Metroblue revising past data to game comparisons?
-
Looks to me like metro blue is revising its past forecasts to make its actual vs forecast comparison results look better.
Example- last night it forecast a low of 5 degC at 5 am for this morning.
Now today the forecast and observation vs forecast shows that it predicted a 5am today temp of 9 degC which just happens to match the actual bang on!
I have been wondering how the AI was getting such low variation numbers in the comparison, when I had noticed it been off quite a bit over the week… Looks to me like the AI is cheating?



-
@teonline Hi there,
We completely get why that looks suspicious at first glance, but rest assured, neither Windy nor Meteoblue "cheats" by rewriting past forecast history to fake benchmark scores!
What you observed is actually the intended behavior of Meteoblue’s Learning MultiModel (mLM) and its real-time post-processing pipeline:
-
How the mLM Model Works:
Meteoblue doesn't just run a single static physics model. Its mLM engine blends output from dozens of global and high-resolution numerical models (like ECMWF, GFS, ICON) with machine-learning algorithms trained on over 100,000 quality-controlled ground stations to correct local biases and terrain effects. -
Short-Range Post-Processing (Nowcasting):
Unlike raw global models that only update every 6 or 12 hours, Meteoblue incorporates continuous hourly post-processing (Kalman filtering and nowcasting) using real-time observations, satellite, and radar data. When you looked last night, you saw an earlier forecast run. As 5:00 AM approached, fresh local observations corrected the trajectory toward the emerging reality (9°C). Naturally, the prediction made right before 5:00 AM will closely mirror the final actual observation. -
How Verification Benchmarks Are Actually Scored:
Official verification reports (such as the ones published monthly by Meteoblue) do not use those last-minute nowcasted numbers to grade their day-ahead performance. They evaluate forecasts locked strictly at fixed lead times (e.g., 24-hour ahead, 48-hour ahead, etc.) against independent ground stations. An earlier forecast error is counted as an error in those statistics—it is not retroactively erased.
You can always compare this against traditional raw NWP models (like raw ECMWF or GFS) using the bottom model bar in Windy, which stick to their fixed initialization cycles without rapid observational blending.
Hope this clarifies how the technology works under the hood!
-
-
Hi. Not surprised by your answer, suspected it had something to do with the continual updating of the AI. But… The charting and comparison numbers in the observations versus forecast display within Windy doesn’t reflect your comments about how the verification data is kept to fixed intervals for comparison. It’s nice that the monthly reports do this- but as a user I’ve never seen these monthly verification reports, so I rely on the in app observation versus forecast display to judge which model is best for my needs.
I would suggest that the in app comparison is currently giving a very biased impression of the AI accuracy to your average user, and should be changed to reflect how the monthly report works. -