Do Machine Learning Models Beat a Constant Benchmark When Forecasting Next-Month Hurst Exponents? Evidence from Saudi-Listed Firms

Financial forecasting under uncertainty is fundamental to strategic investment decision making. Among the factors that influence investment decisions, market efficiency plays a critical role because it determines whether historical information can provide predictive value for future market behavior. The Hurst exponent is widely used to measure persistence and weak-form market efficiency, where values close to 0.5 indicate random behavior and values above or below 0.5 indicate persistent or mean-reverting dynamics. Despite extensive research using the Hurst exponent to characterize market persistence and market-efficiency-related behavior, limited studies have investigated whether future Hurst exponent estimates can be forecast using machine learning techniques and whether these models can outperform a simple constant benchmark. This study addresses this gap by developing a machine learning framework to forecast the next-month Hurst exponent using firm-level descriptive, financial, technical, and statistical indicators and comparing their performance against a constant benchmark. A dataset comprising 141 Saudi-listed companies, 705 firm-month observations, approximately 147 predictor variables, and five monthly periods was constructed. Monthly Hurst exponent values were estimated using the Rescaled Range (R/S) method and used as the target variable. Linear Regression, Ridge Regression, Elastic Net, Support Vector Regression (SVR), Random Forest, and XGBoost were evaluated using an expanding-window time-series validation strategy with two chronological forecast origins. Model performance was assessed using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R2). None of the evaluated machine learning models outperformed the constant benchmark, which achieved the lowest RMSE (0.227) and MAE (0.179). Ridge Regression performed relatively best among the machine learning models, with an RMSE of 0.236 and an MAE of 0.187, but it still failed to improve upon the benchmark, and all evaluated machine learning models produced negative R2 values. These results indicate that the selected firm-level descriptive, financial, technical, and statistical indicators provided limited incremental forecasting value relative to a simple constant benchmark within the adopted empirical design. Given the five-month observation period and limited number of chronological forecast origins, the findings are best interpreted as exploratory evidence regarding the ability of machine learning models to outperform a constant benchmark in forecasting next-month Hurst estimates, rather than as definitive evidence concerning market efficiency or the general predictability of financial markets. The study provides an empirical benchmark for future investigations into forecasting Hurst exponent estimates and changes in market persistence.