What MAPE Cannot See: Directional Forecast Bias and Operational Risk in Electricity Markets

Abstract

Forecast accuracy metrics such as Mean Absolute Percentage Error are incomplete guides to operational risk when prediction errors are asymmetric. Using hourly ERCOT data from January 2019 through January 2024, we evaluate seven forecasting models. Accuracy and directional bias can diverge: the lowest-MAPE specification has MPE of 0.23% versus 0.03% for its seasonal counterpart, while tree-based MPE reaches 2.22% in August and 4.6% to 6.3% in FARWEST. We develop a four-type bias typology and a governance framework for bias-aware forecasting that generalize to settings where prediction errors carry asymmetric costs.