Financial Anomaly Detection for the Canadian Market

Battle Count: 91

The paper is primarily focused on detecting financial stress events and market crashes rather than direct trading signal generation. However, the anomaly detection framework could inform risk management overlays, crisis-aware portfolio rebalancing, and volatility regime detection. The methods are more suited for systemic risk monitoring than for generating alpha. The Canadian market focus and the nature of the events detected (crashes, not gradual trends) limit direct applicability to high-frequency or mid-frequency trading strategies.

The pipeline involves multiple non-trivial components: CCM correlation computation, directed flag complex construction, persistent homology computation (pyflagser), GNN training with GINE convolutions (pytorch-geometric), and knowledge distillation frameworks. The TDA component requires understanding of algebraic topology. The GNN architectures require careful hyperparameter tuning. However, the authors provide complete code on GitHub, reducing practical implementation barriers. The combination of multiple methods (TDA + GNN + PCA + anomaly detectors) adds integration complexity.

Code is publicly available on GitHub (https://github.com/njmead811/Financial-Anomaly-Detection-for-the-Canadian-Market). Data sourced from Yahoo Finance (yfinance package). Uses standard libraries: pyflagser for TDA, scipy for Mahalanobis, scikit-learn for LOF/PCA, pytorch-geometric for GNNs. Hyperparameters are explicitly listed. However, some architectural modifications from original papers (e.g., replacing GIN with GINE) require careful reimplementation.