The Limits of Machine Learning in Daily Return Forecasting: Evidence from Volatility Regimes, Tail Risk, and Multi-Asset Markets

The study analyzes volatility regimes, tail risk and machine learning forecastability of seven financial assets (BTC, ETH, the S&P 500, NASDAQ, Gold, Crude Oil, and US Dollar Index) from 2017 to 2026. Structural break analysis shows that financial instability is characterized by variance, not by a change in the mean. EGARCH is preferred in five out of seven assets, with significant leverage effects in the equity indices. According to EVT analysis, ETH and BTC are the highest-tail-risk assets, while Crude Oil has the heaviest tail. Kupiec and Christoffersen VaR back testing results show that Gold’ s actual exceedance rate at the 95th percentile is almost twice the expected exceedance rate. XGBoost doesn’ t statistically outperform Linear Regression (DM = \(-\)0.403, p = 0.687), and the most important predictive element is a 5-day rolling mean return. The results show that most predictive information in daily returns is contained in linear autoregressive features.