feature-engine

Derive per-symbol feature vectors from streaming NBBO events with deterministic replay.

Updated Mar 2, 2026

Search Tags:#reproducibility #incremental #backtest #feature-engineering #nbbo #per-symbol #stateful-processing

Ready to run

npx skills add https://github.com/Leiisawesome/feelies --skill feature-engine

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: feature-engine
Source: https://github.com/Leiisawesome/feelies/tree/main/.cursor/skills/feature-engine
Command: npx skills add https://github.com/Leiisawesome/feelies --skill feature-engine

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This Skill provides a deterministic, stateful feature engine that incrementally updates per-symbol feature state from L1 NBBO events, reducing recomputation and ensuring reproducible results across warm-up, cold-start, and recovery scenarios.

Core Features & Use Cases

Incremental per-symbol state management with deterministic replay guarantees
Support for rolling window, count-based, and EWMA style features with bounded memory
Clear per-feature warm-up, staleness detection, and a strict feature–signal contract to enable reliable signal generation

Quick Start

Run the feature engine on a stream of NBBO quotes and trades to observe per-symbol feature vectors being emitted as state evolves.

Frequently Asked Questions about feature-engine

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I compute incremental per-symbol features from NBBO market data?▼

To compute incremental per-symbol features from NBBO market data, use a stateful feature engine that processes streaming quotes and trades to emit evolving feature vectors with deterministic replay guarantees.

What is per-symbol state isolation in feature engineering?▼

Per-symbol state isolation in feature engineering ensures each symbol maintains independent lifecycle phases—Init, Warm-up, Active, Stale, and Reset—preventing cross-contamination during stateful processing of concurrent market data streams.

How does deterministic replay guarantee reproducibility for streaming market data features?▼

Deterministic replay guarantees reproducibility by applying versioned provenance and a checkpoint/restore contract via a FeatureSnapshotStore, ensuring identical feature vectors are generated across cold-start, warm-up, and recovery scenarios.

Can I manage bounded memory usage when calculating rolling window and EWMA features?▼

Yes, you can manage bounded memory usage for rolling window, count-based, and EWMA features by enforcing strict per-symbol memory limits and staleness detection during incremental state updates.

Does feature engineering support a strict boundary between features and trading signals?▼

Yes, feature engineering supports a strict feature–signal boundary, isolating deterministic state management and feature vector emission from downstream signal generation logic to enable reliable quantitative analysis.

When should I not use incremental state management for market data backtesting?▼

You should not use incremental state management for market data backtesting if your workflow requires unbounded historical windowing or lacks a checkpoint/restore contract to handle state staleness and recovery.