Overconfidence

Definition

Overconfidence is systematic overestimation of one’s own information, skill, and precision — the best-documented behavioral driver of trading losses. Its market signature is OVERTRADING: Barber-Odean (2000) showed the most active quintile of retail accounts underperformed the least active by ~6.5%/year, with friction converting confident churn into measured wealth transfer. For an AI-agent platform the bias is doubly relevant: it lives in users, and its analogue (overfit conviction) lives in models.

How it works / structure

• The evidence chain: activity predicts underperformance (Barber-Odean 2000); demographic overconfidence proxies predict activity (2001 — the gender study); post-success trading escalates (documented self-attribution — wins are skill, losses are noise); precision overconfidence shows as intervals too narrow (forecasters’ 90% intervals contain reality far less than 90% of the time — documented across domains).
• Platform counters (engine-executable): friction-share accounting per account (cost drag as % of returns — the Barber-Odean diagnostic), trade-frequency baselining vs the strategy’s replay cadence (churn above the strategy’s own design frequency is the flag), forced falsifier declaration before entry (qualitative-analysis — conviction without a falsifier is the bias in its natural habitat), and calibration scoring (predicted vs realized hit rates, tracked per user and per model).
• The model analogue: in-sample confidence, out-of-sample humility (lens-quantitative overfitting discipline is overconfidence-proofing for machines).

When it applies

Every account and every model — this is the platform’s default-on behavioral diagnostic; win-streak periods especially (self-attribution compounds after success — bias-recency amplifies); high-frequency discretionary styles where the documented damage concentrates (strategy-day-trading-styles).

Risk profile & failure modes

• The bias measuring itself: users (and builders) exempt themselves — the evidence is ABOUT the population reading it; diagnostics must be mechanical, not self-assessed.
• Confusing conviction with edge: strength of feeling has zero documented correlation with outcome accuracy; the platform’s thesis format exists to replace felt conviction with stated falsifiers.
• Overcorrection: paralysis-by-humility (never sizing up on genuinely replayed edges) is the mirror error — calibration, not minimization, is the target.

Evidence & limits

Barber-Odean’s account-level studies are the canonical field evidence; calibration and self-attribution literatures are broad and replicated. Effects are population-level — individual exceptions exist and cannot identify themselves ex ante, which is precisely the bias. Platform diagnostics measure behavior (frequency, friction, calibration), never intent.

Falsifiable-thesis examples

Illustrations only, not signals:

• “Accounts in this platform’s top activity decile will show lower net-of-friction returns than the median-activity decile over the year” — falsified by the cohort accounting.
• “My stated 80%-confidence theses resolve true at less than 70% (calibration check)” — falsified by the tracked score.

Cross-references

• The amplifiers: bias-recency (fresh wins), bias-disposition-effect (the exit-side companion)
• The damage habitat: strategy-day-trading-styles
• The mechanical counters: qualitative-analysis (falsifiers), risk-fixed-fractional (sizing that ignores conviction), lens-quantitative (the model analogue)