Weekly weather anomalies explain about 14% of the variance in withdrawal-season U.S. natural gas storage changes

Summary

Using EIA weekly Lower-48 working-gas storage figures and a national temperature-anomaly series built from NOAA daily station data compared against the 1981-2010 climate normal, a simple linear regression finds that the deviation of weekly temperature from normal explains about 14% of the week-to-week variance (R² = 0.141) in U.S. natural gas storage changes during the withdrawal season (November-March, n=175 weeks, 2011-2025). Each 1°C of anomalous warmth is associated with roughly 14 Bcf less gas withdrawn that week (coefficient +14.15 Bcf/°C, p<0.001). The relationship is robust to leaving out any single winter (leave-one-out coefficient range 12.5-15.0, no sign flips) and survives controlling for calendar month (partial r=0.355 vs. zero-order r=0.376). Outside the withdrawal season (April-October, injection months), the same temperature-anomaly measure explains essentially none of the week-to-week variance (R² = 0.2%, not significant) — injection-season storage builds are driven far more by production, LNG exports and power-sector burn than by ambient temperature swings. Pooling all 52 weeks together, without separating season, produces a misleadingly small R² (1.2%) because the sign and strength of the weather effect flips between the heating season and the rest of the year.

Why 14%, and why that is a lower bound, not the full story

The literature (EIA's own Natural Gas Weekly analysis, the STEO Natural Gas Module, and academic work such as Chen (2023) in the Journal of Futures Markets) treats weather — specifically the deviation of population-weighted heating/cooling degree days from the 30-year normal — as the primary short-run driver of weekly storage withdrawals, but never the only one. This analysis' own weather measure is a simplification of EIA's official methodology in one important way: it averages NOAA daily station temperatures unweighted across all reporting U.S. counties, rather than population-weighting them the way EIA's own heating/cooling-degree-day series does. A population-weighted measure would likely raise the R² further, since it would better track where the heating/cooling demand actually sits (dense population centers in the East/Midwest) rather than treating a sparse rural county the same as a dense metro county. Two variants were tested for robustness: (1) a piecewise heating-degree-day (HDD) version built from daily-level rectified temperatures gave a lower R² (7.9%) than the simple linear anomaly, likely because daily HDD summation adds noise the weekly-average version smooths out; (2) an in-sample random-forest fit reached R²=0.45, but that number is inflated by overfitting on only 175 observations and is not a genuine out-of-sample estimate — it is reported only as a loose upper bound on how much of the variance is theoretically recoverable from temperature alone with a flexible nonlinear form.

The remaining ~86% of week-to-week variance in withdrawal-season storage changes is attributable to factors this analysis does not model: production levels, pipeline/LNG export volumes, power-sector gas burn for electricity generation (itself partly weather-driven but also driven by non-gas generation availability), storage-region-specific dynamics (this analysis uses the Lower-48 national aggregate only), and the specific severity/duration of individual cold snaps that a weekly-averaged anomaly measure smooths over.

Data and method

• Storage: energy.eia_natural_gas_storage, region 'Lower 48 States' (R48), storage_type_code='SWO' (working gas), weekly report_date. Net weekly change = volume_bcf(t) − volume_bcf(t-1).
• Weather: weather.weather_daily_by_county (NOAA GHCN-Daily, aggregated to county-day, averaged nationally per day, then to the Sat-Fri week ending on each storage report_date) compared against weather.climate_normals_monthly (NOAA 1981-2010 monthly normals, averaged nationally by month). Anomaly = weekly average temperature − the calendar month's national normal average temperature.
• Regression: OLS, net_change_bcf ~ anomaly_c, run separately for withdrawal-season weeks (Nov-Mar) and injection-season weeks (Apr-Oct), plus a pooled full-year specification for comparison.
• Robustness checks: leave-one-winter-out sensitivity analysis (no sign flip, coefficient range 12.5-15.0 across the 11 non-trivial winters tested) and partial correlation controlling for calendar month.