A case to focus on causal discovery algorithms against non-causal feature selection techniques.
"Notably, causal methods demonstrate remarkably low prediction errors during volatile periods, such as the global financial crisis (GFC) and the COVID-19 outbreak, showcasing their robustness in regime shifts. These causal models select more stable and meaningful sets of predictors throughout the sample period, indicating their reliability and consistency across varying market conditions...
...We have found empirical evidence that causal feature selection methods have been more stable regarding prediction error in most financial crises we labeled. The prediction error metrics that we compare show lower values for all the causal feature selection models when compared to the prediction-based features selection model.?"
Causality-Inspired Models for Financial Time Series Forecasting https://arxiv.org/abs/2408.09960?utm_source=substack&utm_medium=email
Turnleaf Analytics (Forecasting Macro and Inflation with Machine Learning and Alternative Data)
7 个月A good paper and definitely the way to go but we must be mindful that the assumptions behind the PCMCI (an other causal discovery algorithms) are rarely fulfilled in practice e.g. stationarity (which can be tackled but it is time dependent), causal sufficiency (rare), causal Markov condition, faithfulness.