Innovations to fundamental stock valuations: Estimating future earnings per share and free cash flows using statistical and machine learning methods

This study proposes innovations to financial valuation models. Fundamental valuation is used by investors to make buy/sell decisions regarding stock issues. Valuation is the process of determining the intrinsic value — the price reasonable to pay for a stock given its future prospects, which are measured with cash flows, given the level of risk an investor bears when buying stock. These cash flows are measured with two key series: Earnings-Per-Share (EPS) and Free Cash Flows (FCF). Correspondingly, the two series are used as inputs in common valuation models: the Forward Price-Earnings and Discounted Cash Flows. It is required that investors estimate the future cash flows — a sensitive process whereby under- or overstating future cash flows is prone to the risk of losing invested equity. Hence, being able to accurately capture the next quarter's value is of utmost importance for investors active in financial markets: it guides the stock selection process. We propose to formulate this as a regression problem, where the target variable is the next quarter's Earnings-Per-Share or Free Cash Flow value. The input features are their respective lags. The main challenge in this problem is the fact that the series are sparse and limited in the number of observations. This is because the fundamental financial data is published every quarter of the year, as required by law. Hence, our estimators have limited training/validation data to learn from. We approach this problem with the selection of 8 Machine Learning (ML) and 5 Statistical (SE) estimators, conducting experiments on a representative sample of 100 U.S. publicly traded companies. Our study contributes in several ways. First, we show that the quantile transformer and the PCHIP interpolation improves model generalization by making more blatant the linear relationship of the target variable with its features and artificially increasing the number of data observations, respectively. Second, we demonstrate that while certain ML estimators do overfit to the small data sets, others perform at the same or better rate than the statistical estimators. Third, building on top of our observations about data patterns and behaviour of a diverse range of estimators, we propose the transfer learning methodology that allows to combine the predictive capabilities of the ML and SE. We demonstrate the effectiveness of our approach based on the reduction across the range of regression error measures, and the improvement in portfolio performance, over a fixed backtesting simulation period.