Average across highs, lows, and days 1–10
Uncertainty and range width
Does a wider NBM range warn of larger errors?
Yes. Within the same forecast day, wide NBM ranges were followed by substantially larger errors than narrow ranges. The difference grew farther ahead and appeared for both highs and lows.
Wide range; narrow was 3.01°F
Wide-range highs missed by more than 10°F
A wide range carried information beyond the number of days ahead.Across the 20 high, low, and forecast-day groups, wide-range forecasts missed by 2.69°F more than narrow-range forecasts on average. Every wide group had more error than its corresponding narrow group.
What model spread means here
The National Blend of Models provides several percentiles rather than a single temperature. This analysis defines forecast spread as the distance between its P10 and P90 forecasts. This predictive range incorporates uncertainty in the NBM system, but it is not a direct measurement of the differences among individual model members.
Range width normally increases with forecast length, just as forecast error does. Comparing a wide day-10 range with a narrow day-1 range would therefore mix two signals. To avoid that problem, the analysis defines narrow and wide separately for every forecast day and for highs and lows. Narrow and wide are the lowest and highest width groups within each combination. They were intended as fifths, but source temperatures are whole degrees and many ranges share the same width, so tied cases stay together: narrow holds 20% to 39% of a combination’s cases and wide holds 20% to 30%. The largest narrow group is day-1 highs, at 39%.
Wide ranges were followed by larger misses
High-temperature error by range-width group
The narrow and wide groups are recalculated at each forecast day, so each comparison holds forecast length fixed.
Narrow high rangeWide high range
| Day | High narrow | High middle | High wide | Low narrow | Low middle | Low wide |
|---|---|---|---|---|---|---|
| 1 | 1.79°F | 2.23°F | 3.20°F | 2.30°F | 2.74°F | 3.52°F |
| 2 | 1.91°F | 2.49°F | 3.78°F | 2.40°F | 2.88°F | 3.71°F |
| 3 | 2.11°F | 2.72°F | 4.02°F | 2.71°F | 3.21°F | 4.02°F |
| 4 | 2.30°F | 3.10°F | 4.65°F | 2.85°F | 3.48°F | 4.44°F |
| 5 | 2.51°F | 3.56°F | 5.33°F | 2.91°F | 3.71°F | 4.92°F |
| 6 | 2.73°F | 4.10°F | 6.20°F | 3.10°F | 4.03°F | 5.37°F |
| 7 | 3.01°F | 4.56°F | 6.88°F | 3.32°F | 4.46°F | 6.06°F |
| 8 | 3.27°F | 5.08°F | 7.47°F | 3.52°F | 4.84°F | 6.55°F |
| 9 | 3.62°F | 5.53°F | 7.89°F | 3.74°F | 5.30°F | 7.33°F |
| 10 | 3.70°F | 5.89°F | 8.36°F | 3.97°F | 5.68°F | 7.76°F |
The relationship appeared at all ten leads. For day-1 highs, wide-range forecasts averaged a 3.20°F error, compared with 1.79°F for narrow ranges. At day 7, the corresponding errors were 6.88°F and 3.01°F. At day 10, wide ranges averaged 8.36°F, more than twice the 3.70°F error of narrow ranges.
Overnight lows followed the same pattern. The wide-minus-narrow difference was 1.22°F on day 1, 2.74°F on day 7, and 3.79°F on day 10. Range width was therefore useful for both temperature products, even though highs and lows describe different 18-hour windows.
The signal was especially strong for large errors
| Day-7 forecast | Narrow range | Wide range | Wide divided by narrow |
|---|---|---|---|
| High missed by more than 10°F | 2.55% | 22.55% | 8.8 times |
| Low missed by more than 10°F | 2.57% | 17.57% | 6.8 times |
On day-7 highs, more than one in five wide-range forecasts missed by over 10°F. Only about one in forty narrow-range forecasts did. Day-7 lows showed a similar, though somewhat smaller, separation.
This is why average error alone understates the value of spread. Wider ranges were associated both with a larger routine miss and with a much greater frequency of severe misses. A current wide range cannot tell a user that a large error will occur, but it identifies a class of forecasts in which large errors were much more common historically.
A narrow range is reassuring, not a guarantee
The narrow groups still contained large misses. About 2.6% of narrow day-7 high and low forecasts missed by more than 10°F. Model guidance can cluster tightly around the wrong answer when the models share inputs, assumptions, or unresolved processes.
Spread also does not tell direction. A wide range signals that the eventual temperature is less precisely located, but it does not say whether the median will be too warm or too cool. Directional bias requires separate historical evidence by location, season, and lead.
Why this result matters for a current forecast
A national day-7 average treats every current forecast alike. Range width is one input to the calibration that estimates a likely temperature range. On the current forecast cards, the background color shows the calibrated mean absolute error from the past year of rolling tests, averaged for the station’s region, the element, and the lead time. It is not specific to the location or season, and it does not directly encode the range width.
The raw P10–P90 endpoints should not be read as exact 10th- and 90th-percentile guarantees. Across the archive, raw temperature ranges contained fewer than 80% of outcomes in many groups. The calibrated range adjusts the historical coverage and bias while retaining spread as information about the current forecast. The validation and tradeoffs of that adjustment are documented in Methodology.
What the comparison does and does not establish
The analysis demonstrates a strong descriptive relationship after forecast day and temperature type are held fixed. It does not prove that a wider range causes error. Both range width and error can rise because the weather situation is harder to represent, and the archive does not include the atmospheric variables needed to identify those causes.
The national result also combines climates and seasons. A given width may have different implications in a mountain valley and at a coastal airport. The relative-width grouping reduces some of that problem because it compares widths within the same forecast product, but a fuller local analysis will require more years of data.
Bottom line: the width of the current NBM range is useful evidence about likely error size. It adds context to a forecast-day average, but it cannot specify the direction of the error or eliminate the possibility of a large miss when the range is narrow.
Methods
The analysis used 21,157,988 usable station forecasts from 710 forecast runs initialized at 12:00 UTC and 1,766 stations. Maximum-temperature leads were 42–258 hours and minimum-temperature leads were 30–246 hours. Width groups were defined separately within each temperature type and lead. Station metrics were averaged within each forecast run before runs were averaged. The overall wide-minus-narrow estimate gives equal weight to the 20 product-and-lead groups. Crossed P10–P90 ranges were excluded.
This analysis uses the temperature verification archive snapshot generated August 12, 2026. Unless noted, results describe past performance of archived, unadjusted NBM guidance. Nearby stations and adjacent dates often experience the same weather, so the observations are not fully independent.