The Performance of Machine Learning Regression Algorithms in Sparse Financial Time-series Datasets

Modern machine learning algorithms capable of finding complex relationships in the time-series data proved useful to model and forecast financial data. Most studies are concerned with forecasting of prices of financial assets, generated every second. On the other hand, fundamental data such as revenues, total assets, per share earnings, etc., is only generated every quarter of a year and is thus in low availability. This study focuses on forecasting of free cash flows (FCF) – the amount of cash available to shareholders after all investments and dues were paid by a business. The FCF is one of the key inputs to Discounted Cash Flow (DCF) model, most frequently applied for fundamental stock valuation purposes.

As well as other types of fundamental business data, the FCF data is supplied every quarter of a year, and since no business been in existence for hundreds of years, the regression problem needs to be solved with as much as 120 sample observations available. Specifically, we examine the performance of popular ML algorithms and that of classical statistical models, such as ARIMA, Simple Exponential Smoothing, OLS, broadly recognized as “golden standard” approach for sparse time-series datasets. We report the performance of both categories of models in a subset of 100 publicly traded companies.