Are Black Boxes Needed to Forecast Equity Volatility Regimes? Three Pre-Registered Studies of a Transparent Fuzzy Rule Base, Machine Learning and Standard Models

Machine learning is often assumed to be necessary for forecasting equity volatility, yet risk governance increasingly requires models that can be audited. This study asks whether a compact, readable fuzzy rule base can forecast volatility regimes as well as a leading machine learning model, and whether the machine learning advantage reported in one sample holds elsewhere. Three pre-registered studies covering 181 stocks in thirteen markets test these questions against standard volatility benchmarks, with protocols fixed and time-stamped before the test data were analysed. The transparent rule base performed essentially as well as the machine learning model in every study and reproduced the relationship that model had learned. The machine learning advantage over standard benchmarks appeared in the U.S. sample but did not replicate in other markets; market-wide realized volatility helped modestly, whereas lagged implied volatility did not; and using the forecasts to scale exposure did not improve risk-adjusted returns. The findings imply that, for this task, institutions need not trade interpretability for accuracy, and that claims of machine learning superiority in finance should be judged by pre-registered replication across markets before they inform risk-management practice.