Benchmarks

The benchmark below compares the Python-facing ferric-alpha API with alphalens on the same deterministic long/short factor fixture. It measures function execution time only; fixture construction is outside the timed block.

Setup

• Machine: Apple silicon macOS 26.3.1
• Dataset: 250 assets, 252 sessions, 63,000 factor rows
• Forward returns: 1D and 5D
• Iterations: 5; result is median wall-clock time
• Rank autocorrelation recheck: median of 7 runs, 15 iterations per run
• ferric-alpha: 0.1.0, release build, Python 3.13.4, Polars 1.42.1
• alphalens: 0.4.0, Python 3.9.25, Pandas 1.5.3, NumPy 1.23.5

Results

• factor_information_coefficient
  • ferric-alpha: 13.616 ms
  • alphalens: 124.205 ms
  • Speedup: 9.1x
• factor_weights
  • ferric-alpha: 15.153 ms
  • alphalens: 43.330 ms
  • Speedup: 2.9x
• factor_returns
  • ferric-alpha: 11.778 ms
  • alphalens: 46.425 ms
  • Speedup: 3.9x
• mean_return_by_quantile
  • ferric-alpha: 21.022 ms
  • alphalens: 60.900 ms
  • Speedup: 2.9x
• quantile_turnover
  • ferric-alpha: 3.018 ms
  • alphalens: 6.136 ms
  • Speedup: 2.0x
• factor_rank_autocorrelation
  • ferric-alpha: 6.823 ms median (6.686-6.920 ms observed)
  • alphalens: 15.497 ms median (15.326-15.864 ms observed)
  • Speedup: 2.27x

Benchmark results are hardware and environment dependent.

Rank Autocorrelation Fast Path

Complete date-by-asset panels use a dense matrix internally. Assets are encoded once, each date occupies one contiguous matrix row, and lagged rows are correlated directly. This avoids per-date string maps while preserving the public Python and Rust APIs.

Sparse panels, changing universes, null factors, and non-finite factors use the general compatibility path. The optimization therefore changes execution, not the metric definition.

The optimized path was checked directly against alphalens 0.4.0 using a shuffled 80-session by 37-asset panel containing tied factor values. For lags 1, 3, and 10, null positions matched exactly and the maximum absolute difference was 1.11e-16 under a 1e-12 comparison tolerance.

Reproduce

.venv/bin/maturin develop --release
.venv/bin/python scripts/benchmark_alphalens_comparison.py --engine ferric

python3.9 -m venv /tmp/alphalens-bench
/tmp/alphalens-bench/bin/python -m pip install \
  'alphalens==0.4.0' 'pandas==1.5.3' 'numpy==1.23.5' \
  'scipy==1.10.1' 'statsmodels==0.13.5' 'empyrical==0.5.5'
PYTHONWARNINGS=ignore /tmp/alphalens-bench/bin/python \
  scripts/benchmark_alphalens_comparison.py --engine alphalens