Benchmarks

Wickra is a streaming-first library: the state machine inside every indicator can take a single new tick and return its updated output in constant time. The numbers below show what that costs in wall-clock time relative to the most-used Python TA libraries.

Reproduce these on your own hardware

pip install -e bindings/python[bench]
python -m benchmarks.compare_libraries

The script auto-detects every peer library installed in your venv and runs them against the same generated input as Wickra. The CI job cross-library-bench uploads the raw report as an artefact on every push.

Batch — single full pass over 20 000 bars

Lower bar = faster. The orange bar is Wickra, the grey bar is the peer.

vs. finta

SMA(20)

95.6µs

343.5 µs

3.6× faster

EMA(20)

64.6µs

223.1 µs

3.5× faster

RSI(14)

126.2µs

1107.1 µs

8.8× faster

MACD(12,26,9)

119.0µs

531.8 µs

4.5× faster

Bollinger(20,2.0)

105.3µs

812.0 µs

7.7× faster

ATR(14)

123.5µs

5144.8 µs

41.7× faster

vs. talipp

SMA(20)

95.6µs

7640.6 µs

79.9× faster

EMA(20)

64.6µs

12160.9 µs

188.2× faster

RSI(14)

126.2µs

15792.2 µs

125.1× faster

MACD(12,26,9)

119.0µs

49788.1 µs

418.4× faster

Bollinger(20,2.0)

105.3µs

130938.3 µs

1243.5× faster

ATR(14)

123.5µs

28816.0 µs

233.3× faster

Streaming — per-tick latency after 5 000 historical bars

This is where streaming-first really shows. A batch-only library has to re-run its full indicator over the entire history on every new tick; Wickra advances its internal state by one step.

RSI(14)

0.1µs / tick

1.6 µs / tick

13.8× faster

Why isn't TA-Lib / pandas-ta on these charts?

Both fail to install cleanly on Windows without C build tooling — which is precisely the install-pain Wickra was built to remove. The bench script auto-detects every peer library it can find and runs them on the same inputs as Wickra; install them in your environment and those rows will show up too.

What the numbers do not say

• Absolute µs values depend on CPU, memory clock, OS scheduler, and the Python / Node / Rust versions — read them as relative speedups between libraries on identical input, not as a universal performance contract.
• Reproduced on: Windows 11 Pro 26200, AMD Ryzen 9 9950X, 64 GB DDR5, Rust 1.92 (release profile, lto = "fat", codegen-units = 1), Python 3.12, Node 20.
• The Wickra figures are the Python binding runtime, not the bare Rust kernel — there is a small PyO3 boundary cost included on each measurement.

See also

• benchmarks/compare_libraries.py — the canonical script the table above runs.
• Bench workflow — nightly run on the GitHub-hosted Linux runner, archived as build artefacts.
• Streaming-vs-Batch (docs) — what the equivalence guarantee actually means.