Methodology

The Problem with Standard HDDs

Most weather services report Heating Degree Days (HDDs) as a simple national average. A cold snap in Montana gets the same weight as a cold snap in Ohio, even though Ohio consumes roughly 10x more natural gas. This creates a disconnect between the weather data traders see and the actual demand signal that drives storage draws and prices.

How GWDDs Are Calculated

GasAlpha calculates Gas-Weighted Degree Days by extracting forecast temperatures from multiple weather models, computing regional heating degree days across the 9 EIA census divisions, and weighting each region by its share of national gas consumption. The result is a single number that reflects where heating demand actually matters for the gas market.

Unlike standard population-weighted HDDs, GWDDs emphasize high-consumption regions where gas actually gets burned during heating season. This makes GWDDs a more accurate proxy for storage draws and price sensitivity.

Weather Models

GasAlpha runs four independent weather models to provide a consensus view and quantify forecast uncertainty:

Storage Projection Model

GasAlpha projects weekly EIA storage changes using a regression model calibrated to the modern market regime. The model uses observed temperature data and GWDD estimates to predict weekly storage draws and injections.

Projections carry confidence tiers that decrease with forecast lead time, reflecting the inherent uncertainty in extended-range weather forecasts.

Data Sources

• Weather Models: NOAA GFS/GEFS via AWS Open Data, ECMWF IFS/AIFS via ECMWF Open Data
• EIA Storage: Weekly Natural Gas Storage Report from the U.S. Energy Information Administration
• Climate Normals: NCEI 1991-2020 daily temperature normals for baseline comparisons
• CPC Temperature Data: Climate Prediction Center daily temperature observations for model calibration
• CFTC Positioning: Commitments of Traders report for managed money net positioning
• Henry Hub Prices: NYMEX natural gas futures via market data feeds

Update Schedule

The GasAlpha dashboard updates automatically 4 times daily, timed to ensure all models from each cycle are available before processing:

Model Availability Windows

Weather models take several hours to process after initialization. The dashboard captures each model as soon as its data becomes available:

Henry Hub prompt pricing updates every 5 minutes during NYMEX trading hours.

Understanding Model Cycles: 00z vs 12z

Every weather model initializes from a snapshot of current atmospheric observations — surface stations, radiosondes, satellite retrievals, aircraft reports, and ocean buoys. These observations are assimilated into the model's starting state through a process called data assimilation. The quality and quantity of observations available at initialization time has a direct effect on forecast accuracy, which is why the two main cycles — 00z and 12z — can produce meaningfully different results even on the same day.

What the Cycles Mean

Model cycles are named for the UTC time of their initialization. The 00z cycle initializes at midnight UTC (6pm CT / 7pm ET the prior evening) and the 12z cycle initializes at noon UTC (6am CT / 7am ET). Both cycles run twice daily for GFS, GEFS, IFS, and AIFS. The models take several hours to process after initialization before forecast data becomes available for download.

Why 00z and 12z Can Differ Significantly

Each cycle ingests a fresh set of observations. Between the 00z and 12z initialization windows, additional upper-air balloon soundings, aircraft data, and satellite passes become available. When the atmosphere is in a transitional pattern — a front moving through, a ridge building, a storm deepening — the 12z run can see that evolution in ways the 00z couldn't. This is why 12z can sometimes show a dramatically different GWDD total than the prior 00z, especially in the extended range beyond day 5.

Which Cycle Should Traders Use?

There is no universally correct answer, but here is a practical framework:

• 00z settled consensus (8am CT) — The most stable daily read. All four models have had a full overnight assimilation cycle with a complete global observation network. This is the number most traders use as their baseline for the day. GasAlpha displays this as the "00z settled consensus" anchor in the hero and maintains a separate 00z-only ranking chart for historical comparison.
• 12z consensus (4pm CT) — The freshest available data. Incorporates morning upper-air soundings and additional satellite passes. More current but also more volatile, particularly in the extended range. Useful for confirming or challenging the morning read before the close of NYMEX trading.
• Run-to-run change — Often more informative than the absolute number. A consistent warming or cooling trend across multiple cycles (00z → 12z → next 00z) is a stronger signal than a single cycle outlier. GasAlpha displays same-cycle changes in the hero to help identify these trends.

The Extended Range Problem

Model skill degrades rapidly beyond day 7. In the day 1–3 range, 00z and 12z typically agree closely and both are reliable. In the day 4–7 range, meaningful run-to-run differences are normal and expected. Beyond day 7, individual model runs should be treated as probabilistic guidance, not point forecasts. A 10-day GWDD total that includes days 8–10 carries substantially more uncertainty than one anchored to the near-term window.

This is why large cycle-to-cycle swings in the 10-day GWDD — even 50–100 GWDDs — are not necessarily a data quality issue. They reflect genuine forecast uncertainty in a volatile atmospheric pattern. The appropriate response is to widen your confidence interval, not to anchor on either run exclusively.

Limitations

GasAlpha is a forecasting tool, not a crystal ball. Important limitations to understand:

• Weather model uncertainty: All NWP models have inherent errors that grow with forecast lead time. Day 1-3 forecasts are significantly more reliable than day 7-10.
• Consumption weights are static: The EIA consumption weights represent average heating season demand shares. Actual daily demand varies with price, industrial load, power generation mix, and other factors.
• Storage model simplification: The regression model captures the temperature-to-storage relationship but does not explicitly model LNG feedgas variability, pipeline maintenance, or industrial demand shifts.