Unknown Is Not Zero: Heat Stress, Data-Center Power-Demand Exposure, and Utility Volatility

Abstract

This paper asks whether service-territory heat stress and disclosed data-center power-demand exposure jointly amplify the volatility of U.S. listed electric utilities. Its principal finding is methodological but empirically consequential: pre-disclosure silence is not a verified zero. The study constructs an auditable point-in-time chain from dated operating or binding customer-load megawatts to the regulated utility, the applicable EIA Form 861 peak-load denominator, a representative NOAA station, daily parent-company returns, and lag-available financial controls. Ten issuers satisfy a broad data-center four-quarter evidence rule over 2021–2025, although the measure is not uniformly AI-specific and one issuer is predominantly exposed to cryptocurrency hosting. Reclassifying 8,685 legacy zerocoded issuer-days as unknown leaves 3,855 positive-disclosure issuer-days and 212 complete portfolio observations across nine variance-recursion segments. The measurement path is non-monotone: the uncentered interaction is 0.143 in the legacy zero-coded panel and 0.060 when OTTR is excluded; after the common-positive restriction it reaches −2.000 at the coefficient bound, while centering yields an interior estimate of −0.118. Before centering, Corr(C,C×A)=0.979 and the interaction VIF is 29.4; centering lowers the VIF to 1.09. In the final Student-t GARCH-X specification, AIC and BIC select the no-interaction model and a 499-replication conditional parametric bootstrap does not reject it (LR=1.638, p=0.324). The current design therefore fails to detect heat-by-load amplification; it does not establish that the population effect is zero. H3 remains untested, and the pre-specified Markov-switching extension is not estimated. Positively, the audit identifies how exposure coding and coordinate construction can manufacture apparent interaction evidence, supplies a reusable regulatory-data architecture, and defines an advance rule under which a longer panel can retest the same hypotheses without redefining the original evidence.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.