A0774
Title: Sparse neural networks and explainability in financial statement analysis
Authors: Lars Fluri - University of Basel (Switzerland) [presenting]
Abstract: An alternative approach is proposed to feature selection and sparse modelling in the context of financial data analysis to predict free cash flow. Utilising deep learning important features (DeepLIFT), a process is introduced for iterative elimination of input features and extends currently used algorithms. This reduces the model complexity and enhances the robustness through the elimination of less significant input nodes. Furthermore, a method for regrowth of nodes using gradient magnitude of previously eliminated features based on state-of-the-art methods is used. Drawing on a dataset of 874 firms from the DACH region over a decade, the model is used to identify forward-looking predictors of free cash flow. Additionally, it evaluates both computational aspects and performance metrics (including in-sample and out-of-sample performance) to measure improvements from the original dense model to the optimized sparse model. By reducing the number of total features present by over 75\% (from 49 to 10 features), out-of-sample $R^2$ decreases by only 13\% while the standard deviation is improved by over 35\%. The contribution is to the evolving field of machine learning applications in finance, proposing an alternative framework for feature selection and model optimization.
