An Empirical Comparison of XGBoost and LightGBM for Capital Expenditure Deviation Prediction in U.S. FERC-Regulated Rate-Base Transmission Projects

Abstract

U.S. transmission investment is becoming increasingly important as the electric grid expands to support reliability, renewable integration, and long-term regional planning. In a rate-base regulatory framework, deviations between expected and actual transmission CapEx flow to customers through formula-rate true-ups, subject to FERC prudency review. Accurate CapEx deviation prediction therefore has practical value for both FERC prudency review and utility capital planning, yet the application of modern gradient-boosting models to FERC-regulated transmission CapEx data remains underexplored. This paper conducts an empirical comparative evaluation of XGBoost and LightGBM against CatBoost, Random Forest, and an elastic-net baseline, using a utility-year panel assembled from FERC Form 1, EIA-860, FHWA NHCCI, and BLS/FRED macro-commodity series spanning 1994 to 2023. With Bayesian-optimized hyperparameters and a 2016 to 2023 held-out window, XGBoost attains MAE 0.113 and R2 0.803, marginally ahead of LightGBM at MAE 0.118 and R2 0.784, while retaining a thirteen-percentage-point advantage over the linear baseline in direction accuracy. Feature-group ablations identify hardware and macro-commodity features as the two largest contributors, with leave-one-out MAE increases of 24.8% and 22.1% respectively. TreeSHAP attributions rank prior-year Construction Work in Progress, copper prices, and Order-1000 era effects among the leading deviation drivers. These findings carry direct practical value for FERC staff reviewing formula-rate filings, utility teams preparing annual CapEx updates, and state commissions intervening in formula-rate proceedings.