Insider Trading Activity Detection Using One-Class Support Vector Machine in Regulatory Frameworks

Securities regulators monitoring trading activity for insider trading face a class-imbalance problem several orders of magnitude more extreme than typical fraud-detection applications: confirmed insider trading cases represent a tiny fraction of total monitored activity, rendering supervised classification trained on labelled positive examples statistically unreliable. One-Class Support Vector Machines (OC-SVM), introduced by Schölkopf et al. (2001), address this by learning a boundary enclosing the region of normal trading behaviour from unlabelled (presumptively normal) data alone, flagging departures from that boundary as anomalous regardless of whether any confirmed positive examples informed training. This paper applies OC-SVM (RBF kernel;$\gamma=0.08)$to account-event observations surrounding 5,760 material corporate announcements across 480 monitored companies over 36 months, yielding 864,000 account-event observations (5,760 events × mean 150 active accounts per event window) of which 184 are confirmed insider trading cases drawn from completed enforcement actions, used exclusively for post-hoc validation rather than training. Eight behavioural features-trade-size anomaly, announcement-timing anomaly, post-trade price movement, holding-period anomaly, cross-account correlation, historical-deviation Z-score, options-volume surge, and network-proximity score-are extracted per observation. At a standard operating tier$(\boldsymbol{v} \boldsymbol{=} \mathbf{0. 0 1}$, flagging 8,640 observations), OC-SVM recovers 126 of the 184 confirmed cases (68.4% recall), outperforming Z-score thresholding (34.8% recall; +96.6% relative), Isolation Forest (56.5%; +21.1%), and Local Outlier Factor (62.0%; +10.3%) at an identical flagging budget. Feature ablation identifies announcement-timing anomaly as the single most predictive signal, its removal alone reducing recall by 20.0 percentage points-more than triple the next-largest individual feature contribution. Deployed in place of a legacy rule-based flagging system, the OC-SVM pipeline simultaneously increases confirmed-case detection (126 versus the legacy system's 57 cases, a 119.2% relative improvement) and reduces investigator review hours by 44.9% (5,472 versus 9,940 hours over the same 36-month period), though survivorship bias in the confirmed-case validation set, adversarial evasion vulnerability, and cross-market-segment generalisation gaps each constrain how far these results extend beyond the validated population.

Hali tarjima qilinmagan