Explainable forecasts for the two characteristic thermal winds at northern Lake Garda:
- Peler — the northerly morning wind, roughly 04:00–09:00;
- Ora — the southerly afternoon wind, roughly 12:00–17:00.
Open the live dashboard · Read the model card · Review data licences
Garda Oracle turns openly reusable weather observations and forecast fields into separate three-day probabilities for Ora and Peler. It also shows live wind at Torbole, the strongest factors behind each forecast and a 30-day comparison with observations. The interface is available in German, English and Italian.
Weather models provide pressure, cloud, radiation and wind fields, but they do not directly answer the question a rider asks: Will the local lake wind actually establish itself?
The project started with a transfer question. An earlier hobby project, Walchi Thermic Oracle, forecasts the local thermal at Lake Walchensee. Rather than copy that model, Garda Oracle reused the methodology: define measurable local outcomes, assemble relevant large-scale weather signals, validate honestly across years and test whether skill survives the move from reanalysis to real forecast inputs.
At Garda the answer was unusually clear. Two independent logistic models substantially outperform a month-of-year climatology:
| Regime | Base rate | Leave-one-year-out Peirce | Climatology Peirce |
|---|---|---|---|
| Ora | 69% | +0.54 | +0.22 |
| Peler | 53% | +0.39 | +0.20 |
On 513 days from 2024–2026, features from the previous day's actual forecast run retained the signal: best-threshold Peirce was +0.59 for Ora and +0.49 for Peler. These results are described with their limitations in the model card.
flowchart LR
OM["Open-Meteo forecast fields"] --> F["Shared feature builder"]
MT["Meteotrentino T0193 live wind"] --> UI["FastAPI dashboard"]
F --> O["Ora logistic model"]
F --> P["Peler logistic model"]
O --> UI
P --> UI
The models are trained offline with scikit-learn. Serving needs none of the ML
stack: coefficients, scaling values, calibration parameters and decision
thresholds are frozen in src/garda/coeffs.py and scored with the Python
standard library.
The same pure-Python feature builder is used during coefficient export and live scoring, which prevents a separate training implementation from silently drifting away from production.
See architecture for the request and training flows.
Prerequisite: uv.
git clone https://github.com/s1st/garda-oracle.git
cd garda-oracle
uv sync --frozen --extra dev
uv run uvicorn garda.web:app --reload --port 8765Open http://127.0.0.1:8765/. No credentials are needed. The local dashboard uses the public Open-Meteo and Meteotrentino endpoints.
Quality checks:
uv run --frozen pytest -q
uv run --frozen ruff check src tests scripts
uv run --frozen mypy srcRaw observations and forecast downloads are not committed. They are recreated using public endpoints:
uv sync --frozen --extra dev --extra ml
uv run python scripts/fetch_torbole.py --end-date 2026-07-03
uv run python scripts/build_labels.py --end-date 2026-07-03
uv run python scripts/fetch_era5_garda.py --end-date 2026-07-03
uv run python scripts/fetch_prevruns_garda.py --end-date 2026-07-03
uv run python scripts/export_garda_coeffs.pyThe dates above reproduce the training snapshot used by the committed model. Providers may correct historical records later, so a future download is not guaranteed to be byte-identical. The committed coefficient golden-vector test protects the served model from accidental changes.
See reproducibility for the full workflow, expected files and validation commands.
src/garda/
├── features.py # forecast retrieval and shared feature construction
├── model.py # pure-Python logistic scoring and explanations
├── coeffs.py # generated model artefact
├── history.py # 30-day day-ahead hindcast versus observations
├── i18n.py # German, English and Italian copy
├── web.py # FastAPI application and caches
└── templates/ # Jinja2 dashboard pages
scripts/ # public-data download, research and model export
tests/ # scorer, feature, history, i18n and web tests
The included Dockerfile builds the same application used by the public site.
It does not contain credentials. GARDA_GATE_SECRET can optionally protect a
raw Cloud Run origin when a reverse proxy injects the corresponding
X-Gate-Secret header. CF_BEACON_TOKEN optionally enables cookieless
Cloudflare Web Analytics. The maintained deployment also accepts the exact
custom host garda.s1st.de without the proxy header so its INWX-managed DNS
can reach Google's custom-domain frontend directly.
The checked-in cloudbuild.yaml contains the maintainer's current GCP resource
names as non-secret deployment configuration. Forks should replace the image,
service and region substitutions with their own resources.
Copyright © 2026 Simon Stieber. Project-authored code, documentation and frozen model parameters are available under the GNU Affero General Public License v3.0 only. Weather data remain subject to their source licences; see DATA_SOURCES.md for exact attribution and distribution boundaries.
This is a decision aid for wind sports, not an official weather warning or safety service. Local wind can differ materially from a single station or grid-cell forecast. Check current observations, official warnings and conditions on site.
Contributions are welcome; start with CONTRIBUTING.md.