Average absolute change
Forecast revisions
How forecasts change as the day approaches
Daily changes became smaller as the forecast window approached. Updates improved average accuracy, though only about three in five changed forecasts moved closer to the observation. Large late revisions identified forecasts with more error still remaining.
Day 2 to day 1
Average remaining day-1 high error
Forecast updates improved accuracy on average, but an individual update was far from certain to help.Among forecasts whose median changed, 57–63% moved closer to the observed temperature, depending on forecast day and whether the forecast was for a high or low.
Following the same temperature window through successive forecasts
A forecast for a Saturday high can appear in many daily NBM runs. Comparing “day 7” from one week with “day 6” for a different date would mix the update with a new weather event. This analysis instead matches the same station, high or low, and valid-window ending time in forecast runs exactly 24 hours apart.
The source temperatures are whole degrees. A zero-degree update can therefore mean that the underlying guidance changed by less than the published precision, or that it did not change. The analysis retains those ties when calculating average update size and separately asks whether forecasts that did change moved closer to the observation.
Average revisions became smaller near the forecast window
Absolute change between daily forecasts
Each point compares two forecasts for the same station and temperature window.
High revisionLow revision
High forecasts changed by 1.89°F on average from day 10 to day 9. The average fell to 1.61°F from day 6 to day 5, 1.23°F from day 4 to day 3, and 0.89°F from day 2 to day 1. Low revisions declined from 1.69°F to 0.77°F across the same span.
This decline is an average over hundreds of forecast dates and many stations. It means that the typical daily update became smaller as more observations entered the forecast system and the valid window approached. It does not mean every forecast changed smoothly or became fixed at a particular day.
| New forecast day | High revision | High improvement rate | Low revision | Low improvement rate |
|---|---|---|---|---|
| 1 | 0.889°F | 62.84% | 0.768°F | 58.60% |
| 2 | 1.072°F | 62.25% | 0.899°F | 62.16% |
| 3 | 1.234°F | 60.30% | 1.011°F | 59.14% |
| 4 | 1.457°F | 60.68% | 1.198°F | 57.84% |
| 5 | 1.608°F | 60.68% | 1.353°F | 57.99% |
| 6 | 1.744°F | 59.43% | 1.506°F | 59.02% |
| 7 | 1.824°F | 58.66% | 1.626°F | 57.96% |
| 8 | 1.834°F | 57.72% | 1.668°F | 57.57% |
| 9 | 1.892°F | 57.39% | 1.691°F | 57.32% |
New forecasts were better on average
| Transition | High revision | Older high error | Newer high error | Changed highs improved | Low revision | Older low error | Newer low error | Changed lows improved |
|---|---|---|---|---|---|---|---|---|
| Day 8→7 | 1.824°F | 5.250°F | 4.781°F | 58.66% | 1.626°F | 4.953°F | 4.543°F | 57.96% |
| Day 2→1 | 0.889°F | 2.633°F | 2.343°F | 62.84% | 0.768°F | 2.981°F | 2.810°F | 58.60% |
From day 8 to day 7, average high error fell by 0.469°F and low error fell by 0.410°F. From day 2 to day 1, high error fell by 0.290°F and low error by 0.171°F. All four changes were improvements at the aggregate level.
The case-by-case result is more restrained. Among values that changed, 58.66% of day-7 highs improved and 57.96% of day-7 lows improved. Day-1 highs performed best in the selected transitions, with 62.84% moving closer. Day-1 lows improved 58.60% of the time. A revised forecast was therefore more likely to help than hurt, but a user should not assume the newest change is automatically in the right direction.
A large late revision was a warning about remaining error
| Day-1 forecast | No revision | Revision of at least 5°F |
|---|---|---|
| High average remaining error | 2.16°F | 4.22°F |
| High missed by at least 5°F | 9.1% | 32.1% |
| Low average remaining error | 2.76°F | 4.49°F |
| Low missed by at least 5°F | 16.1% | 37.3% |
The strongest practical result concerns the size of the latest change. Day-1 high forecasts that had not changed from day 2 averaged a 2.16°F remaining error. When the latest revision was at least 5°F, the remaining error averaged 4.22°F. The chance of missing by at least 5°F rose from 9.1% to 32.1%.
Overnight lows showed the same pattern. With no latest revision, day-1 lows averaged a 2.76°F error and missed by at least 5°F in 16.1% of cases. After a revision of at least 5°F, average error rose to 4.49°F and the 5°F-miss rate reached 37.3%. The large revision improved many forecasts, yet it also identified situations in which uncertainty had not been fully resolved.
The signal also appeared farther ahead
| Day-7 forecast | No revision | Revision of at least 5°F |
|---|---|---|
| High average remaining error | 4.19°F | 6.63°F |
| Low average remaining error | 4.21°F | 5.81°F |
At day 7, high forecasts with no daily revision averaged 4.19°F of remaining error, compared with 6.63°F after a revision of at least 5°F. Low forecasts averaged 4.21°F and 5.81°F, respectively. The relationship was therefore not limited to the final update.
These bins describe association, not a calibrated alert probability. The archive does not show that changing the forecast creates the later error. A difficult or rapidly evolving weather situation can cause both a large update and a larger remaining miss.
How a forecast user can interpret a change
A zero revision is modestly reassuring because that displayed group had lower historical error. A large late revision deserves more caution even after accepting the new value. The practical response is to check the likely range, the current NBM range width, and any decision threshold rather than focusing only on whether the revised median moved warmer or cooler.
This is especially important for threshold decisions. If a forecast changes from 31°F to 36°F, the newest value may be more accurate on average, but the large revision marks a group with above-average remaining error. A likely range crossing freezing conveys that uncertainty more honestly than the 36°F median alone.
The NBM range-width report provides a complementary signal. It tests the width of the predictive range within one forecast run, while this report tests instability across successive runs. Both are historical indicators of error size; neither specifies the error direction in an individual case.
Bottom line: revisions generally shrank and improved accuracy as the forecast approached. Large late changes remained useful warning signs because forecasts in that group still missed by more than stable forecasts.
Methods
Forecast pairs match exactly on temperature type, station, and valid-window end. The newer forecast run is 24 hours after the older one, both forecasts must be usable, and pairs crossing an NBM version boundary are excluded. Each transition contains 694 matched forecast dates. Station metrics are averaged within a date before dates are averaged. Improvement rates among changed forecasts treat ties separately. Confidence ranges use three-valid-date moving blocks.
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.