The collector also runs automatically every 15 minutes. Use this when newer morning data may improve the afternoon forecast.
Lake wind--
Gust--
Direction--
Temperature--
Loading latest lake observation…

Wind report
Plan a session

Your window

These are your own settings, not this location's threshold - the forecast tier shown above is unchanged.

01 · Forecast

Forecast vs observed

HourFcst windFcst gust Wind scale Observed avg / peak
Forecast range Forecast wind Forecast gust Observed mean-to-peak span Good hour
Open the full report Evidence, operations & evaluation — collapsed

11:00–20:00 local time (Europe/Zurich)

Today

Tomorrow

Daily best · longer-range guidance

Today's wind readings

Lake station · sustained wind and gusts · knots

Observed wind at Lake Silvaplana during the current Europe/Zurich calendar day. The solid and dashed lines are measurements; the dotted line is what the model forecast for each hour, shown only where the forecast was issued before that hour had started.

Historical wind readings

Published lake observations · choose a prior date

Browse measured wind and gusts from earlier dates, with the forecast that was on the board before each hour started. The wind and gust lines are station observations; the dotted line is the model.

Loading historical lake observations…

MeteoSwiss local forecast

Silvaplana · official three-day outlook

Daily summary

Today · observed measurements
Tomorrow · forecast

Lake water temperature

Live performance

Verified real-world forecasts

How the deployed model has actually done, checked against the real Silvaplana lake reading (kitesailing.ch primary) with MeteoSwiss Segl-Maria (SIA), ~4km from the lake, as the reference when no lake reading exists.

Station & ingestion health

Operational status of every data source behind the forecast above
Lake sampler (kitesailing.ch, primary ground truth)
Attempt history, collection coverage & failure categories
Summit stations (Corvatsch, once confirmed)
No summit station data yet.
Forecast station data
Ground-truth source mix (verified predictions)

Signal visuals

The dataset at a glance — rideability mix, seasonal shape, and what the model weighs
Rideable vs not (all labeled hours)
Session rate by calendar month
Top model weights (|magnitude|)
Model evaluation & diagnostics

Frozen holdout evaluation

Not the same as live performance above

A separate model, trained only on past seasons and scored once against a season it never trained on, then frozen - this is the honest answer to "how would this do on data it's never seen," and it does not change as the live model keeps learning.

Passo del Bernina feature status
Run fresh evaluation

Opens GitHub Actions in a new tab. Select "Run workflow" and enable the run_backtest option.

Feature ablation

Diagnostic, not model selection

Each row is its own model, trained fresh with only the listed features, scored once against the same frozen holdout above. This explains roughly how much each feature group contributes - it is not used to pick a winner, and the deployed model always uses the full feature set.

Technical details

Operational alert tiers (frozen holdout)

Reproducibility & deployment

What the model listens to

The signals it weighs, strongest first. These shift automatically as the model learns.

Rideable days by month

Wind history & model confidence

Actual wind (knots) Model confidence Verified live forecasts

Known limitations

  • Historical retraining and the live loop now share the same SIA-first labeling criterion (lake reading first, Segl-Maria/SIA as reference). SIA is ~4km from the lake; its exact equivalence to the spot is still being measured as lake readings accumulate.
  • Backtest features come from Open-Meteo's 0-hour archive, not a genuine forecast with up to five days of lead time - expect live accuracy below the frozen holdout's, especially on the compact longer-range cards.
  • The 2026 holdout is a partial season; trust the live accuracy number as it accumulates.

Choose wind report

Configured winds and monitoring locations

The reports here are not all the same kind of report. Silvaplana runs the learned model: a probability per hour, verified against the lake and retrained from its own results. Corralejo runs a forecast-only heuristic — a deterministic 0–10 score with no trained model, no verification history and no calibration behind it, on wind thresholds still measured at the GCFV airport rather than at the spot. Other locations may appear as preliminary pages: measured climatology and analyzed sessions only, with no forecast of any kind — the page says so itself. Each location keeps its own profile, observations, forecasts, model and dashboard data, so their records never mix. A location appears here only once its own dashboard payload actually exists.

exported 12:55 · data 12:55 UTC