How these figures are produced

A post-processing of the physical model, not a replacement

The AI does not produce a forecast from scratch: it corrects that of the official physical model.

Every day at 06:00 UTC, the pipeline retrieves the reference physical forecast for each station, then a post-processing model trained on that same station’s history corrects its output over the next 48 hours.

For significant wave height, this reference is not the same everywhere: five wave models are compared station by station — Météo-France (MFWAM), ECMWF WAM, DWD EWAM and GWAM, NCEP GFS-Wave — and the one that forecasts the site best becomes its reference. The AI is therefore compared with the best of the five, not with a model chosen once and for all; the selected model is named in the table and on each station page. For water level, the reference is a harmonic tidal analysis.

The post-processing model itself is chosen automatically, station by station, at training, and its identity is named in the same place as the physical reference. Three families exist: a gradient boosting (histogram gradient boosting, “hgb”), a regularised linear regression (“ridge”), or a separate gradient boosting per lead-time band — 1 to 12 h, 13 to 24 h, 25 to 48 h (“hgb-per-lead”). A station running ridge is not one where the model failed to do better: it is the project’s comparison floor, linear regression already capturing the gain at that station without a more sophisticated model adding anything measurable. The model name is published only when it can be identified; otherwise none is invented and nothing is displayed in its place.

The model takes as input the physical forecast itself, the recent error of that forecast against observations, and a 10 m wind forcing. For tide gauges, it predicts only the residual — the difference between the observation and the harmonic — and the published level is rebuilt as “harmonic + predicted residual”.

The next day, the forecast issued the day before is compared with the observations that actually arrived, and the day’s mean absolute error is computed for the AI and for the physics alike, on exactly the same points. Nothing is rewritten after the fact: a day without usable observations stays marked as missing and enters no average.

All timestamps of the scoreboard, its API and its charts are in UTC.

A station that gains nothing publishes no AI figure

A station is published only if the model gains at least 5% of MAE over the physical forecast there. The others stay displayed, without a figure.

Below that threshold, the model finds no usable signal in the current data for that station. It stays listed on the scoreboard, tracked, and without an AI figure — removing it would make the others unverifiable. The scoreboard table names the ones currently in this case, and marks it on each row.

A station can also pass the threshold while carrying an additional caveat: the model does not beat its own baseline there once that baseline is debiased. Its displayed gain is then first of all a constant removed from the physical forecast, not meteo-oceanic knowledge. These stations are flagged as such, in the table and on their detail page.

What these figures are not yet worth

Four caveats that bear on the displayed gain itself. They are measured, not rhetorical.

Where the data comes from

The attributions below are licence obligations attached to the data used, not acknowledgements.

Display a station’s verdict on your own site

A compact iframe badge with the station’s 7-day MAE and a link to its detail page on the scoreboard. The attribution line under the badge is part of the code to copy: a link placed inside the iframe remains, for a search engine, a link from our site to itself — only the one placed in your page signals the source.

OceanData Consulting