pangwanabervely-stack/covid-case-forecasting-zimbabwe

Time-series forecasting of weekly COVID-19 cases in Zimbabwe using ARIMA and Random Forest, with a comparative analysis of model performance and failure modes. # COVID-19 Weekly Case Forecasting Zimbabwe **Author:** Bervely Pangwana **Tools used:** Python (pandas, numpy, matplotlib), statsmodels (ARIMA), scikit-learn (Random Forest) ## Project Overview This project forecasts weekly COVID-19 case counts for Zimbabwe using two different modeling approaches a classic statistical time-series model (ARIMA) and a machine learning approach (Random Forest with engineered lag features) and rigorously compares their performance against each other and a naive baseline. **Note:** This is a technical forecasting exercise using historical public health data for skill demonstration. It is not intended for clinical or policy decision-making. ## Objectives - Build a time-series forecasting pipeline from raw case data to evaluated predictions - Apply and compare a classic statistical method (ARIMA) against a machine learning method (Random Forest with lag features) - Critically evaluate model performance, including understanding *why* a model underperforms, not just reporting whether it did ## Data Source Our World in Data COVID-19 dataset, filtered to Zimbabwe, resampled from daily to weekly case counts to reduce reporting noise. ## Methodology 1. **Data Preparation:** Filtered global dataset to Zimbabwe, cleaned missing/negative values, resampled daily counts to weekly totals. 2. **Train/Test Split:** Held out the final 12 weeks as a test set (never seen during training). 3. **Baseline:** Naive forecast (next week = last known week) as the benchmark to beat. 4. **Model 1 - ARIMA:** Tested for stationarity (Augmented Dickey-Fuller test), selected the best (p,d,q) order via AIC grid search, fit and forecasted. 5. **Model 2 - Random Forest:** Engineered lag features (previous 4 weeks + rolling mean), trained a Random Forest Regressor, and generated forecasts recursively (each prediction feeds into the next week's input features). 6. **Evaluation:** Compared all three approaches using MAE and RMSE. ## Key Findings **Model Performance …