Method

What we compare

We save NWS forecasts over time and compare temperatures with accepted station readings from the same place and published forecast period. Largest miss is the biggest signed difference among compatible forecasts captured in the seven preceding civil days. A positive value means the forecast was higher; a negative value means it was lower. Absolute-error ties use the earliest capture, then stable record order.

This is captured history, not a complete record of every forecast or a ranking of providers. Missing data is never treated as zero. Earlier dates without captured NWS forecasts are unavailable; later captures do not recover those forecasts.

Reading the results

The temperature dials show the prediction range, the latest prediction captured before the target day, and the measured high or low from the selected NWS interval. Other forecast intervals remain in history but cannot compete for that result. Unavailable means no eligible forecast history; validity problems remain unscored.

An unfinished measurement period shows actuals so far without a final miss. An overnight low can remain in progress the following morning. After its interval closes, sufficient observations permit a completed comparison. Incomplete Observation keeps available actuals and comparisons visible but excludes them from aggregate accuracy and completed-day search eligibility. Too sparse withholds summary actuals and misses; individual readings remain inspectable. Late readings and corrections reassess coverage and the result.

Coverage uses audited station reporting cadence, not fetch counts: hourly for KNYC and KATT, and five-minute for the other references. Each accepted timestamp can occupy one nearest expected reporting slot; duplicate readings cannot fill extra slots. Live slots are evaluated after their matching window closes. More than 25% missing slots or a continuous absence of at least three hours makes coverage too sparse. Otherwise, an internal reading gap over 70 minutes or a boundary gap over 60 minutes makes it incomplete. These initial product thresholds are revisited as evidence accumulates.

Unsupported is different from missing data. NWS provides combined liquid-equivalent precipitation and snowfall, but does not supply rain-only forecast amounts. The rain metric is unsupported; combined precipitation is not relabeled as rain. Trace, ranges, missing observations and partial periods remain distinct.

P marks predicted precipitation; O marks the official recorded result. A missing positive icon does not mean it was dry. Snowfall depth and liquid-equivalent precipitation are different quantities.

Sources

Reporting days

Official climate reports use fixed local standard-time days. Calendar dates and the seven preceding capture days use city civil time. Temperature measurements instead follow the selected NWS daytime or overnight interval, including its next-morning end. During daylight saving time these periods can differ, including ordinary days and 23/25-hour transitions.

Accepted temperature readings must have verified quality, supported units and the matching station reference. The start timestamp is included and the end is excluded. Conflicting temperature versions prevent a result. We never interpolate missing readings or substitute a daily climate extreme. Samples can miss peaks between readings, and whole-Celsius reports can round differently from official climate highs or lows. Daily climate reports remain separate with their own period and finality; existing precipitation comparisons still require compatible complete periods. Inspect the day’s readings and Reporting periods for the underlying evidence.

Forecast change

Covered future dates show the latest comparable prediction minus an earlier prediction. We choose the earlier capture with the largest absolute difference from the latest, then display the signed difference. For example, 80, 85, 83 becomes +3°F; 80, 89, 83 becomes −6°F. This is movement in the forecast, not accuracy or a better/worse judgement.

One comparable capture is Tracking; two unchanged captures show zero. Captures come from the seven preceding city civil days, never after the current time or the target day begins. There is no backfill. Earlier ties use the newest earlier capture, then stable record order. Comparisons use unrounded values; chart sampling does not choose the result.

Only identical quantities, units, reference locations and valid intervals are compared. NWS provider high/low periods belong to the city-local date when the period starts, including overnight lows; they are not relabeled as complete daily extrema. Snow needs a complete target-day amount period. New or incomplete periods cannot borrow a different period’s history. Expand Reporting periods for exact intervals.

The chart reference for a future date is the latest prediction, with no observed result. Failed checks leave the last captured value and timestamp. During the temperature interval, measured readings develop without a final comparison; its closing time and observation coverage determine the subsequent result. The chart and expandable snapshot history remain available alongside the dial.

Checks and revisions

A check records an actual retrieval attempt. A revision is a changed forecast representation or a distinct announced provider update. Repeat checks retain their own receipts and remain eligible under the same scoring rules. A successful NWS 304 references its prior response without moving the publication time.

Hourly checks do not mean NWS publishes hourly. Served intervals may merge equal hourly values and do not establish native resolution. Failed or missed checks remain gaps. Before recorded schedule history, coverage uses nominal hourly slots with unknown historical cadence; no missed forecasts are replayed.

Captured history

These dates and counts describe NWS forecasts actually saved here, including repeat checks. They do not establish complete coverage between the first and last capture or imply earlier NWS history exists.

As of Sep 8, 2026, 8:21 PM UTC