ofilemfetane10/indabax-botswana-2026-forecasting

Phase 1 submission for the IndabaX Botswana 2026 AI Hackathon. A multi-source forecasting pipeline combining Baltic Dry Index, Brent crude oil prices, Botswana policy rates, FAO food price indicators, and regional inflation data to forecast Botswana food inflation and analyse human capital risk channels. # IndabaX Botswana 2026 Forecasting Challenge ## Overview This repository contains our Phase 1 submission for the IndabaX Botswana 2026 AI Hackathon. The project develops a multi-source forecasting framework to predict Botswana food inflation and explore how global economic shocks may propagate into human capital outcomes. The solution integrates shipping, energy, monetary policy, domestic food prices, and regional inflation indicators into a unified forecasting pipeline. ## Problem Statement Botswana is a small open economy that is highly exposed to global shocks. Changes in shipping costs, oil prices, regional food inflation, and monetary policy can influence domestic food prices and eventually affect household welfare, education outcomes, and health outcomes. The objective of this project is to: 1. Forecast Botswana food inflation. 2. Identify important economic drivers of inflation. 3. Compare classical machine learning and deep learning forecasting approaches. 4. Explore regional inflation spillovers using cross-country HCP indicators. 5. Produce evidence that can support downstream human-capital forecasting. ## Datasets Used This benchmark combines all five challenge datasets: - Baltic Dry Index (shipping costs) - Brent Crude Oil Prices - Botswana Policy Rate - FAO Botswana Food Price Indicators - HCP Cross-Country Food Inflation Indicators ## Methodology ### Feature Engineering The Baltic Dry Index was transformed into monthly indicators capturing: - Monthly mean - Monthly standard deviation - Monthly maximum - Monthly minimum - Monthly range - Last observed monthly value - Intramonth return - Momentum indicators - Volatility indicators Lag structures of: - 1 month - 2 months - 3 months - 6 months - 9 months - 12 months were used to capture delayed transmission effects. ### Forecasting Models Two forecasting approaches were evaluated: #### Classical Machine Learning - Gradient Boosting Regressor - XGBoost-style forecasting pipeline - Rec …