Financial markets are inherently nonstationary systems, characterized by structural breaks and drastic regime changes, making concept drift a significant challenge for predictive modeling. Conventional machine learning approaches often fail to address this, proving slow or incapable of adapting to such non-stationarities and requiring computationally expensive retraining. This work introduces GroCH, an adaptive meta-learning algorithm designed for online incremental learning on high-frequency financial data streams. GroCH aims to maintain robustness during periods of market instability by effectively handling various types of concept drift. It leverages a hybrid approach, combining unsupervised learning for concept representation with supervised models to ensure predictions are consistently performed with up-to-date models.

The algorithm’s efficacy is validated through experiments on price-movement direction prediction. These tests utilize both semi-synthetic data streams that simulate financial series with known regime changes, and real-world data from the SPDR S&P 500 exchange-traded fund at various intraday resolutions. GroCH demonstrates competitive performance relative to the relevant literature, highlighting its potential for financial forecasting amid structural changes.