How to handle Regime Changes (by ex HFT quant trader)

Defining Regime Change and Non-Stationarity00:00:01

Introduction to the concept of regime change in trading. This section defines it mathematically as non-stationarity, where statistical properties—mean, variance, skewness, and kurtosis—are not constant over time, making standard predictive models ineffective.

Testing Stationarity in Financial Data00:03:46

An empirical look at Bitcoin returns, illustrating how structural changes (e.g., major market events like the FTX collapse) cause massive fluctuations in data distribution, proving that financial time series are rarely stationary.

Feature Engineering and Memory Encoding00:09:43

Techniques for improving supervised learning models by encoding historical data. Using moving averages and lag-based features allows models to capture current market dynamics rather than relying on global averages that fail during regime shifts.

Online Learning with Passive-Aggressive Regressors00:14:27

Introduction to online learning, where models update their weights continuously as new data points arrive. The passive-aggressive regressor adjusts to prediction errors in real-time, effectively allowing the model to switch between momentum and mean-reversion styles.

Reinforcement Learning for Trading00:20:40

Modeling trading as a multi-armed bandit problem. This section introduces policy gradient learning (REINFORCE algorithm) and the critical role of entropy regularization to prevent models from getting stuck in local optima, ensuring they remain adaptive to changing environmental rewards.

Conclusion and Summary00:34:24

Final overview of strategies discussed: using hidden states for supervised learning, online learning for real-time error correction, and reinforcement learning with entropy regularization for long-term adaptation to changing market regimes.