What the forecaster learned
A 24-tree random forest predicts the peak load of the next minute. How well it does, what it looks at, and what happens without a clock.
Accurate where traffic has a shape
Forecast against what happened, on 15 traces the model never saw. Typical error is the mean absolute error in requests per second.
| Pattern | Typical error | Fit (R²) | Within 20% of actual |
|---|
Forecast against what happened
For every ten seconds, the forecast of the next minute's peak against the peak that arrived. Where the lines split, the forecaster was wrong.
Mostly it watches the rate now
Share of the forest's decisions each signal accounts for. The position in the cycle matters less overall, but it is what lets the model expect a scheduled sale.
What the clock is worth
Requests slower than 300 ms. The same forest retrained with the position-in-cycle features switched off, run on the same unseen traces.
| Pattern | Reactive | Forest, no clock | Forest, with clock |
|---|
With no clock the forest can only extrapolate the trend, so it catches a ramp but not a sale that starts every hour at the same minute. Trained on synthetic traffic: 24 trees, depth 9, at least 8 samples per leaf. Train it on your own Prometheus data before trusting it.