End-to-End Exchange Rate Forecasting System (USD/NGN)

Overview

Exchange rate volatility significantly impacts businesses, investors, and individuals in emerging economies like Nigeria.

This project implements a production-oriented machine learning system for forecasting the USD/NGN exchange rate, designed to handle the full ML lifecycle:

• Feature engineering
• Model training and selection
• Monitoring and drift detection
• Automated retraining
• API-based inference

Unlike typical notebook-based projects, this system is built to be robust, extensible, and deployment-ready.

---

Key Features

• 📊 End-to-end ML pipeline (data → training → deployment → monitoring)
• ⚙️ Shared feature pipeline to prevent training-serving skew
• 🤖 Multi-model training (Linear, Random Forest, XGBoost)
• 🏆 Automated model selection based on RMSE
• 📉 Drift detection using statistical tests (KS-test)
• 🔁 Automated retraining pipeline
• 🚀 FastAPI-based prediction service
• 🧾 Prediction logging (SQLite, Postgres-ready)

---

Quick Start

Run Locally

git clone https://github.com/Udeibom/exchange-rate-forecasting.git

cd exchange-rate-forecasting

python -m venv .venv source .venv/bin/activate # Linux/Mac .venv\Scripts\activate # Windows

pip install -r requirements.txt python training/train_all.py

uvicorn api.main:app --reload

Make a Prediction

POST http://127.0.0.1:8000/predict

---

Forecasting Objective

• Target: USD/NGN closing exchange rate
• Task: Regression
• Horizons:
  • 1-day ahead (primary)
  • 7-day ahead (secondary)

---

Evaluation Metrics

• RMSE (primary)
• MAE
• MAPE

Time-series validation is strictly enforced (no data leakage).

---

System Architecture

         ┌────────────────────┐
         │   Raw FX Data       │
         └─────────┬──────────┘
                   ▼
         ┌────────────────────┐
         │ Feature Pipeline   │
         └─────────┬──────────┘
                   ▼
┌────────────────────────────────────┐
│   Model Training (LR, RF, XGB)     │
└──────────────┬─────────────────────┘
               ▼
     ┌────────────────────┐
     │ Best Model Selector│
     └─────────┬──────────┘
               ▼
     ┌────────────────────┐
     │ Model Registry     │
     └─────────┬──────────┘
               ▼
     ┌────────────────────┐
     │ FastAPI Inference  │
     └─────────┬──────────┘
               ▼
     ┌────────────────────┐
     │ Monitoring & Drift │
     └─────────┬──────────┘
               ▼
     ┌────────────────────┐
     │ Automated Retrain  │
     └────────────────────┘

---

Feature Engineering

A shared feature pipeline ensures consistency between training and inference.

Features include:

• Lag features (1, 7, 14, 30 days)
• Rolling mean and standard deviation
• Returns and volatility
• Calendar-based features

The pipeline follows a scikit-learn-style API (fit, transform).

---

Models

• Linear Regression (baseline)
• Random Forest Regressor
• XGBoost Regressor

Models are evaluated using time-series cross-validation, and the best model is automatically selected.

---

Model Lifecycle Management

Model Registry

• Stores trained models and pipelines
• Tracks feature schema
• Enables reproducible inference

Monitoring & Drift Detection

• Rolling error tracking
• KS-test for distribution shift
• Backtesting against naive baseline

Automated Retraining

Triggered when:

• New data is available
• Drift thresholds are exceeded

---

Prediction API

Endpoint

POST /predict

Flow

Request → Feature Pipeline → Model → Prediction → Logging

Predictions are stored in SQLite (Postgres-ready).

---

Project Structure

. ├── api/ │ └── main.py ├── data/ ├── features/ │ ├── build_features.py │ └── pipeline.py ├── training/ │ ├── train_all.py │ ├── retrain.py │ └── scheduler.py ├── evaluation/ │ ├── metrics.py │ ├── drift.py │ ├── backtesting.py │ └── report.py ├── models/ ├── artifacts/ ├── README.md └── requirements.txt

---

Key Engineering Decisions

• Shared feature pipeline to eliminate training-serving skew
• Time-series validation to prevent data leakage
• Automated model selection based on RMSE
• Drift detection (KS-test) for monitoring
• Modular architecture for scalability

---

Future Improvements

• Docker + cloud deployment
• CI/CD pipelines
• Alerting system (Slack / Email)
• Real-time data ingestion
• Shadow model evaluation

---

Author

Caleb Udeibom
Machine Learning Engineer