How the WY2026 forecasts did

The season in one line

Of the 16 forecasts we published for WY2026 that can now be graded against a closed season, ours came closer than NRCS’s in 15. The typical miss was 40% for ours against 159% for theirs. A season is a small sample and we will say so every year: one season neither proves nor refutes a forecast, which is why the multi-decade record below is the number that matters.

Every graded forecast, point by point

What each forecast said, and what the river then delivered. A negative miss means the forecast was low.

Volumes in thousand acre-feetA volume: one acre covered a foot deep, about 326,000 gallons -- roughly a suburban household's outdoor use for a year. (kAF) over each issue’s own forecast period. “Won on” names what earned this point a published number: the number means our central value was clearly closer over past seasons; the odds means our stated range was much better calibrated while the central value held level.

What this is, and what it is not

This is a calibration product. Replayed across 25,886 past forecasts — every year predicted using only the years before it — our central number is a dead heat with NRCS’s (20.3% typical miss against theirs at 20.4%). Where we are genuinely better is the odds: our stated range holds 81.3% of the time against a nominal 80%, where NRCS’s holds 74.2% — and the gap is widest in dry years, which are the years the number matters most.

So: do not read this as a more accurate forecast. Read it as the same forecast with honest error bars. NRCS remains the official forecast and the authority; we hold the verifying record and publish what it says. We fit 79 forecast points and publish a number at only 30 of them — those whose own replay beat NRCS. Everywhere else the honest answer is that NRCS’s forecast, unadjusted, is the best available, and that is what the district pages say.

The ceiling, stated plainly: most of what is left of an April forecast’s error is weather that has not happened yet, and seasonal precipitation forecasts for the interior West have close to no skill. No method fixes that, ours included. Coefficients are frozen once a year in December and never refit mid-season, so a number cannot quietly change under you. Built 2026-09-01. The five-decade record behind all of this →

Do the odds hold up? Year by year

Both forecasts publish a range they say the volume will land inside about 80% of the time. This is how often it actually did, in each of 34 years, replayed without ever seeing the year being forecast. The dashed line is the 80% both are claiming. Across the driest third of years ours holds 78.9% against NRCS’s 68.6% — the gap that is the whole point of this model.

Where ours fails, stated plainly. Our worst years are 2002 (15.3% against NRCS’s 34.6%), 1995 (45.5% against NRCS’s 33.1%), 2015 (48.9% against NRCS’s 47.6%). These are the extremes — the historic 2002 drought among them — where both forecasts broke and ours sometimes broke worse. A model calibrated on the bulk of the record does not get the tails for free, and a year outside anything in the training record is exactly where a stated range should be read with suspicion. We publish these rather than the average alone, because the average is not what a drought year feels like.

Who made this, and what it stands on

The outlook model is developed and published by John Crawford (Hydrohistorian, a True Ascent Labs project). It is not an NRCS product, is not endorsed by NRCS or any agency, and carries no official standing — it is an independent second opinion computed from public records. Anything on this page that is wrong is ours, not theirs.

Sources. Seasonal water-supply forecasts and snowpack data: the NRCS National Water and Climate Center (AWDB — SNOTEL snow-water equivalent, soil moisture, and the SRVO forecast and naturalized-volume archive). Observed streamflow: Colorado’s Division of Water Resources published gage record. Basin snow and precipitation: the University of Arizona 4 km SWE/precipitation product. Drought index: gridMET (Climatology Lab, University of Idaho). Each is used as published; none of these organisations has reviewed this work.

Method credit. The approach — pooled gradient-boosted quantile models trained across many forecast points at once, stacked on the official forecast — follows the published solutions of the U.S. Bureau of Reclamation / DrivenData Water Supply Forecast Rodeo (2023–24), whose winning code is open source, and the peer-reviewed water-supply-forecasting literature, including NRCS’s own machine-learning work (Fleming et al.). Built with LightGBM (Microsoft, MIT licence). The verification method, the per-point publication rule and every number on this page are our own. How this site is built →