Benchmarks

Reference timings from pytest-benchmark (Apple Silicon M-class laptop, 10 cores, Python 3.12, Numba 0.61). Re-run on your machine with:

uv pip install -e ".[dev,bench]"
pytest tests/benchmarks --benchmark-only

Single firm

Operation	Mean	Throughput
distance_to_default (scalar)	~18 µs	55 kops/s
prob_of_default (scalar)	~1.2 µs	800 kops/s
fit(naive)	~38 µs	26 kops/s
fit(jmr_iterative)	~76 µs	13 kops/s
fit(vassalou_xing) (snapshot)	~83 µs	12 kops/s
MertonModel.{set,get}_params	~290 ns	3.4 Mops/s

Vectorised math

Operation	Mean	Throughput
distance_to_default on 1 000 000 inputs	2.5 ms	400 Mops/s

Panel calibration (joblib threading, all cores)

Panel size	Method	Mean
1 000 firms	naive	~0.05 s
1 000 firms	jmr_iterative	~0.3 s
10 000 firms	naive	~0.5 s
10 000 firms	jmr_iterative	~3 s
100 000 firms	naive	~5 s
100 000 firms	jmr_iterative	~30 s

Calibration on a 252-day series

Method	Mean
vassalou_xing (series)	~200 ms
duan_mle (no surv. corr.)	~600 ms

Portfolio Monte Carlo

Workload	Mean
Vasicek analytic IRB, 5 000 firms	~4 ms
Basel IRB capital, 5 000 firms vector	~0.5 ms
MC simulate, 5 000 firms × 100 000 sims	~3 s

Backtest metrics on 1 000 000 (PD, default) pairs

Metric	Mean
auc	~50 ms
brier	~10 ms
ks_statistic	~70 ms
roc_curve	~80 ms
calibration_curve	~60 ms
hosmer_lemeshow	~120 ms

These are all well within the v1.0 GA targets. CodSpeed compares each PR against the previous main commit and flags regressions >5%.