How to spot outliers: an Ensemble Anomaly Detection Framework

Battle Count: 55

The paper is primarily focused on operational risk and model validation in investment banking rather than direct trading strategy development. However, it is highly relevant to quantitative trading infrastructure: ensuring the integrity of risk metrics (VaR, ES, credit sensitivities) that underpin position sizing, hedging, and capital allocation decisions. The detection of stale values, sign flips, and scaling errors directly impacts the reliability of inputs to trading systems. The FRTB and Basel III compliance angle is critical for any desk using internal models for regulatory capital.

The framework involves multiple components: six statistical/ML detection methods, a deterministic filter, ECDF-based normalization, weighted aggregation with qualified-majority voting, and per-dataset calibration. Each method requires separate implementation and tuning. The calibration procedure (method selection, weight optimization, aggregation rule choice) adds complexity. However, individual methods (Z-scores, Isolation Forest, LOF) are well-established and available in standard libraries. The main complexity lies in the ensemble composition, normalization, and dataset-specific calibration rather than in any single algorithm.

The paper uses proprietary data from UBS Investment Bank (183 credit-derivative trades, 129 trading days) that is not publicly available. The methodology is well-described with formulas and architecture diagrams, but empirical validation cannot be reproduced without access to the same data. The anomaly injection protocol is described in detail, but calibration parameters (severity coefficient k=6, MAD thresholds) are dataset-specific. No code or repository is provided.