← Back to Caster
Waiting for the first learning cycle…
Completed live learning updates—
Targets waiting for observations—
24-hour rain training examples—
A learning update means one measurement and horizon, not an independent weather event. New forecasts are saved once per hour. Completed results are checked every five minutes.
Three-hour temperature comparison
Each forecast uses the same starting observation and is checked against the same temperature reading three hours later.
Waiting for matched forecast results.
Average absolute error over the last 28 days, in °C. Lower is better. Nearby forecast periods overlap.
Cloud drift
Loading the camera trial…
Directions show where the cloud or wind is coming from. A difference can indicate directional wind shear. The camera estimates direction from visible cloud movement.
Three-hour rain trial
Waiting for completed rain windows.
Both forecasts use the same station readings. The second also uses cloud drift. Lower Brier scores are better. Rain means at least 0.2 mm during the forecast window.
Under Pressure
Pressure forecast trial at 3, 6 and 12 hours. Average error is measured in hPa. Lower is better.
Loading the pressure trial…
Coverage is the share of captured Caster forecasts for which the trial had enough usable pressure history.
Live accuracy
Lower error is better. Temperature errors are °C, pressure mbar, humidity percentage points, wind km/h. Rain uses Brier score (lower is better). Daily maximum is evaluated after the whole local day finishes.
Latest candidate predictions
These are experimental, not the main site forecast. 24-hour rain probabilities are withheld until at least 500 complete, reset-free examples are available.
How it learns and what it cannot claim yet
A small incremental model learns relationships between recent pressure, temperature, humidity, wind, rain history, time of day and season. Each target is revealed only after its forecast period ends. Saved forecasts are never rewritten. Missing or flagged observations do not teach the model.
The historical seed was replayed chronologically, with delayed outcomes. Early tests found useful improvements for short-term temperature, but no consistent improvement across every measurement or horizon. This is not evidence that it beats the current Shadow model. There is no automatic replacement of the main model.
Rain windows begin at the next five-minute boundary after issue. Temperature and wind targets use the archive observation nearest the saved target time. Daily maximum forecasts are issued during the previous day's 23:00 hour. No outside forecast feeds are used.