About Wickra Feature Store

Wickra Feature Store folds OHLCV and microstructure event streams into ML-ready feature/label matrices over the 514 streaming indicators of the Wickra core. A build is a JSON document — data, not code — so the exact same feature build runs in every one of ten languages and returns byte-for-byte identical results.

What makes it different

• The build is data. A serde FeatureSpec — a universe, an ordered list of indicator / price / microstructure feature columns, and forward-looking label columns. Because it is data, it crosses the C ABI and WASM unchanged.
• 514 indicators. Any feature column can reference any Wickra indicator (Rsi, Ema, Macd, BollingerBands, …), emitted one row per bar.
• Forward-looking labels. Forward returns over a horizon and triple-barrier labels are computed from future bars and aligned to the row that predicts them.
• Parallel and deterministic. The universe is folded in parallel with rayon, or sequentially on the WASM path — both produce a byte-identical matrix, pinned by a golden corpus in CI. Optional z-score / min-max column scaling is part of the spec.

Why it exists

Feature engineering for market data is usually one language, one hard-coded pipeline. Wickra Feature Store defines the feature surface once, in Rust, and exposes it as a JSON-over-C-ABI data API to Rust, Python, Node.js, WASM and — over a C ABI — C, C++, C#, Go, Java and R. The spec itself is portable JSON, so the same pipeline runs anywhere, and columnar output (CSV, Arrow, Parquet) drops straight into a training pipeline.

Open source

Released under the MIT OR Apache-2.0 license — permissive, OSI-approved, free for any use including commercial. Source, issues and releases on GitHub.

Disclaimer

Wickra Feature Store is a software library, not a trading system, and is provided as-is with no warranty. It transforms market data into features; it does not give financial advice. Use it at your own risk.