Hydrohistorian
The NRCS runoff forecast report card
The statewide verdict
Across all 69 graded points, the median point’s April even-odds forecast (the “50% exceedanceA forecast stated as odds of AT LEAST this much: the 90% exceedance value is a near-floor the river beats nine years in ten, the 10% value a near-ceiling it reaches one year in ten, and 50% is even odds -- as likely to come in above as below.” number — as likely to come in above as below) misses the delivered volume by 20% in a typical year. The stated likely range (the 90% value the river should beat nine years in ten, up to the 10% value it should reach one year in ten) actually contains the outcome 78% of the time at the median point — a perfectly calibrated forecast would say 80%. 29 points run more than 5% high on average and 7 run more than 5% low. New to these numbers? How to read a runoff forecast →
This is a grade of the forecasts, never of the forecasters — these are hard rivers, and a 17–25% typical miss on an April guess at a summer’s runoff is real skill. What this page adds is the number NRCS does not publish beside each forecast: how it has verified, point by point, over five decades.
The dry-year problem
Split every point’s record into its own driest, middle and wettest thirds and the forecasts stop looking even-handed: in dry years the April even-odds forecast runs a median of +30% high; in wet years it runs 10% low. The stated 10–90% band, nominally right 80% of the time, holds in only 73% of dry years at the median point. The operational reading is uncomfortable: the years a water manager most needs the number are the years it is most likely to promise water that never comes. Regression to the mean and thirsty post-drought soils are the textbook explanations — the size of the effect here is the measurement.
Warm bars are statewide-dry years (median point delivered under 80% of its normal year), blue bars wet ones. The driest years on record here — 1977, 2002, 2012, 2013, 2018 — all sit far above the line; 2002’s statewide median miss was +96%. And the skill is not trending better: 1950s 25%, 1960s 16%, 1970s 18%, 1980s 23%, 1990s 20%, 2000s 21%, 2010s 21%, 2020s 22% median miss — the 2020s are so far the roughest decade in the archive, which is what a snow–runoff relationship degrading under aridification would look like.
Where the drought bias lives
Each forecast point, colored by its dry-year bias -- how far the April forecast typically runs above (red) or below (blue) what the river delivers in that point’s own driest third of years. Larger dots carry more graded years; a hollow gray dot is a flagged ⚠ point that misses badly even when graded on its proper basis (see the method notes). Click any dot for its numbers.
What a month of waiting buys
The same points, graded at every issue date: the January outlook misses by a median 28%, June’s by 16%. Most of the improvement arrives late, when the snowpack’s fate is already sealed — which is exactly why the dry-year problem above matters: the early-season number is the one decisions get made on, and it is the least certain.
Do the stated odds mean what they say?
Each forecast publishes five exceedance levels — "90% chance the volume is at least X". Pooling 1,997 point-years that published all five: a value stated with 90% confidence was actually reached 82% of the time; one stated at 10% was reached 10%. A calibrated forecast plots on the diagonal.
Accuracy against calibration
Right is a bigger typical miss; the dashed line is where a calibrated band should sit. Larger dots carry more graded years. Every dot is named on hover and listed in the table below. Reservoir inflows and transbasin-diverted rivers are graded against NRCS’s naturalized volumes rather than their gage (marked in the table); any point whose typical miss still exceeds 100% is left off this chart and flagged ⚠ below (see the method notes).
All 69 points, best typical miss first download CSV ↓
The record-adjusted benchmark
This site also publishes, at some points, a record-adjusted version of NRCS’s April number — the same forecast, re-anchored on three things the record knows: how that point’s April issues have verified, the winter’s baseflow at the gage (a soil-moisture and carryover proxy), and the district’s April 1 snowpack. Coefficients are frozen every August from completed seasons only, so a season’s adjustment can never quietly change under a reader.
It appears only where it has earned it: each point’s model is replayed over every past season without ever seeing the year it predicts, and the adjusted number is shown only where that replay beats NRCS by a clear margin — currently 23 of 67 fitted points. Replayed strictly forward in time (each year predicted using only earlier years), those points’ typical April miss falls from 30.1% to 22.5%, and in their driest third of years from 60.9% to 36.8%. Everywhere else the honest statement stands on the page instead: NRCS’s own number, unadjusted, is the best available.
Honest limits: about two-thirds of the dry-year bias is post-April weather and no April-dated number can remove it; a point whose winter gage record is too ice-gapped to compute baseflow shows no adjusted number that year (a model fed partial inputs is not the model the replay graded); and only the April benchmark issue is adjusted — January–March issues need their own predictor windows, planned for next season. The adjustment is computed by this site and is not an NRCS product; it verifies here annually, win or lose, same as NRCS.
Method, sources and honest limits
Forecasts: NRCS National Water and Climate Center archive (AWDB), element SRVO — the seasonal streamflow-volume forecast, exceedance values at 90/70/50/30/10%. For each point and year, the LAST issue published in April is graded, against the period that issue itself declared (NRCS has changed the window over the decades; each year verifies against its own). Observed: the daily record DWR publishes for the forecast point’s own gage, summed over the declared period; years with under 90% daily coverage are dropped, never silently summed. Median miss is the median over years of |50% forecast − observed| as a percent of observed. Bias is the median signed error: positive means the forecast typically ran high. In-band is the share of years the observed volume fell between the 90% and 10% exceedance values — nominally 80% — computed only over years that HAVE a published band (early decades often carried only the 50% number). Dry-year bias is the median signed error over the driest third of that point’s own gradable years (points with fewer than nine show a dash). The yearly chart shows only years with ten or more graded points; its dryness index is the median across points of the year’s volume over that point’s own median year. Only points whose gage record this site holds are graded (69 of ~80 Colorado forecast points), and points with fewer than five gradable years are left out entirely. The one caveat that matters: at some points NRCS forecasts an ADJUSTED (naturalized) volume -- corrected for upstream storage and diversions -- while the gage records the regulated flow that actually passed, and grading one against the other is a definitional mismatch, not a verdict. Points whose gage-based miss exceeded 100% (reservoir inflows and transbasin-diverted rivers) are therefore graded against NRCS’s own naturalized monthly volumes (AWDB element SRVO, the series NRCS itself verifies against, reaching the 1920s at the long points) and marked (naturalized) in the table — a fair grade of the forecast, though no longer an independent one, since the verifying series is NRCS-derived. A point still flagged ⚠ after that regrade genuinely misses that badly. Skill archive refreshed monthly; this rendering 2026-10-01. Per-point year-by-year charts live on each district’s Snow & runoff tab.