Skip to content

Latest commit

 

History

History
86 lines (65 loc) · 3.36 KB

BENCHMARKS.md

File metadata and controls

86 lines (65 loc) · 3.36 KB

Benchmarks

benches/allocate.rs times allocate with criterion for the methods that run a solver or a clustering step:

Method What it measures
MinVariance The active-set QP solver, long-only
MinVarianceGroups The same with two overlapping groups that bind: the first half of the assets at most 30% together, and every third asset at least 40%
MinCvar The two-phase simplex LP
Hrp Single-linkage clustering and recursive bisection
Herc Single-linkage clustering, the gap statistic and equal risk contribution across clusters
RiskParity Newton's method on ½yᵀΣy − Σᵢ bᵢ ln yᵢ
cargo bench
cargo bench -- MinVariance/100

Data

The returns are generated from a fixed seed, so every run sees the same inputs:

  • three common factors, each asset with its own exposures between 0.5 and 1.5;
  • a daily drift between 0.02% and 0.08%;
  • noise with a standard deviation between 0.5% and 2%;
  • two periods per asset, which keeps the sample covariance well conditioned.

Each timing covers the whole of allocate, including the sample covariance.

Baseline

Measured at 0.3.1 on an Apple M5 with Rust 1.92. Each time is criterion's point estimate.

Method 10 assets 50 assets 100 assets 500 assets
MinVariance 28 µs 35 ms 18 s
MinCvar 54 µs 24 ms 0.96 s Not run
Hrp 1.5 µs 62 µs 1.4 ms
Herc 2.2 µs 0.19 ms 4.6 ms
RiskParity 2.6 µs 0.24 ms 20 ms
  • MinCvar: at 0.3.1 it was benchmarked at 10, 50 and 100 assets. At 200, its LP's phase 1 stalled on a degenerate vertex: every pivot left x where it was, until the iteration cap returned NotConverged after about 75 seconds.
  • MinVariance: a solve takes about 500 times longer from 100 to 500 assets, where the size alone would account for about 125 times.

Changes since the baseline

MinCvar's starting point

The LP used to start every scenario weight at 0, which left its asset rows without a residual: the degenerate start above. It now starts the scenarios where the equal-weight portfolio loses most at their cap.

MinCvar 10 assets 100 assets 500 assets
0.3.1 54 µs 0.96 s Not run
Now 47 µs 28 ms 11 s

Groups

Measured on 2026-11-12, on the same machine, when MinVarianceGroups was added. MinVariance is the same benchmark as above, after the solver work of 0.3.2 to 0.3.7.

Method 10 assets 100 assets 500 assets
MinVariance 12 µs 2.4 ms 0.23 s
MinVarianceGroups 20 µs 0.32 ms 7.1 ms

From 100 assets it's faster than without groups: the binding groups leave the solver fewer free weights to factor. A group that doesn't bind costs a slack variable, which the solver leaves out of its factored solves.

Running time

  • cargo bench: about seven minutes, most of it MinVariance and MinCvar at 500 assets. Those two sizes, where a solve takes over a tenth of a second, get criterion's minimum of 10 samples rather than its default of 100.
  • cargo test --all-targets: runs each benchmark once, unoptimized, at the smallest size only, so CI checks that the benchmarks build and run without timing them. Timings on shared CI runners are too noisy to compare.