Self-Adaptive Machine Learning Models for Financial Risk Forecasting: Handling Non-Stationarity in Banking and Cryptocurrency Time Series

Financial markets rarely sit still. They drift, jump, and occasionally snap into entirely new regimes, with volatility behaving differently depending on the moment. These shifts steadily erode the usefulness of static risk forecasting models. Approaches built on fixed parameters or scheduled retraining tend to lag behind reality, especially when markets are under stress and conditions change fast. This paper investigates whether self-adaptive machine learning models can generate reliable and practical volatility forecasts in live settings, without relying on repeated full retraining cycles. Using daily closing prices from the S&P 500 index and Bitcoin as examples of traditional and cryptocurrency assets, the task is set up as a strictly out-of-sample, one-step-ahead volatility forecasting problem. We compare static models, periodically retrained models, and fully self-adaptive models across calm periods and clearly defined stress regimes, including the COVID-19 crash and major cryptocurrency boom and bust cycles. The results show that self-adaptive models deliver stronger performance across both asset classes. They achieve lower forecast errors, more stable errors during stress, quicker recovery after shocks, and sharply lower computational demands. Recursive EWMA-based models stand out for their solid performance using constant memory and lightweight updates, while online gradient-based learners adapt flexibly without drifting into instability. Taken together, the evidence points to continuous adaptation as a structurally superior approach for financial risk forecasting under non-stationarity. The study demonstrates that self-adaptive models can be deployed in real-time risk systems and sets the stage for future work that connects adaptive forecasting with evolving explainability and regulatory reporting requirements.