🇬🇧 English · 🇪🇸 Español
Computational Performance Benchmarks
puremacro is engineered for ultra-fast, zero-overhead execution using vectorized NumPy and SciPy routines.
Runtime Benchmarks (Apple Silicon M-Series / Linux x86_64)
The table below reports execution times across representative macroeconomic models and econometric estimators:
| Engine | Runtime (ms) | Target Reference | Implementation Details |
|---|---|---|---|
| VAR(4) + Cholesky IRF (\(T=1000, N=5, H=40\)) | 1.07 ms | MATLAB / statsmodels | Pure NumPy vectorized MA recursion |
| Sequence-Space HANK GE Solve (\(T=40, N_a=50\)) | 111.8 ms | Julia SSJ / Python SSJ | EGM + \(O(T^3)\) direct matrix inversion |
| DICE Climate Simulation (\(150\text{ yr}, 3\text{-box}\)) | 0.89 ms | GAMS / DICE2016 | Vectorized reservoir transition cycle |
| Mixed-Frequency DFM Nowcast (\(T=120, N=15\)) | 4.84 ms | NY Fed / MATLAB | EM SVD missing-value convergence |
| Elastic Net Coordinate Descent (\(T=200, P=40, 30\lambda\)) | 39.08 ms | glmnet / scikit-learn | Pure NumPy soft-thresholding loop |
| Bandpass (CF) + Beveridge-Nelson (\(T=500\)) | 3.17 ms | MATLAB / Stata tsfilter |
Linear-time state space projection |
The timings above are the reference CPU NumPy/SciPy code paths. The
hardware-accelerated alternatives are documented elsewhere: the backend=
keyword of the spatial solver in
docs/spatial_and_trade_ge.md and the continuous-time
HJB solver in docs/vfi_hjb_continuous.md.
Reproducing Benchmarks
Run the automated benchmark suite locally: