RESEARCH
APRIL 2026 · WP BF-2026-01
↓ PDF
N = 1,000 TRADERS · 24 MONTHS
RESEARCH
In a 24-month exclusive observational study of 1,000 prop traders, AI-assisted traders showed lower drawdowns, fewer emotionally-driven exits, higher Sharpe ratios, and lower rule breach rates. This paper identifies seven cognitive biases responsible for the performance gap.
Based on Kahneman & Tversky's Prospect Theory (1979), AIProp's Behavioral Bias Index (BBI) operationalises seven measurable biases from platform trade-log data. Each bias has a direct, quantifiable cost to prop trader outcomes.
Primary outcomes from the 24-month observational study (April 2024 – March 2026). Statistical methods: independent samples t-test, chi-square, Wilson score confidence intervals. Significance threshold: p < 0.05 two-tailed.
| Metric | Manual (n=490) | AI-Assisted (n=510) | Difference |
|---|---|---|---|
| Mean Sharpe Ratio | 0.62 (SD 0.41) | 0.89 (SD 0.35) | +0.27 (+44%) |
| Avg Max Drawdown | 7.8% (SD 2.9%) | 4.3% (SD 1.8%) | −3.5 pp (−45%) |
| Loser/Winner Hold Ratio | 4.1× (CI 3.8–4.4×) | 1.7× (CI 1.6–1.8×) | −2.4× (−59%) |
| Rule Breach Rate | 18.4% (CI 15.1–22.1%) | 12.2% (CI 9.5–15.4%) | −6.2 pp · p<0.01 |
| Emotional Exits | 61.7% of exits | 37.2% of exits | −24.5 pp · p<0.001 |
| Profit Factor | 1.21 (SD 0.38) | 1.58 (SD 0.29) | +0.37 (+31%) |
| Risk Adherence Index (RAI) | 61.4% | 88.9% | +27.5 pp (+45%) |
The AI-assisted cohort comprises three meaningfully different sub-types. The Hybrid AI + Human Oversight configuration — combining automated risk enforcement with human thesis judgment — achieved the best outcomes across all metrics.
| Sub-Cohort | N | Breach Rate (95% CI) | Sharpe | RAI |
|---|---|---|---|---|
| Manual (reference) | 490 | 18.4% (15.1–22.1%) | 0.62 | 61.4% |
| AI-Assisted Discretionary | 183 | 15.1% (10.4–21.1%) | 0.81 | 79.2% |
| Rule-Based EA | 198 | 13.6% (9.5–18.8%) | 0.91 | 91.3% |
| Hybrid AI + Human Oversight | 129 | 8.5% (4.8–14.4%) | 0.97 | 94.1% |
The Hybrid sub-cohort achieved an 8.5% breach rate — less than half the manual cohort's 18.4%. This is consistent with the hypothesis that AI risk enforcement combined with human thesis judgment is the optimal configuration — structural rules make emotional patterns impossible, while the human retains edge in entry selection.
The BBI is a composite score (0–100) aggregating seven sub-scores derived entirely from platform trade-log data. Higher scores indicate stronger bias expression. Weights were calibrated via logistic regression on breach events (training dataset 2022–2023, N = 3,400 accounts).
| BBI Component | Operational Measurement | Weight |
|---|---|---|
| Disposition Score | Loser/winner hold time ratio per session | 22% |
| Loss Aversion Score | % of trades where stop-loss was moved further after entry | 20% |
| Overconfidence Score | Trade frequency deviation from 30-day baseline + position size variance | 18% |
| Anchoring Score | Target Override Rate: % of trades with declared exit target but manual close elsewhere without thesis change | 14% |
| Mental Accounting Score | Position size uplift after session profit > 1.5× daily average | 12% |
| Herding Score | % of entries post-momentum (after 1.5× ADR move, no prior plan) | 8% |
| Revenge Trading Score | % of entries within 30 min of stop-out, equal or larger position size | 6% |
Disposition effect 2.7× more severe than academic benchmarks. Manual cohort loser/winner hold ratio of 4.1× vs Odean (1998) benchmark of ~1.5×. For $50K–$100K accounts, the implied annual performance cost is $2,200–$4,400.
73% of breach events in the manual cohort (66/90) were preceded by a BBI-tagged behavioral event — loss aversion stop-removal, revenge entry, or mental accounting oversize — in the same session. Strategy is not the primary problem. Discipline is.
AI-assisted trading associated with 41% fewer emotionally-driven exits: 37.2% vs 61.7% (−24.5 pp; chi-square = 142.3; p < 0.001). Improvement was largest in Rule-Based EA and Hybrid sub-cohorts where execution rules make emotional patterns structurally impossible.
AI-assisted traders associated with 45% lower drawdowns (4.3% vs 7.8%) and a 34% lower rule breach rate. Hybrid sub-cohort achieved 8.5% breach rate vs 18.4% for manual — consistent with AI risk enforcement combined with human thesis judgment being the optimal configuration.
Risk Adherence Index (RAI) is the strongest single behavioral predictor of account outcomes (r = 0.74, p < 0.001). Mean RAI 88.9% (AI) vs 61.4% (manual; +27.5 pp). AI's primary mechanism is structurally enforcing the risk adherence that manual traders fail to maintain under emotional pressure.
Get the full PDF of Working Paper BF-2026-01 — including the complete BBI framework, sub-cohort analysis, statistical methodology, and all five evidence-based conclusions from AIProp's 24-month study of 1,000 prop traders.