QuantNova.jl

Appearance

Benchmark Methodology ​

This page documents how QuantNova performance numbers are produced and how to reproduce them.

Principles ​

  • Warmup first: all benchmarks run warmup iterations to remove JIT compilation effects.

  • Median reporting: we report the median runtime across repeated runs.

  • Fixed inputs: parameters and random seeds are fixed in each benchmark script.

  • Single-process by default: scripts do not enable explicit multithreading or GPU backends.

Benchmark Suites ​

1) QuantNova Comprehensive Suite (Julia) ​

Script: benchmarks/comparison/comprehensive_benchmark.jl

Timing method: @elapsed in Julia, times converted to microseconds, median reported.

Key parameters (fixed in the script):

  • European Black‑Scholes: S=100, K=100, T=1, r=0.05, σ=0.2 Runs: 10,000 (warmup: 1,000)

  • American binomial: 100 steps Runs: 500 (warmup: 50)

  • SABR implied vol: F=100, K=100, T=1, α=0.2, β=0.5, ρ=-0.3, ν=0.4 Runs: 10,000 (warmup: 1,000)

  • Batch pricing: 1,000 options, K∈[80,120], T∈[0.1,2.0], σ∈[0.1,0.5] Random.seed!(42) Runs: 100 (warmup: 10)

  • Greeks (AD): compute_greeks on a European option Runs: 1,000 (warmup: 100)

  • Monte Carlo: 10,000 paths, 50 steps for European and Asian Runs: 20 (warmup: 2)

  • American LSM: 10,000 paths, 50 steps Runs: 10 (warmup: 2)

  • Backtesting / Factor / Statistics: synthetic data, Random.seed!(42) 5 years of daily data (252 * 5)

Run:

bash
julia --project=. benchmarks/comparison/comprehensive_benchmark.jl

2) QuantLib C++ Comparison (Direct C++) ​

Scripts:

  • benchmarks/comparison/quantlib_benchmark.cpp

  • benchmarks/comparison/quantlib_benchmark_extended.cpp

Timing method: std::chrono in C++, median reported.

Parameters: Same as the Julia benchmarks (e.g., S=100, K=100, T=1, r=0.05, σ=0.2).

Compile & run (from script comments):

bash
clang++ -std=c++17 -O3 -I<HOME>/dev/QuantLib -L<HOME>/dev/QuantLib/build/ql \
  -lQuantLib -o quantlib_benchmark benchmarks/comparison/quantlib_benchmark.cpp

DYLD_LIBRARY_PATH=<HOME>/dev/QuantLib/build/ql ./quantlib_benchmark

3) QuantLib Comparison via PyCall (Julia) ​

Script: benchmarks/comparison/quantlib_comparison.jl

This uses QuantLib’s Python bindings through PyCall. It verifies correctness first and then benchmarks.

Notes:

  • Theta units differ: QuantLib returns per‑day theta; QuantNova reports per‑year. The script converts QuantNova theta to per‑day for a fair comparison.

Run:

bash
julia --project=. benchmarks/comparison/quantlib_comparison.jl

4) Python Baselines (pandas / statsmodels / vectorbt) ​

Script: benchmarks/comparison/python_benchmark.py

Timing method: time.perf_counter, median reported.

Dependencies: numpy, pandas, scipy, statsmodels, optional vectorbt.

Run:

bash
python benchmarks/comparison/python_benchmark.py

Reproducibility Tips ​

  • Run on an idle machine when possible.

  • Report CPU model, OS, and Julia version with your results.

  • If you change JULIA_NUM_THREADS or enable GPU backends, note it explicitly.

Where Results Appear ​

Performance summaries in the README and docs are derived from these scripts and should be updated whenever benchmarks are re‑run.