rebiraolin/ethiopia-fx-analysis

# Predictive Modeling of FX Regime Shifts: Ethiopia Case Study *Using Random Forest on scale-invariant features to detect currency devaluation events from sparse, multi-frequency data.* > **🚀 Try the Live Interactive Dashboard →** --- Figure 1: Model Predicted Probability vs. Actual Official Rate during the 2024 FX Liberalization. The model, trained exclusively on 2017–2023 data, shows a probability spike to 0.42 in late June 2024 — weeks before Ethiopia's historic exchange rate reform. --- ## Abstract Ethiopia's FX market presents a fundamental **data scarcity challenge**: official exchange rates are reported daily (~8,000 rows), but critical parallel (black market) rates exist only as monthly observations (~83 rows), creating a 30:1 frequency mismatch that prevents conventional time-series modeling. This project harmonizes the two series using **PCHIP monotonic interpolation** and forward-filling, then engineers 14 **scale-invariant features** — rate-of-change, normalized volatility, SMA deviation, and premium momentum — to enable prediction that transfers across exchange rate regimes. A Random Forest classifier, trained exclusively on the pre-liberalization era (2017–2023, where only 1.7% of days experienced jumps), achieves an **F1-score of 0.535** and **55.2% recall** on the 2024 regime shift, demonstrating that scale-invariant feature engineering can detect structural breaks in managed exchange rate systems before they occur. --- ## Table of Contents - The Data Challenge - Methodology - Phase 1: Data Harmonization - Phase 2: Feature Engineering - Phase 3: Model Training — The Scale-Invariant Pivot - Results - Model Performance - Feature Importance - The "2024 Liberalization" Test - Project Structure - How to Run - Key Takeaways & Limitations - License --- ## The Data Challenge Predicting exchange rate regime shifts in developing economies is uniquely difficult because the data landscape is fragmented, sparse, and structurally inconsiste …