Practical Guide to Feature Engineering for FX: Price, Order‑Book, Sentiment & Macro Inputs

Introduction — Why feature engineering still wins in FX

Machine learning (ML) models for foreign‑exchange (FX) trading succeed or fail largely based on the features fed to them. Good features encode the market microstructure, persistent statistical patterns and the external drivers that move currency prices; poor features amplify noise, overfitting and look‑ahead bias. This guide presents pragmatic, implementation‑focused feature ideas across four families: price/time‑series, order‑book (microstructure), sentiment (text & social), and macroeconomic inputs — and it explains practical engineering, selection, and monitoring rules for each.

What you'll get: concrete feature definitions, data and latency tradeoffs, normalization and leakage checks, and a short production checklist so you can move from ideas to backtests to live execution.

Note: many high‑frequency and microstructure findings are supported by academic and practitioner work on limit order books and order‑flow imbalance; where possible we reference modern sources and engineering patterns for reproducibility.

1. Price & time‑series features — the foundation

Price features remain the backbone of FX models at all horizons. Useful, well‑tested components include:

• Normalized returns and lags: log returns at multiple horizons (e.g., 1s, 1m, 5m, 1h, daily) and lagged return columns to capture momentum and mean reversion.
• Rolling statistics: rolling mean, standard deviation, skewness, kurtosis, and z‑scores computed over multiple window sizes; use exponential windows for faster adaptation.
• Volatility features: realized volatility, Parkinson/Garman–Klass estimators, rolling RV and vol‑of‑vol; useful both for sizing and as inputs to regime detectors.
• Trend & cycle filters: moving average crossovers, Hilbert transforms, and detrended price oscillators for regime tagging.
• Volume‑aware metrics: signed volume imbalances, VWAP distance, and volume‑weighted returns when trade ticks are available.

Engineering notes:

• Always compute returns on a uniform grid (time buckets or event buckets) and align all features to the same target horizon to avoid leakage. The Feature Engineering Bookcamp and practical time‑series guides discuss these patterns in detail.
• Standardize/robust‑scale each feature per asset and per regime window to keep model weights stable across pairs.
• Monitor feature decay—momentum and microstructure signals often have short half‑lives and require more frequent recalibration.

2. Order‑book & microstructure features — extracting short‑horizon signal

If you have depth data (top‑of‑book or multi‑level), the order book is a rich source of predictive features for short horizons. Core features and design patterns:

• Best‑bid/ask spread & mid: instantaneous spread and midprice; also compute microprice (imbalance‑aware mid) to capture queue pressure.
• Order‑flow imbalance (OFI / MLOFI): net changes to bid and ask sizes across levels (L1–L10). Multi‑level order‑flow imbalance improves short‑horizon mid‑price forecasting in empirical studies.
• Depth ratios & weighted depth: weighted sum of quantities across levels (e.g., depth ratio or depth imbalance), often normalized by typical depth percentiles for the symbol. Practical implementations use a top‑N weighted scheme where nearer levels get exponentially higher weight.
• Trade sign and impact curves: signed trade flow (buyer vs seller initiated) and measured immediate impact (future mid move vs size bucket) are powerful features for execution and alpha modeling.

Practical cautions:

• Order‑book features require careful timestamp syncing and robust null‑handling—missing L2 data or crossed books must be filtered.
• Microstructure features are latency‑sensitive: compute them where data arrives (edge/VPS) if you intend to trade on sub‑minute signals.
• Alpha decays quickly; incorporate rolling recalibration and treat these features as high‑frequency signals with short retrain cycles. Practitioner write‑ups and open toolkits show reproducible pipelines to compute spread, imbalance and microstructure analytics.

3. Sentiment & macro inputs — leading context and regime indicators

Sentiment and macro data provide longer‑horizon context and are essential when modeling carry, directional views and regime shifts.

Sentiment features (news & social)

• News sentiment scores: article‑level polarity (−1 to +1), surprise tags (hawkish/dovish), and entity tagging (e.g., central bank, inflation) from providers or in‑house NLP pipelines. Aggregate at minute/hour/day as appropriate.
• Event flags & embeddings: convert headlines and speeches into event flags (rate‑cut, rate‑hike, QE, sanction) and semantic embeddings (sentence transformers) to measure similarity to historical episodes.
• Social momentum: volume and velocity of social mentions, attention spikes, and network‑level signals (retweet/reshare patterns); useful for crypto‑linked FX flows and risk‑on/off transitions.
• Recent practitioner reviews show AI‑driven real‑time sentiment pipelines and the growth of predictive sentiment tools in 2024–2025; these systems combine multi‑source feeds for earlier detection of tone shifts.

Macro features

• Real rates & yield spreads: short and long rate differentials (e.g., 2y–10y and cross‑currency short rates) and break‑even inflation are core drivers for FX trends.
• Surprise indices: the difference between actual releases (CPI, NFP) and market consensus; create rolling surprise aggregates to detect persistent cycles.
• Liquidity & positioning: COT extremes, repo rates and interbank liquidity proxies provide institutional positioning context that often precedes sustained FX moves.

Engineering and model‑risk notes

Modern advances allow language models to assist feature generation (e.g., entity extraction, event summarization and candidate feature proposals), but these systems need careful validation to avoid hallucinations and spurious correlations. Recent academic work explores LLM‑assisted feature generation and multimodal financial foundation models — promising but requiring strong interpretability and governance in production.

Practical pipeline checklist

1. Define prediction horizon (H) and align all features to the H target; test for leakage with strict backtest time‑blocking.
2. Implement robust preprocessing: outlier trimming, winsorization, and per‑symbol robust scaling.
3. Feature selection: use information coefficient (IC), feature permutation importance and stability tests across regimes instead of naive forward‑step selection.
4. Monitor feature drift and schedule retraining thresholds (IC decay, validation loss, population drift metrics).
5. Document data lineage and embed explainability (SHAP/Integrated Gradients, surrogate linear models) for human review and regulatory traceability.

Conclusion — balance simplicity and novelty: start with proven price and volatility features, add microstructure inputs if you control low‑latency data, and use sentiment/macro features to inform regime or allocation overlays. Use LLMs and multimodal tools as accelerants for feature discovery, but keep human‑in‑the‑loop validation and strong monitoring to avoid overfitting and hallucinated features.