Predictive Analytics for Sales Forecasting in Emerging Market Retail: An Ensemble Learning and Time-Series Approach for A&Z Supermarket

Abstract

Evaluation of forecasting models with 5000 daily sales records (2020-2023) from A&Z Supermarket located at Oyo State, Nigeria. Ten (10) product IDs from five (5) stores. The dataset incorporates a total of 11 predictive variables. This includes price, competitor price, macroeconomic index (simulated), promo indicators, holiday flags and rolling averages. The forecast models are assessed through a train/test split with a 70/30 ratio. They are forecasting for a 7-day horizon. Further, rolling window validation prevents leakage. According to descriptive analysis, the average sales quantity resulted in 26.49 sales units with a standard deviation of 11.58. The day of sales displays cumulative spikes, mainly during weekends as well as holidays.  The research compares various traditional models such as ARIMA, SARIMA, and Linear Regression with machine learning models like Random Forest, XGBoost, LSTM, and GRU and has a RMSE ranging from 10.55 to 14.36 and R² values ranging from %u221A0.740 to 0.884. Bagging, boosting, stacking techniques are used to improve the performance of predictive models. The stacking ensemble achieved the lowest RMSE (8.14). The performance of LSTM was observed to be best as depicted by (RMSE= 8.75, R²= 0.91) as it outperformed Prophet and XGBoost. The lack of generalizability can be attributed to a narrow dataset that incorporates only a single retailer and a simulated macroeconomic index. An analysis of the business KPIs indicates a decline in inventory costs of 17.8% and of stockouts by 15.4%. Such improvements along with a 9.5% increase in ROI are encouraged through the use of LSTM (Long-Short Term Memory) investments. The results indicate that hybrid ensemble-time-series models are superior to conventional methods and offer important practical conveniences in optimizing inventory and forecasting demand in retail settings in emerging markets.