Beyond Trend Prediction: Early Unveiling of the ۲۰۲۰ Tehran Stock Exchange Crisis through Anomaly Detection with Isolation Forest and One-Class SVM in a Shifting Macro-Policy Landscape

Predicting crashes in volatile markets like the Tehran Stock Exchange (TSE) is challenging. This research shifts focus to early anomaly detection, evaluating Isolation Forest (IF) and One-Class SVM (OCSVM) on their ability to identify signals preceding the ۲۰۲۰ TSE crisis, using ۱۴ technical indicators and pre-crisis training data. Both models showed sharp anomaly score drops, potentially signaling the crisis ۱۲۰ days early. However, Precision-Recall analysis showed IF marginally outperformed OCSVM (AUC-PR=۰.۰۸ vs. ۰.۰۷), though both struggled to balance crisis detection with low false alarms. A moving-window alert system achieved full crisis recall but with low precision. Findings highlight anomaly detection's supplementary risk management potential, emphasizing the critical need for improved threshold calibration for actionable alerts, with IF emerging as a possibly more adaptable choice.

Keywords:

Anomaly Detection , Early Warning System , Financial Crises , Tehran Stock Exchange , Machine Learning , Systemic Risk