When bots mislead markets: asymmetric contamination risk in sentiment-based ınvestment decisions

Abstract Social media sentiment analysis has become central to understanding investor behavior, yet Large Language Model (LLM)-generated bots may systematically distort these measurements. This study examines how different bot behaviors create measurement biases that mislead investment decisions by contaminating perceived investor sentiment. Using a fixed test set simulation design—where models are trained on clean historical data but exposed to bot-contaminated streams at inference time—applied to 1091 financial tweets and 320 independent tests across four bot scenarios (benign paraphrase, positive amplification, negative amplification, and noisy bots), we document highly asymmetric contamination effects. Negative amplification induces critical platform distortions at only 20% bot penetration, while positive amplification requires 40%, revealing a 2:1 tipping-point asymmetry in critical contamination thresholds. This pattern aligns with negativity bias documented in behavioral finance: Investors and sentiment models disproportionately weight negative information. Critically, we find a decoupling phenomenon: model performance metrics (F1-score) remain stable even as measurement error increases substantially, suggesting firms may fail to detect contamination using standard evaluation frameworks. These empirically derived thresholds (20–40% bot penetration) offer actionable benchmarks for firms deploying sentiment-based investment tools, and highlight the need for multi-metric, contamination-aware evaluation frameworks in behavioral finance research.