01 / THROUGH TIME
Does the outlook get closer?
Mean absolute difference from the zero-hour forecast.
Nearest forecast within ±4 hours
Whiskers / shading: exploratory 95% intervals. Empty years stay empty.
FORECAST RESEARCH / WEATHER
Forecasts, in hindsight.
How the outlook changes between six days away and the hour itself.
Forecast reference, not measured weather. Differences use the stored zero-hour forecast. Gaps and changing seasonal coverage limit claims about improvement.
Loading station summaries…
02 / DISTANCE TO TARGET
What changes as the day approaches?
Selected metric across all six nominal leads.
Inspect the selected lead offsets
Offsets are rounded to the nearest hour for this display; matching uses exact seconds.
03 / DIRECTION
Warmer, or cooler?
Share of paired forecasts at the selected lead.
04 / SEASONAL SIGNATURE
A different bias in summer?
Mean signed difference by target month and lead. Select a cell to focus the study.
CoolerWarmer
Positive values mean warmer than the near-term forecast. All selected years combines the page's year range, weighted by paired-hour count. This year selector affects only this chart; all twelve months stay visible. Hover or focus a cell for its sample count and dates.
05 / YEAR BY YEAR
How does the seasonal bias change?
Mean signed difference by year and target month · 6 days ahead, ±4 hours · °F
CoolerWarmer
Rows follow the page's year range; columns always show January–December. This grid has its own lead selector and follows the station, quality and sample settings. Each cell is a paired-hour-weighted mean; missing or future months stay empty. Select a cell to focus the study on that year, month and lead. Both bias grids use the same color scale.
06 / WHAT THE ARCHIVE CAN SUPPORT
Coverage is part of the result.
How many tracked hours have a stored near-term forecast? Select a month to explore it.
07 / INSPECT THE EVIDENCE
Compare the leads for a day.
Daily mean differences and sample counts across all six leads. Quality and sample settings apply.
Methodology & interpretation Reproducible choices, visible limitations
A consistent comparison
Each target hour contributes once per lead. Select the nearest nominal lead within ±4 hours; exact matches win and equal-distance ties use the older forecast. A missing zero-hour reference stays missing. Temperatures are normalized to °F.
What “closer” means
The reference is another forecast from the same system. Shared errors can remain invisible. These results measure forecast agreement, not accuracy against a sensor.
Uncertainty, without false precision
Pointwise 95% percentile intervals use 1,000 resamples of fixed calendar blocks. At least 30 sampled days and eight occupied blocks are required. Try 7, 14, and 28 days. Intervals condition on retained samples; gaps and unequal seasons can bias comparisons. They are exploratory and unadjusted for multiple comparisons.
Quality screening
Temperatures more than two sample standard deviations from the site/month/local-hour baseline are flagged. Baselines pool retained zero-hour forecasts across years within ±2 local hours and require 30 samples and nonzero variation. Sparse groups skip this screen. Broad temperature bounds and adjacent reference jumps are also flagged. Flagged inputs are excluded by default, preserved in the audit log, and can be restored with the quality filter. These rules can remove real extremes and affect apparent performance.
Time & coverage
Dates follow each site's timezone; repeated daylight-saving hours stay distinct. Nominal lead uses the first hour in the saved bundle, not a verified issuance time. Winter in the year chart means January, February, and December of that calendar year.
Sources & provenance
Recorded NWS hourly forecasts from Trackit. NWS product documentation ↗ · Verification metrics ↗