How accurate are temperature forecasts?

Raw NBM error grows with forecast length

A single accuracy percentage cannot describe the full distribution. Average error shows the normal scale of a miss. The within-5°F rate answers how often a forecast stayed reasonably close. The at-least-10°F rate captures the less common outcomes that matter most when a decision depends on crossing a temperature threshold.

The decline is gradual rather than a cliff at day 7. Day-1 highs averaged a 2.34°F error; day-3 highs averaged 2.97°F; and day-7 highs averaged 4.78°F. By day 10, the average reached 6.11°F. Overnight lows followed a similar curve.

The largest errors grew much faster than the average. Only 1.36% of day-1 high forecasts missed by at least 10°F. The rate rose to 6.41% at day 5, 11.88% at day 7, and 19.87% at day 10. A day-10 forecast was therefore not merely a few degrees less precise on average; it also carried a substantially larger chance of a major miss.

What the day-7 result means in practice

For a routine plan with several degrees of flexibility, the two-thirds within-5°F figure may be more useful than the average. For frost protection, energy demand, outdoor work limits, or heat precautions, the tail of the distribution matters more. The historical 10°F-miss rate describes how often errors of that size occurred across the archive. It is not the probability for a particular forecast today.

The uncertainty shown with a current forecast can refine that historical baseline. The model-spread analysis finds that forecasts with wider NBM ranges had much larger errors than narrow-range forecasts issued at the same lead.

Winter was harder than summer in this archive

December day-7 errors were 2.76°F larger than July for highs and 2.91°F larger for lows. The contrast was visible even at day 1, though it widened farther ahead. January and February also had larger errors than July in the selected results.

Monthly estimates contain 54–62 forecast runs, providing about two years of data for each month. They describe the weather and model versions present in this archive. A longer record is needed before treating the differences as a stable climatology for every winter and summer.

Regional differences became larger farther ahead

The Northern Rockies and Plains had the largest errors among the broad regions shown. Its day-10 high error was 8.30°F, compared with 4.99°F in the Southeast and 5.04°F in the West. At day 1, the regional differences were much smaller.

The regions are state groupings used by the calibration system. They summarize broad patterns but do not capture every local climate or terrain feature. The reporting network is also denser in the eastern United States. For a place-specific view, the local accuracy pages show the station record behind each represented city.

Overnight lows were harder nearby; highs were harder farther ahead

Lows had 0.47°F more average error at day 1 and 0.29°F more at day 3. The difference was unclear at day 5. By day 7, highs had 0.23°F more error, and the gap grew to 0.40°F at day 10.

There is an important timing detail. A high and low carrying the same forecast-day label do not have exactly the same lead: the high-temperature window ends 12 hours later. Some of the longer-range high disadvantage may come from that additional lead time. The archive cannot separate that timing effect from physical differences between forecasting daytime highs and overnight lows.

The raw 80% range covered fewer than 80% of outcomes

The unadjusted NBM P10–P90 range is intended to describe the central 80% of possible outcomes. In the archive it contained 76.2% of day-1 highs and 69.3% of day-1 lows. At day 7, coverage was 76.8% for highs and 72.7% for lows. Outcomes exactly on a whole-degree endpoint count as inside.

This shortfall helps explain why ModelSpread adds a calibrated likely range to current forecasts. Details of that product audit, including the coverage-versus-width tradeoff and the absence of prospective results, now appear in the calibration validation methodology.

Download the report metrics and source hashes.