Inflation Nowcasting
High-frequency ML models for real-time inflation estimation in Sub-Saharan Africa
Inflation Nowcasting applies machine learning to high-frequency, non-traditional data sources to produce real-time estimates of inflation — bridging the gap between monthly official statistics and the rapidly changing economic reality on the ground.
Motivation
In many African economies, official Consumer Price Index (CPI) data is released monthly with a 3–6 week lag. For central banks, businesses, and policymakers, this creates a blind spot in fast-moving inflationary environments (e.g., post-pandemic supply shocks, currency devaluations). Nowcasting — predicting current conditions using leading indicators — addresses this gap.
Approach
- Data sources: Commodity spot prices, currency exchange rates, satellite-derived agricultural productivity indices, web-scraped supermarket price panels, mobile money transaction aggregates
- Models: XGBoost, LightGBM, and LSTM ensembles benchmarked against ARIMA and naïve baselines
- Evaluation: Rolling-window backtesting across Nigeria, Kenya, Rwanda, and Ghana CPI series (2018–2024)
Key Findings
- Gradient-boosted models trained on high-frequency data reduce mean absolute error by 31–47% versus AR(12) baselines depending on the country
- Satellite-derived crop indices are the single highest-importance feature for Nigeria and Kenya; mobile money flows dominate for Rwanda
- Weekly granularity is achievable with the proposed pipeline; daily is feasible for countries with richer web-scraped coverage