markuskay/exchange-rate-forecasting

Deep learning models for forecasting Naira-to-Dollar exchange rates and their impact on petrol prices in Nigeria Make sure Public is selected # Deep Learning Models for Forecasting Naira-to-Dollar Exchange Rates # and Their Impact on Petrol Prices in Nigeria A dual-output deep learning system that simultaneously forecasts the USD/NGN exchange rate and predicts the downstream impact on national petrol prices in Nigeria, deployed through an interactive Streamlit web dashboard. ## Overview Nigeria's persistent Naira depreciation from 150 NGN/USD in 2010 to 1,548 NGN/USD by 2025 has driven petrol prices from NGN 65 to NGN 950 per litre, worsening inflation and household hardship. This study develops a fine-tuned dual-output Bidirectional LSTM model with Multi-Head Attention to forecast both the exchange rate and petrol price impact simultaneously using 12 engineered macroeconomic features. ## Model Architecture - Dual-output Bidirectional LSTM with Multi-Head Attention mechanism - Shared encoder feeding two task-specific output heads - Exchange rate prediction feeds directly into the petrol price head enforcing the real-world economic transmission relationship - Noise-based data augmentation expanding training set by factor of 3 - Compared against Transformer model which failed on this dataset scale ## Performance Results | Metric | USD/NGN Forecast | Petrol Price Forecast | |--------|-----------------|----------------------| | R² | 99.96% | 99.94% | | RMSE | 0.005458 | 0.006055 | | MAE | 0.004266 | 0.004880 | | MAPE | 3.74% | 9.17% | ## Key Features - 12 engineered macroeconomic input features including FX Volatility, CBN Forex Reserves, Brent Oil Price, Inflation Rate and lag variables - Real-time dual prediction with threshold-based alert classification - Stable zone, moderate warning and high depreciation alerts - Sensitivity analysis tab showing petrol price response across different Brent oil price scenarios - Interactive Streamlit dashboard with three tabs ## Dataset - Monthly macroeconomic records sp …