Tail Risk and Crash Detection in Cryptocurrency Markets : An Extreme Value Theory and Deep Learning Approach

Sammanfattning

This thesis investigates crash detection in cryptocurrency markets bycombining Extreme Value Theory (EVT) with machine learning and deeplearning. Cryptocurrency returns show heavy tails, volatility clustering, andregime changes, making extreme downside events both practically importantand difficult to detect. The thesis focuses on BTC/USDT and ETH/USDTusing 5-minuteOHLCV(Open,High,Low,Close,Volume)datafromBinanceover the period January 2019 to December 2023.The thesis follows a two phase methodology. In Phase 1, the tail ofcryptocurrency log-returns is characterised using the Peaks-over-Thresholdapproach. The Generalized Pareto Distribution (GPD) is fitted to tailexceedances, and the Hill estimator is used as a non-parametric check of heavytailedness. The results show positive GPD shape parameters for both BTC andETH,confirmingpower-lawtail behaviour and justifying our method for crashdetection methods.In Phase 2, crash detection is approached as both a supervisedclassification problem and an unsupervised anomaly detectionproblem. Threemodels are compared: a LightGBM classifier using statistical features,an LSTM Autoencoder trained only on normal market regimes, and aTransformer Encoder classifier trained with an EVT-informed focal loss.Crashes are labelled algorithmically using a drawdown based rule, and allmodels are evaluated on identical train/test splits using metrics appropriatefor imbalanced classification, with more emphasis on the Area Under thePrecision-Recall Curve (AUPRC).The results show that the Transformer achieves the strongest overallcrash detection performance, outperforming both LightGBM and the LSTMAutoencoder on AUPRC for both assets. However, an ablation studycomparing EVT-informed focal loss with standard focal loss finds onlymarginal differences, suggesting that the Transformer architecture itself isthe main driver of the performance improvement, while the current EVTweighting mechanismdoes not provide clear additional predictivebenefit. Thethesis concludes that EVT is valuable for characterising cryptocurrency tailrisk, but that more refined methods are needed to integrate EVT directly intoour deep learning model for crash detection.