VERIFICATION
How the forecasts performed: weather models on identical samples, raw and calibrated; whether the ensembles' ranges and odds hold up; issued storage estimates against EIA; and the historical winter record.
Recent weather · matched comparisons
Lower 48, population-weighted. Both entries are scored on the same initialization, cycle, valid date and lead day. Choose a model's calibrated version or a combination to see what calibration and blending change.
Mean absolute error by days ahead
Lower is better. The normal is the error of forecasting the 1991–2020 normal on those same samples.
MAE is the average size of the error; bias is forecast minus observed, so positive means too many degree days. Skill is 1 − model MAE ÷ normal MAE; above zero beats the normal. Counts are forecast run/day pairs, which can share an observed date and are not independent samples. Small samples should be treated cautiously.
Ensembles · are the ranges and odds reliable?
A 10th–90th percentile range should hold the outcome 80% of the time. Raw ensemble members sit too close together in the first days; the calibrated members have their spread set against their recent error.
Outcomes inside the 10–90 range · percent by days ahead
On target at 80%. Below it, the range is too narrow and the forecast too confident.
CRPS is the probabilistic counterpart of the mean absolute error, in the same units; lower is better. Skill compares it with a climatological ensemble (the same calendar day in 1991–2020). Spread ÷ error is 1 when the members' spread matches the error of their mean, and below 1 when the ensemble is too confident. Run/day pairs share observed dates and are not independent; one window is one season.
When members say “above normal”
For each share of members above the 1991–2020 normal, how often the outcome was above normal, with the number of cases. Reliable odds match their column: about 20% in the 10–30% column, about 80% in the 70–90% column.
Brier skill compares the odds with always quoting the sample's own frequency; above zero is better. Cells with fewer than five cases are left blank.
Calibration in force
Each model's daily maximum and minimum are corrected at every grid cell from the forecasts that verified in the last 30 days, before degree days are taken. Ensemble members are then spread about their mean by the factor shown.
The last column is where the blend still takes the corrected forecast: it compares the two versions on the runs they share and keeps the raw one unless the correction beats it, because a bias fitted to 30 days goes stale through a seasonal turn. Heating degree days, Lower 48. A model reading “no leads” is still published calibrated on its own — the blend simply is not using it.
How much of the error is the measurement
Every model's error on the same day carries one term that belongs to the observed record rather than to the forecast: nClimGrid reports a station observation day and the forecasts carry a local solar day, and its 5 km cells keep variation a 25 km grid cannot. Because that term is the same for every model, it can be separated from the error each model owns.
The floor is read at day 0, where a model has already seen the weather and its own error is near zero, and held constant: it does not depend on how far ahead the forecast is. It is an upper bound — error the models share because they share initial conditions cannot be told apart from shared measurement error by comparing models alone. “Spread ÷ forecast error” is what the published spread factor would be aiming at if it aimed at the error the model owns; above 1 the members are too close together, below 1 too far apart. The published ranges are not changed by this: see the methodology.
How the calibration, the blend and the pooled ensemble are made.
Issued storage estimates · observed record
The last estimate issued strictly before each report's release, with its original issue time. This is the record of published estimates.
Estimate and reported change · Bcf
The summary covers up to 52 settled reports; the table shows the latest 12. Lead time and observed-weather coverage vary, so these estimates should be assessed separately from the fixed-horizon winter reconstruction. EIA truth uses the currently downloaded inventory record and may include revisions.
Historical winters · archived forecasts and reconstruction
November–March. Gas-heating-weighted HDD skill and weather-driven storage estimates at seven, three and zero days before Friday week end.
Gas-heating HDD · error by forecast day
Day 1 is each run's first complete daily value. Each model is compared with the normal on its own available sample. Model samples differ; this chart does not rank the models on identical dates.
Storage reconstruction · Bcf
For each model and horizon, the all-observed benchmark uses the same weeks, storage fit and release cutoff. It measures the model's error when all weather is known.
What the winter reconstruction assumes
Historical observations and storage reports use their current revised values. Original data vintages and actual arrival timestamps are not preserved, so this is a reconstruction of the information available at issue time.
Read the evaluation method and release-calendar assumptions.