ML models for equity returns are business cycle dependent

 

Machine learning is being viewed by many as the solution to all forecasting problem. The initial evidence from several studies is positive but the computing costs and the barriers to entry for using these models is high. It takes work to implement ML programs. A new study focuses on forecasting skill of ML models of the SPX index but decomposes performance between recession and non-recession periods. The paper finds that performance decreases during recessionary periods especially when volatility is higher. 

This is unfortunate, yet is should not be surprising. Recessions are not frequent, in the data set tested there were seven recessions, so the periods for training models during recessions are smaller. If you don't see the phenomenon, you cannot model it. This applies to all modeling. We also know that parameters may change during recession periods so using relationships during normal periods may fail. See "Stock Price Predictability and the Business Cycle via Machine Learning"

The authors use recurrent neural networks which include Long Short-Term Memory (LSTM), bidirectional LSTM (BLSTM), and gated Recurrent Unit (GRU) model as well as multi-layer perceptrons (MLP).  The models that have macro variables that account for recessions and expansions in the economy show some improvement especially during periods of expansion. As has been noted, equities often bottom before the trough in growth and then start to rally while the economy is still in recession. Just identifying recessions may not be enough.

Recessions are just different and harder to model.  Just because you think you have a good model does not mean you are ready for recession.