jameskingsley/nigeria-inflation-2026

# Nigeria (NG) Inflation Forecasting System (MLOps 2026) An end-to-end MLOps ecosystem designed to forecast Nigerian inflation rates. This project goes beyond simple modeling by implementing an automated "Model Shootout" pipeline that retrains, evaluates, and hot-swaps the best-performing model into production every month. ## System Architecture The project is built on a distributed microservices architecture: * Data Layer: Managed via ClearML Datasets, versioning historical inflation data from the National Bureau of Statistics (NBS). * Orchestration Layer: Prefect 3.0 schedules a monthly training flow. It handles retries, caching, and infrastructure-as-code deployments. * The "Shootout" Logic: Every month, three architectures compete on the latest data: * ARIMA (Statistical): Captures linear auto-regressive trends. * Prophet (Additive): Handles yearly seasonality and economic shocks. * XGBoost (ML): Captures complex non-linear relationships. * Model Registry: ClearML tracks every experiment. The winner (lowest MAE) is auto-tagged as production. * Inference Layer: FastAPI (hosted on Render) pulls the production model and provides a /predict endpoint. * Presentation Layer: A Streamlit dashboard provides real-time visualization for stakeholders. ## Project Structure #### Deployment & Usage ###### Research & EDA View the notebook/eda.ipynb to see the Augmented Dickey-Fuller (ADF) tests, stationarity transformations, and initial model benchmarking. ###### Running the Pipeline To start the automated worker that polls for retraining tasks: ###### API Inference The API supports hot-swapping. When the pipeline finds a new winner, it calls the /refresh endpoint to update the API without downtime. * Endpoint: GET /predict?months=12 * Endpoint: POST /refresh (Internal use) ###### Model Performance The current production model is ARIMA, achieving a Mean Absolute Error (MAE) of 12.85. The system is designed to pivot to Prophet or XGBoost automatically if econom …