Do electricity prices predict the returns and volatility of cryptocurrency mining companies? Evidence from GARCH and machine-learning models

We test whether U.S. wholesale electricity prices contain independent predictive information for the daily returns and volatility of eight publicly listed cryptocurrency mining stocks over 1054 trading days. Evaluating all 16 predictor subsets within AR-X GARCH and deep learning (LSTM, GRU) frameworks, 384 estimated models in total, we document a novel factor-substitution pattern: for Cipher Mining and TeraWulf, AIC model selection identifies Mid-Columbia electricity returns as the dominant predictor, with the S&P 500 absent from the optimal specification. For a complementary group of four stocks, lagged electricity volatility is associated with reduced subsequent conditional variance, a pattern consistent with operational curtailment hedging. A deep learning model (LSTM or GRU) achieves lower RMSE than AR-X GARCH in seven of eight stocks; Diebold–Mariano tests indicate that forecast accuracy differences are not statistically significant, supporting the interpretability of the econometric framework. Our findings reveal a previously undocumented segmentation of the mining universe along energy cost exposure with implications for portfolio construction.