Forecasts you can check.

A retrospective test on public data: the system forecast 90 days ahead from 13 past dates, and the forecasts were scored against what actually happened. It is not a live track record. The live track record (below) is separate and only counts days after each forecast was issued.

183 series: 63 Wikipedia articles from a fixed list, and 20 cities x 6 weather variables. Forecast start dates every 14 days; 3 early dates (Oct-Dec 2025) were used only to pick each series' simple reference method, and the 10 later dates (25 Mar to 29 Jul 2026) were scored. Numbers below are ratios of average absolute error: below 1.00 means the system's error was lower than the reference, above 1.00 means higher. Each series counts equally (geometric mean across series). Each forecast used only data up to its start date.

Versus each series’ best simple reference (picked on the early dates)

The reference is one of seven simple methods (repeat last value, same weekday, same day last year, recent average or median, climatology). Not a fixed method, so read the other two tables as well.

Versus repeating the last value

A weak reference at long range, so these ratios flatter the system.

Versus a seasonal reference

Wikipedia: repeating the same weekday last week. Weather: the typical value for that date in the previous four years (climatology).

Where it lost, and a few clear wins (31-90 day window)

Ratio versus the selected reference. Worst 8 of 183 series, and the 5 best excluding event articles (Easter, Christmas and similar had no event inside the test window, so their very low ratios say nothing about event skill and are left out).

How to read this

• On this test the system's average error was lower than the selected simple reference on most series: about 5 to 20 percent lower for weather, more for Wikipedia pageviews. That is a descriptive result for these 183 series, not proof for others.
• The Wikipedia gap is probably overstated: references were picked on Oct-Mar data and scored on Mar-Nov, so seasonal mismatch favours the system. Against the seasonal reference it is about 0.55 to 0.66.
• Rain: the smooth line is the median, the typical outcome. It is not a forecast that a particular day will be wet or dry, and the system does not predict individual storm days. The benefit measured here is a small drop in average error, and the band is the real picture.
• Bands (10th to 90th percentile) are raw and uncalibrated. Median coverage across series is 0.78 to 0.89, but it varies a lot by series and horizon, so do not read them as exact 80% ranges.
• Only simple references were tested, not professional weather services or other statistical models. Data are current revised values, not as first published. Forecast start dates overlap, so error bars would be wider than a simple count suggests.
• Full per-series results and code are in the private repository.