Benchmarks

These are the numbers we hold ourselves to. Every one is reproducible from the harness in benchmarks/; lower is better for error metrics unless noted. This is pre-1.0 software, so treat them as a spike baseline, not a leaderboard claim.

Forecasting (MASE, lower is better)

Standard synthetic suites (trend + seasonality + noise), scored against a seasonal-naive floor of 1.0.

Model	MASE	Notes
seasonal-naive	1.11	the floor
theta-classic	1.03	bundled statistical model, zero dependencies
distilled Chronos student	0.89	native blob, no runtime at serve time
Chronos (teacher, ONNX)	~0.80	for reference; needs the ONNX build

The point of the middle two rows: distilling the foundation model into a native student recovers most of the teacher’s accuracy (0.89 vs 0.80) while serving in the zero-dependency core. See Distillation.

Anomaly detection (VUS-PR, higher is better)

Univariate track of the TSB-AD benchmark.

Detector	median VUS-PR
sub-pca (bundled)	~0.44

sub-pca (subsequence PCA, Jacobi eigensolver in C) lands at or above the published SOTA band on the univariate suite, from a dependency-free build.

Tabular prediction

On TabArena-style tasks, a distilled gbt student matches or beats a tuned XGBoost baseline, while running inside SQLite with no external library. Full per-task tables live in the harness.

Interval calibration

On a smooth synthetic series at a nominal 90% level:

Interval method	empirical coverage
residual (Gaussian)	0.57
conformal	0.90

The default Gaussian band is overconfident on smooth data; the conformal band hits nominal coverage. You can reproduce this on your own data with backtest(). See Auto-selection & conformal intervals.

Reproducing

git clone https://github.com/PureStorage-OpenConnect/sqlite-predict
cd sqlite-predict && make loadable
cd benchmarks && uv run python run.py

Results are written to benchmarks/results/. If your numbers differ, open an issue with your platform and the generated table.