In recent discussions I have examined the Kelly Criterion and Meta-labeling as means of risk management and now I am focusing on regime-based models. Regime-based models are trading models that are supposed to perform well in specific market conditions. The idea is that a combination of multiple, regime-specific strategies can outperform a single, all-weather trading strategy. However, in my opinion this is a misguided notion: managing multiple regime-specific strategies is just as challenging, if not more so, as having a single, more robust trading strategy.


ZNZTng5oreszYtPnC-uryenQWgJR6LVOtJxNcI4FWtR-y1J1R9VBFATXtMbGB3CCksLqGJWFs7IwUcomlc6wNzD09Lm0B1bjpyH2OxBdwMxlvZUHlxITBcMrUbZF6l9hz7KOzTsdfnh8uk2iUZGyNK8Conceptual comparison of end to end vs regime-based strategies


The main issues with regime-specific strategies are:

  • Changing market dynamics: Financial markets are constantly changing, with new information and events affecting prices and volatility. Time-specific strategies that worked well in past market regimes may not be effective in the future or may perform poorly during unforeseen regime shifts.
  • Difficulty in detecting regime shifts: It can be challenging to identify, or even forecast, regime shifts accurately in real-time, leading to poor performance and increased risk. Even if you have N perfect regime-specific trading strategies, knowing when to switch between them is the real challenge, which is somewhat related to the meta-labeling problem. You basically need a meta-strategy "wiser" than the main strategy currently in use to determine when to switch.
  • "Regimes" are mental constructs: The concept of regimes is simply a way of explaining why different rules apply in the current market and they may not be clear-cut. Conditions can change rapidly (e.g., during the Covid pandemic) or slowly (e.g., during the 2010-2020 bull run), resulting in a mix of past regimes or new, unseen ones.


In my opinion, regime-based trading is a form of overfitting. You have a strategy that works well under specific conditions and hope to figure out the best time to apply it in the future, which is basically market timing. 

Instead of developing multiple, market-specific strategies, I prefer to focus on developing machine learning (ML) models that account for overall market conditions (e.g., VIX, yield curve, etc.) and embed them in the training and targets of my model. I prefer simple and agile models that can be retrained quickly (e.g., weekly or even daily) to incorporate market changes immediately.

I argue that traders should adopt a more flexible, adaptable approach that adjusts to changing market conditions in real-time. This could involve using ML algorithms that learn from data and adapt to changing conditions, or using a combination of fundamental and technical analysis to make trading decisions.

I have included a toy ML-based algorithm that incorporates the VIX value as a regime feature as a starting point for further development. Note that I would have preferred comparing a regime-based and an end-to-end strategy, but I did not have the opportunity to develop one. If the topic is of interest, I can look into comparing a simple In-and-Out classic strategy with an equivalent ML-based version.