Performance

What a forecast costs. Accuracy comparisons live on the benchmarks page; this is the runtime axis, measured on one machine with every model given the same core budget. There are two regimes and they disagree, so both are reported.

Streaming: one new observation arrives

laplace is online. A new observation costs a single state update whose cost does not depend on how much history preceded it. A foundation model carries no state, so it re-runs its whole forward pass over a shifted window every time. This is the regime a live deployment runs in.

Batch: one forecast for a series never seen before

This is the benchmark protocol, and it is laplace at its worst: it must consume the whole history to warm up, while a foundation model amortises a batched forward pass across many series. Best batch size shown per model.

Every measurement

Caveats

• CPU only, 1 torch thread. A GPU changes the foundation-model numbers substantially and does not change laplace's.
• The two regimes disagree, and neither is wrong. Batched, a foundation model can be cheaper per series than laplace replaying a history. Streaming, laplace is orders of magnitude cheaper. Which number applies depends entirely on whether you are scoring a benchmark or running a service.
• Rust and Python differ by roughly 18×. Ratios here are against the Rust core (skaters_fast), which is what ships; the pure-Python reference is correspondingly slower.
• Horizon costs laplace. Going from k=1 to k=12 costs it roughly an order of magnitude, because the multi-scale ensemble runs a full candidate pool per decimation stride per horizon.
• Not size-ordered. The models differ in parameter count and in output mechanism (quantile heads versus sampled paths), so this ranks implementations as configured, not architectures in the abstract.
• Dependencies are not in the table. laplace is pure Python (and JavaScript, and Rust) with no weights to download and runs in a browser; the foundation models need torch and hundreds of megabytes of weights.

Reproducing this

Every number on this page is generated, not typed. The script measures each model and writes both this page and the machine-readable results beside it, so re-running it refreshes the page and nothing here can drift from the data.

PYTHONPATH=src:benchmarks python benchmarks/perf_compare.py

perf_compare.py — the measurement script · perf_results.json — these numbers as data

Knobs, as environment variables: PERF_L context length, PERF_H horizon, PERF_N series count, TORCH_THREADS core budget for the neural models. Each model is warmed up before timing and reported as the median of repeats.

Measured 2026-08-12 on Darwin arm64, Python 3.11.15, skaters 0.16.0+8ddc8d5, context 256 points, horizon 12, 64 series, median of repeats.