Mixed-frequency feedforward neural network with optimal exogenous variables for bitcoin mining energy forecasting

Abstract

Abstract This paper analyzes the electricity consumption of Bitcoin mining as a component of blockchain-based financial infrastructure and develops a hybrid forecasting framework that combines a Neural Network Autoregressive model with Exogenous Inputs (NARX) and Mixed Data Sampling (MIDAS). The specification embeds nonlinear state dependence within a feedforward neural network structured as a NARX and exploits mixed-frequency information from daily and monthly indicators to forecast weekly electricity consumption. A key methodological contribution lies in reframing exogenous variable selection as a ranking-based optimization problem grounded in individual explanatory power. To support this, a large language model (LLM)-assisted screening procedure is used to construct a theory-consistent pool of candidate predictors drawn from the finance, energy and cryptocurrency literature. From this pool, an optimization-based strategy identifies a parsimonious subset of variables that minimizes forecast error within the NARX–MIDAS framework. Empirical results demonstrate that the optimized model significantly outperforms benchmark specifications, achieving reductions of 15–20% in root mean squared error and 10–12% in mean absolute error. Beyond predictive performance, the proposed framework yields interpretable insights into how macroeconomic conditions, policy-related uncertainty and financial market dynamics influence Bitcoin mining activity. These findings have direct implications for risk management, energy planning and regulatory oversight in blockchain-based financial systems, highlighting the value of integrating LLM-assisted knowledge extraction with rigorous optimization-driven forecasting methodologies.