Same Marginals, Same Covariance, Different Tails? Electricity Basis Risk in NYISO and ERCOT

A position that is long power in one zone and short in another carries basis risk, and most of that risk sits in a few days each year. We ask which modeling choice improves day-ahead tail forecasts of such spreads: empirical instead of parametric spread shocks, or a more flexible model of how the two zonal prices move together. Thirteen procedures are run on two NYISO and two ERCOT spreads over 2020–2025 (2,192 days each), with all settings chosen on NYISO development data and frozen before testing. The core experiment fixes both zonal forecast distributions and their covariance, so any change in tail accuracy must come from the shape of the dependence. That shape matters little: conditioning it on the current market state moves mean tail loss by less than a third of a percent in every spread. Filtered historical simulation (FHS) improves on a parametric ARX–GARCH spread model in all four spreads, by 0.5–4.6%, though no gain survives a post-hoc correction across 312 comparisons. The route matters in ERCOT, whose zonal prices move almost in lockstep. There every copula built on zonal forecasts has 30–243% higher tail loss than direct ARX–GARCH, and FHS beats the constrained copula on South–North by 25.9% (simultaneous 95% interval 3.2–43.3%). ERCOT failed a numerical grid check, so its copula results are exploratory.