We’ve got another round of enhancements in ForecastWatch.
The new app is now the home for forecast accuracy
Forecast accuracy browsing, CSV export, and long-term trends all live in the new ForecastWatch platform. The older app.forecastwatch.com address and the legacy pages on the original site now redirect there automatically: a bookmark to the old drilldown view lands on Monthly Insights, the old trends view lands on Long-Term Trends, and the daily forecast-error pages land on Daily Insights. Pages without a direct successor send you to the new app’s home. Existing bookmarks keep working, just pointed at the equivalent area of the new app, so there’s nothing to relearn or re-save.
Data Export presets save on the first try
Saving a Data Export configuration as a preset now works whether or not you’ve adjusted every filter first. Opening Data Export and immediately choosing “Save as preset” succeeds, capturing the metric, days-out, and location values shown as defaults rather than only the ones you happened to touch. A saved preset now stores the exact filter set the export itself uses, so re-applying it reproduces the same export. The Save button gives a clear prompt when a genuinely required field, like the month, is still missing.
Methodology docs: a left rail that scrolls, and links you can share
The methodology documentation’s left navigation now scrolls on its own when it runs longer than the screen, so every section stays reachable. On a narrow window the navigation stays compact and the document area shrinks to fit instead of the menu stretching. Each documentation page now has a stable in-platform link: visiting https://app3.forecastwatch.com/docs/collection/providers/ opens the Providers page directly, and the address bar updates as you click through. You can browse to any page, copy the address, and share it with a colleague, and it opens that exact page inside the platform.
UK SYNOP observations and scores, restored for the spring 2026 feed gap
Forecast accuracy for 37 UK SYNOP stations is now recoverable for the 27 March to 26 May 2026 period. Our observation feed stopped delivering the UK SYNOP network at 14:00 UTC on 26 March and resumed on 27 May, leaving a two-month gap with no temperature, precipitation, wind, or cloud observations, and therefore no forecast accuracy scores for those stations and dates. We recover that gap from OGIMET, an independent archive of the same WMO GTS bulletins. The recovery runs the full pipeline: recovered observations are backfilled as actuals, every affected provider forecast is rescored against them, and the monthly regional and provider accuracy aggregates for the UK, Europe, and global are re-aggregated to fold the recovered scores in. Daily high/low, precipitation, wind, and cloud actuals, the per-forecast scores, and the rolled-up accuracy numbers customers see are all populated for the affected stations and dates exactly as a normal day would be. Coverage is near-complete: 23 of the 37 stations recover at least 99 percent of hours and the rest recover 82 to 99 percent, totalling about 53,000 verified hourly observations. Recovered temperatures match the live feed to within 0.1 degrees C. UK SYNOP observations, scores, and aggregates are correct for this window from the day of the backfill onward.
Open-Meteo forecasts, downscaled to true station elevation
Every model we source through Open-Meteo now downscales to each station’s exact elevation. This covers the Open-Meteo blend and the individual models we pull from Open-Meteo: ECMWF (IFS and AIFS), GFS, ICON, GEM, JMA, KMA, and their ensemble versions, 15 providers in all.
Open-Meteo lapse-rate-corrects temperature to whatever elevation it has for a point. We now hand it the real station elevation rather than letting it estimate from its 90 m terrain model. For most stations the two agree within about 6 to 10 meters and the forecast barely moves, a fraction of a degree. At high-relief mountain stations, where a terrain model and the actual station can sit at very different heights, this removes a systematic temperature bias of up to several degrees.
Because the correction takes effect at deploy time and we are not rescoring history, the measured temperature accuracy of these Open-Meteo-sourced models at affected mountain stations shifts slightly from the deployment date forward. Scores before that date reflect the old DEM-estimated elevation; scores from the deployment date onward reflect true station elevation and are the ones to rely on for these stations.
Clearer attribution when a day is excluded from scoring
When a station-day is excluded from scoring and no audit row records the specific reason, the monthly station-detail and observations APIs now return a neutral category: "unrecorded" / source: "unknown", labeled “Excluded from scoring; reason not recorded.” Previously these days were tagged observation_data_gap, which implied the observation was missing or bad. In most of these cases the observation is perfectly valid and the exclusion happens at the score level, so the old label pointed at the wrong thing. The neutral values clarify that the observation may be fine and the day was simply not scored. The /v1/observations endpoint continues to correctly return no invalidation for those same days, and the API Portal endpoint docs now list the new unrecorded / unknown values. This change is additive to the response contract; existing categories and sources are unchanged.
