An end-to-end automated data pipeline to forecast foreign exchange (FX) currency pair movements using regression and classification models. This project leverages real-time market data, macroeconomic indicators, and lightweight AutoML to deliver directional predictions.
This project focuses on predicting FX currency pair movements using:
- Regression models for base pairs (e.g., EUR/USD)
- Classification models for synthetic pairs
- Macroeconomic feature engineering, including BTC correlation
The pipeline is orchestrated using Apache Airflow, stores time-series data with ArcticDB, and uses PyCaret for fast, modular model training.
- Python – Data handling, modeling, automation
- PyCaret – Automated regression and classification
- Polygon API – Real-time FX and BTC market data
- ArcticDB – High-performance time-series data storage
- SQLite – Local relational storage for structured data
- Apache Airflow – DAG scheduling for model runs
- Pandas, NumPy, Seaborn – Data manipulation & visualization
- Keltner Channel indicators
- Rolling window statistics (mean, volatility, momentum)
- BTC/USD correlation as macro proxy
- Lag features and price ratios
- Labeling rule for directional classification
| Model Type | Target | Metric | Score |
|---|---|---|---|
| Regression | Base pairs | R² / MAE | ~0.68 / ~0.02 |
| Classification | Synthetic pairs | F1 Score | ~0.82 |
| Overall | Directional acc. | Accuracy | ~78% |
Tested on holdout set over a 30-day period.
