A Comparative Analysis of Unsupervised AI Anomaly Detection Algorithms for Identifying Suspicious Trading Patterns Around U.S. Corporate Earnings Announcements

Abstract

Suspicious trading activity around U.S. corporate earnings announcements remains a central concern for the Securities and Exchange Commission. This study conducts a comparative evaluation of four unsupervised AI anomaly detection algorithms---Isolation Forest, One-Class Support Vector Machine (OC-SVM), Long Short-Term Memory (LSTM) Autoencoder, and UnSupervised Anomaly Detection (USAD)---applied to identifying suspicious timing patterns around quarterly earnings disclosures. The analysis is built on a twelve-feature multi-dimensional representation that combines price, volume, and order-flow indicators, computed over a [-20, +2] trading-day event window. Model fitting is fully unsupervised throughout; label information from a held-out validation split of the training period is used only at the threshold-calibration stage, and percentile-only operating points that do not consume any labels are reported alongside the F1-maximizing operating point for transparency. Ground-truth labels are constructed from 127 prosecuted insider trading cases, drawn from a four-stage pipeline applied to SEC Litigation Releases issued during the 2015--2023 sample period, with the case-selection conventions of a published academic sample (1996--2013) adopted purely as a methodological template. The main benchmark is reported on the subset of NASDAQ-listed S&P 500 constituents for which complete LOBSTER Level-1 order-book coverage was available in the study data environment, comprising 217 tickers, 6,884 event windows, 61 positives across the full period, and 22 positives in the 2021--2023 test split, with the full 14,416-window sample (46 test positives) used as a sensitivity analysis. Empirical results indicate that, within this sample, deep temporal detectors directionally favor classical baselines, with USAD attaining a balanced-threshold F1 of 0.58 (5-seed std 0.05) and ROC-AUC of 0.82, against 0.46 (std 0.04) and 0.72 for the Isolation Forest baseline. Bootstrap 95% confidence intervals overlap meaningfully across detectors, and the small number of test positives limits resolution, so the reported numbers should be interpreted as an exploratory benchmark rather than a definitive ranking. The study further quantifies the sensitivity-to-false-positive trade-off under three threshold configurations, reports flagged-window counts and false-positive rates so that downstream review burden can be assessed directly, and provides computational efficiency benchmarks at one-hour, one-minute, and one-second aggregation, yielding practical reference values for regulatory surveillance applications in which review capacity is the binding operational constraint.