Benchmarks

Looking for the indicator library's numbers?

This page is about Wickra Gym. Wickra's own indicator benchmarks — the comparison against TA-Lib, talipp, pandas-ta and the other Rust TA crates — live at wickra.org.

Micro-benchmarks for wickra-gym-core, measured with criterion via the gym-bench crate:

cargo bench -p gym-bench                         # parallel (rayon) tensor precompute
cargo bench -p gym-bench --no-default-features   # sequential precompute

The benches cover the two hot paths: the one-time feature-tensor precompute (walk the dataset, compute every observation feature per bar, pack into the fixed tensor) and the rollout (reset + N step calls — the O(1)-per-step index into that tensor plus the fill/PnL reward).

Results

Indicative single-machine numbers (parallel precompute; criterion median of 100 samples). They measure rollout and precompute throughput, not a cross-framework speed comparison — the product's value is deterministic, O(1) rollouts, not raw speed.

Rollout (reset + 100 × step)

Observation dim	median (100 steps)	Throughput
5	~43.1 µs	~2.3M steps/s
20	~42.6 µs	~2.3M steps/s
50	~43.3 µs	~2.3M steps/s

Step throughput is flat across the observation dimension — that is the whole point of the fixed feature tensor: step() is a constant-time index that does not depend on how many features the observation carries.

Feature-tensor precompute (build_tensor)

Dataset size	median	Throughput
1,000 bars	~20.6 µs	~48M bars/s
10,000 bars	~204 µs	~49M bars/s
100,000 bars	~2.58 ms	~39M bars/s

Precompute is roughly linear in the number of bars — the one-time O(1)-per-bar pass that turns every later step() into a pure index.

Method

• Machine and OS vary; treat the absolute numbers as indicative and re-run locally for your hardware.
• The tensor is byte-identical between the parallel (rayon) and sequential (--no-default-features) precompute, so the feature flag changes scheduling, not results.
• The nightly bench.yml workflow re-runs this on a schedule and uploads the report as a CI artifact.

The numbers above are the ones in the repository's BENCHMARKS.md, measured with the commands it names.