Which model to trust for DC

Every model graded on Washington's own weather: storm-total snow and liquid at DCA, IAD and BWI and across the region, the 500 mb pattern two days out, daily highs and lows at the area's airports, and whether it holds the cold-air wedge east of the Blue Ridge. The Now page uses these records to weight today's forecasts.

Lower is better for the misses, higher for the pattern score. The best value in each column is in bold. Snow is each storm's total at the three airports; highs and lows are the last 30 days at DCA, IAD and BWI, one day ahead; the pattern score is the 48-hour 500 mb anomaly correlation over the lower 48, last 30 days; wedges count every observed cold-air-damming event, one day ahead. A blank means the model isn't scored on that test (NBM has no 500 mb field, for instance) or has too few cases yet.

Snow and liquid: storm by storm

Every storm in the catalog, scored at DCA, IAD and BWI. One "case" is one storm at one station: the model's storm-total forecast at a given lead time, compared with what the airport measured.

forecast–observation pairs from storms. Liquid is the fairer test: most models' snowfall is derived at about 7:1 from frozen precipitation, so a snow score mixes precipitation skill with a snow-ratio guess.

How far off, on average

Average absolute miss of the storm total, by lead time. Lower is better. Points with fewer than 6 cases are left out. Hover for every model's value at a lead time.

Too high or too low?

The average signed error. A model can have a small bias but a large average miss if it errs both ways.

Rankings

1 day out

3 days out

With storms, most of them minor, these rankings are suggestive rather than settled. Each new storm this winter adds three cases per model and lead time.

Placement: model maps vs. the NOHRSC snowfall analysis

The National Operational Hydrologic Remote Sensing Center's gridded snowfall analysis (NOHRSC v2, about 4 km, averaged to the same 0.25° grid) gives a complete observed map to compare against, including where nobody reports. Snow only; storm so far, at the lead chosen here. The analysis is built from the same observers plus a radar and HRRR first guess, so it isn't fully independent of them.

Wrong amount or wrong place?

Left to right: too little or too much snow overall (SAL amplitude; ±0.67 means half or double). Up: how far the forecast's snow would have to be moved to match the analysis, once both have the same total. This is the earth mover's distance, a distance-aware cousin of the L2 error that doesn't punish a near miss twice. Small dots are single storms; big dots are model averages.

Which way was the band off?

Each dot is where the center of the model's sat relative to the analysis's: up means the band was too far north. Hollow dots mark storms whose snow area runs off the edge of the grid, which pulls the center inward.

At what scale is the forecast useful?

The Fractions Skill Score compares the fraction of each neighborhood covered by the in the forecast and the analysis. At 28 km it asks "did it put the snow in the right grid box?"; at 250 km, "did it get roughly the right part of the region?" Above the "useful" line (0.5 + half the observed coverage) the forecast beats a random placement of the same amount. Where a line crosses it is the smallest scale at which that model can be trusted.

The 500 mb pattern: is the RRFS as good as the NAM was?

The NAM was retired on November 3, 2026, after 20 years as NCEP's regional model. Its replacement is the Rapid Refresh Forecast System (RRFS): 3 km, all of North America, out to 84 hours. Its ensemble, REFS, runs to 60 hours. Forecasters long treated the NAM's 48-hour 500 mb pattern as its "wheelhouse," so this tab grades the RRFS on that test against the global models, including ECMWF's AI model (AIFS), which many now consider the best of them. Every 00Z and 12Z run is archived from NOMADS and verified against the ECMWF and GFS analyses. Built .

Anomaly correlation (ACC) measures how well the forecast's pattern of ridges and troughs, relative to normal, lines up with the analysis. 1 is perfect, and 0.6 is the classic floor for a useful forecast. At 48 hours, good models score above 0.95, so the differences that matter are in the second and third decimal places.

Day by day at 48 hours

Each dot is one 48-hour forecast. Lines are a 7-day running mean (14 forecasts). The S1 score (Teweles and Wobus, 1954) compares height gradients, so it rewards placing troughs and ridges correctly whatever their depth. It's the score NCEP used for decades to track its regional models, including the NAM. RRFS has no scores on the North America–Atlantic sector: its domain doesn't cover enough of it.

NAM's last season vs. RRFS now

Raw scores rise and fall with the weather (winter patterns are larger and easier to score well on), so each model is shown relative to the GFS over the same period. The top row is the NAM's final cool season, to , from the NOAA archive on AWS. The bottom row is the last 30 days. If RRFS sits where the NAM did, or to the right of it, it has kept the NAM's edge.

Skill by lead time

Mean score by forecast hour, using the score chosen above. RRFS runs to 84 h and REFS to 60 h, so their lines stop there. The dotted line marks 48 h.

Where the 48 h heights run high or low

Mean error on the 1° grid, against the ECMWF analysis, over the last days. A persistent blue patch means the model digs troughs too deep there (or builds ridges too weakly). Over a few weeks this reflects the weather regime as much as the model itself; the pattern only means something once it holds over a season.

Is REFS confident for the right reasons?

Each dot is one 48-hour REFS forecast: its member spread against the error of its weighted mean. A well-calibrated ensemble scatters around the dashed line. Points well above it are overconfident misses, where the members agreed and were wrong together. Convection-allowing ensembles usually have too little spread at 500 mb.

RRFS's worst 48 h forecasts

CONUS, lowest anomaly correlation first. A bust is more interesting when the global models did well at the same valid time: that points to the RRFS itself rather than a hard pattern.

Near the ground: DMV temperatures and the wedge

The 500 mb pattern is only half of what the NAM was known for. The other half was cold-air damming: the wedge of cold, dense air that northeast winds trap east of the Blue Ridge, which keeps Dulles and Charlottesville at 33° and freezing rain while the mountains warm up. Coarser models tend to scour the wedge out too early and run too warm. This section scores daily highs and lows at 12 airports and tests each model on every observed wedge.

Daily highs and lows

The average miss of each model's daily high (circles) and low (triangles), against the official climate-report values from ACIS. Leesburg (JYO) has no climate report, so its highs and lows come from the hourly observations. RRFS forecasts are archived every 3 hours, so its highs and lows can come out a little less extreme than the hourly models'. Hover a dot for its warm or cold bias.

Does it hold the wedge?

An hour counts as a wedge when three things hold together: north-to-northeast wind at two of Dulles, Charlottesville and Lynchburg; pressure at Lynchburg and Roanoke at least 2 hPa higher than at Elkins, across the mountains; and Lynchburg at least 3 °C colder than Norfolk. An event is at least 6 such hours (October to April). Each model's own forecast fields go through the same test. The chart shows how much earlier or later than observed each model ended the wedge, and how much too warm it ran at the wedge stations during the real thing, averaged over every event.

Every observed wedge, newest first. ✓ means the model had a wedge too, followed by how many hours early (−) or late (+) it ended it and how much too warm or cold it ran under it. ✗ means it never showed one. REFS isn't scored here: its archived mean carries temperature but not the wind and pressure the test needs.