A Cognitive Digital Twin for Detecting and Explaining Behavioral Biases in Retail Investing: System Design and Evaluation

Abstract

New retail investors are exposed to behavioral biases that prior studies associate with lower net returns. This paper presents a Cognitive Digital Twin (CDT) that detects three of them, the disposition effect, overconfidence, and loss aversion, while a person trades in a simulated market, then explains them back to the user. It records every decision in a fourteen-round simulation over twelve Indonesia Stock Exchange (IDX) stocks, scores three indicators that operationalize established behavioral finance constructs, each with a 95% confidence interval from 500 bootstrap resamples, and updates a per-user profile that drives plain-language feedback graded by severity. Scoring is deterministic: fixed formulas, not a learned or anomaly-detection model. Because an investor’s true bias is unobservable, the detector cannot be graded against a known answer. We evaluate it two ways. First, synthetic investors with deliberately planted biases supply the ground truth real data lacks. Recovery reaches rank correlation 0.97 for overconfidence, 0.92 for the disposition effect, and 0.77 for loss aversion once three sessions are combined, up from 0.82 and 0.63 for the latter two on a single session. Second, a study with 16 novice investors finds the overconfidence score agrees best with users’ self-assessment (ρ = 0.71, n = 11), and that returning users’ profiles are stable. Both checks point to overconfidence as the most reliably detected bias.