Benchmarks

Not a leaderboard zoneboost is trying to win — its actual value proposition is exact, zero-approximation attribution (explain()), not necessarily topping accuracy on tabular benchmarks the way gradient boosting often does. This page reports the real gap (or lack of one) rather than assuming it.

Methodology

Each model uses its own library's out-of-the-box defaults — only random_state is set for reproducibility, no tuning that favors either one.
Cross-validation, not a single train/test split, so results reflect genuine variance across splits rather than one split's luck — the same "never grade yourself on your own homework" principle already central to zoneboost's own cross-fitting.
Real data: California Housing (a standard tabular-regression benchmark, randomly subsampled to 3,000 rows with a fixed seed purely so the comparison runs in a couple of minutes — a disclosed speed tradeoff, not a hyperparameter change to either model).
Compared against LightGBM. InterpretML's Explainable Boosting Machine (EBM) — architecturally the closest existing interpretable model, see How It Compares — was deliberately left out of this comparison: its default outer_bags=8 spawns a separate parallel worker process per bag per CV fold, and the fixed process-spawn overhead in the environment this was run in dominated wall-clock time even on small datasets, independent of any real compute cost. That's an environment/parallelism-backend artifact, not a finding about EBM's accuracy, so it isn't reported here as a misleading number.

Regression: California Housing

3,000-row random subsample, 3-fold cross-validation.

Model	RMSE	R²	Fit time (s)	Predict time (s)
ZoneBoostRegressor	0.5510 ± 0.0212	0.7677 ± 0.0107	5.26	0.267
LGBMRegressor	0.5221 ± 0.0275	0.7912 ± 0.0166	1.11	0.003

LightGBM's out-of-the-box RMSE is about 5% lower and fits roughly 5x faster. Neither model was tuned. zoneboost's own value here isn't matching or beating LightGBM's raw accuracy — it's that every one of those predictions decomposes exactly into main effects and named interaction terms via explain(), with no sampling and no approximation, which LightGBM has no built-in equivalent for.

Reproducing

pip install -e ".[benchmark]"
python benchmarks/compare_lightgbm.py

The script also includes a classification comparison (Breast Cancer Wisconsin) and a synthetic-data demonstration of ZoneBoostRegressor.explain() correctly attributing a known interaction that LightGBM has no way to name — see benchmarks/ in the repository for the full, runnable script and its documented scope.