Improving the explainability of autoencoder factors for commodities through forecast-based Shapley values.

Area of Science:

• Machine Learning
• Econometrics
• Financial Modeling

Background:

• Autoencoders are powerful dimension reduction tools in machine learning, analogous to Principal Component Analysis (PCA).
• Their application in finance for nonlinear factor models is growing due to flexibility and performance.
• A key limitation is the reduced explainability compared to PCA.

Purpose of the Study:

• To improve the explainability of autoencoders in nonlinear factor models.
• To introduce a novel Shapley value-based approach for assessing latent factor relevance.
• To identify significant nonlinear latent factors in financial markets, specifically the commodity market.

Main Methods:

• Utilizing Shapley values to quantify the contribution of each nonlinear latent factor.
• Implementing a forecast-based Shapley value approach for out-of-sample accuracy measurement.
• Applying the methodology to factor-augmented models in the commodity market.

Main Results:

• The Shapley value approach effectively measures the relevance of nonlinear latent factors.
• Identified the most influential latent factors for individual commodities based on forecasting performance.
• Demonstrated improved explainability of autoencoder-based factor models.

Conclusions:

• Shapley values offer a robust method to enhance the interpretability of autoencoders in financial factor modeling.
• The forecast-based approach provides valuable insights into latent factor importance for out-of-sample predictions.
• This technique is particularly useful for understanding complex dynamics in markets like commodities.