probability than your pre-market estimate, the market is correcting for your overconfidence. **Track raw model output vs outcome:** Build a calibration chart from historical model predictions. Where model exceeds actual frequency: apply a calibration correction. Where model matches: trust the output. ## The Elite Level Standard The best bettors are neither overconfident nor underconfident — they have well-calibrated uncertainty. A 60% estimate is correct approximately 60% of the time. This calibration takes years of deliberate tracking to achieve and is one of the clearest markers of analytical maturity.