Generated by python/benchmarks/bench_vs_arch.py. Every number below comes from the one machine listed here;
the ratios compare runs on that machine only and say nothing about another one.
The matched-work rows fit the same simulated series (GARCH(1,1) with Student-t(8) innovations, default_rng(7))
with the same specs: constant mean, sGARCH(p,q), gjrGARCH(1,1), eGARCH(1,1) and EWMA (lambda estimated, arch's
EWMAVariance(None)), each with normal, Student-t or GED innovations, plus sGARCH(1,1) with skewed Student-t
innovations (sstd against arch's skewt, the same family with another parameterization). volengine runs with Context(threads=1) and var_init="backcast", the arch convention;
arch runs with rescale=False and fit(disp="off"), one fit after the other. Each timing is the total wall
time of the fits it covers.
- Machine: AMD Ryzen 7 5700X 8-Core Processor, 8 cores, 16 hardware threads, 32 GiB RAM, Windows-10-10.0.19045-SP0
- Versions: volengine 1.0.0, arch 8.0.0, numpy 2.5.4, Python 3.14.7
- Date: 2026-11-17
- Work: 20 series of 5000 observations, 19 specs, 380 fits per library; median wall time of 5 repetitions.
| Run | Wall time (s) | Fits per second | Relative to arch on this machine |
|---|---|---|---|
| arch, sequential (matched work) | 16.80 | 22.6 | 1.00x |
| volengine, threads=1 (matched work) | 3.14 | 120.9 | 5.35x |
| volengine, threads=0 (all cores; not matched work) | 0.59 | 642.1 | 28.39x |
| Model (matched work) | Fits | arch (s) | volengine, threads=1 (s) | Relative to arch on this machine |
|---|---|---|---|---|
| sGARCH | 200 | 9.60 | 1.87 | 5.14x |
| gjrGARCH | 60 | 4.05 | 0.48 | 8.45x |
| eGARCH | 60 | 2.61 | 0.67 | 3.88x |
| EWMA | 60 | 0.53 | 0.12 | 4.31x |
| Spec | Worst relative loglik difference over the series |
|---|---|
| sGARCH(1,1) norm | 5.1e-11 |
| sGARCH(1,1) std | 4.8e-11 |
| sGARCH(1,1) ged | 4.3e-11 |
| sGARCH(2,1) norm | 1.6e-09 |
| sGARCH(2,1) std | 5.2e-09 |
| sGARCH(2,1) ged | 3.5e-09 |
| sGARCH(1,2) norm | 7.5e-09 |
| sGARCH(1,2) std | 3.6e-08 |
| sGARCH(1,2) ged | 1.5e-08 |
| sGARCH(1,1) sstd | 4.2e-11 |
| gjrGARCH(1,1) norm | 5.1e-11 |
| gjrGARCH(1,1) std | 4.9e-11 |
| gjrGARCH(1,1) ged | 3.6e-11 |
| eGARCH(1,1) norm | 1.0e-09 |
| eGARCH(1,1) std | 4.4e-10 |
| eGARCH(1,1) ged | 5.2e-11 |
| EWMA(1,1) norm | 5.1e-11 |
| EWMA(1,1) std | 3.3e-11 |
| EWMA(1,1) ged | 2.3e-11 |
Notes:
- The optimizers differ (L-BFGS in VolEngine, SLSQP in
arch), so the times include different numbers of iterations. The second table shows that both reach the same optimum: the largest relative difference of the log-likelihood over the 20 series is below 1e-4 for every spec (the script fails otherwise). - The
threads=0row is not matched work:archdoes not parallelize a fit, so that row measures the batch parallelism of VolEngine, not the algorithms. Fits are bit-identical for any thread count. - The run was not isolated: a normal desktop session, no pinned cores and no fixed frequency.
- Reproduce with
VOLENGINE_LIB=<path to the library> python python/benchmarks/bench_vs_arch.py --out docs/performance/python-vs-arch.mdin thepython/.venvenvironment.