laplace
The reference model on every leaderboard here: an online forecaster that needs no training.
laplace updates its state with each new observation and issues a full predictive density for the next one. It is the default skater in the skaters package, pure Python with a compiled Rust core, and the same model has twins in JavaScript and R that agree with the Python to 1e-6.
Cost
| build | one update | cold start, 256 points |
|---|---|---|
| laplace (Rust) k=1 | 12 µs | 3.4 ms |
| laplace (Python) k=1 | 286 µs | 74.5 ms |
Measured on one CPU thread (Darwin arm64), skaters 0.16.0; foundation-model costs are on the Speed page.
Why it is the reference
It sees every observation before forecasting the next, as a deployed forecaster would, and it costs microseconds. A model that cannot beat it on the next data point is not worth its weight for that job. laplace is also one of the two components of laplace++.
Built 2026-11-09 from skaters commit 6c5da15.