A hybrid model for predictive decision-making in enterprise data analytics

Abstract As competition in the business world is getting tougher, many businesses are confronted with the need to effectively manage increasingly large-scale datasets for predictive decision making for enterprise systems. With the increased sophistication in data variation, collection and analysis due to the emergence of new technologies, there is need for a hybrid predictive decision-making model that will handle the massive amounts of large-scale data from enterprise information systems and applications across various industries to guarantee better decision-making. This research work is aimed at developing a hybrid predictive decision-making model for enterprise data analytics, leveraging on the synergies of Prophet, ARIMA, and Random Forest to improve the prediction accuracy of seasonal and non-linear data in enterprise systems. Decision-making processes could be enhanced via the provision of more reliable forecasts. Obviously, single model predictive approaches lack the robustness and flexibility needed to capture patterns accurately; hence it has become crucial to implement a hybrid model that will leverage the synergies of different forecasting techniques to provide accurate and actionable insights, as well as empower organizations to make informed decisions based on comprehensive data analytics. For this study, a historical supply chain dataset of a retail outlet was used for data analysis. The data was split using the train_test_split while RandomizedSearchCV was used for hyperparameter tuning. Also, the Apache Spark open-source technology was adopted to assess the feasibility of our model for real-time enterprise application, with the Star schema adopted to organize data into measurable facts and descriptive dimensions to simplify data analysis and optimized for SQL data querying. To capture seasonality and linear components of data, the Prophet and ARIMA time series models were used while Random Forest was used for non-linear relationships. The models were trained and evaluated individually after which they were combined and blended using an ensemble approach of a weighted mean average. Evaluation of the hybrid predictive decision-making model was made by comparing the model’s predictions against actual values using RMSE, MAE, MAPE and R 2 evaluation metrics. The result revealed that combining the forecasting models gave a lower error RMSE value of 4.83, MAE value of 3.32, lower MAPE of 1.92% and, R 2 of 99.2% as against the individual model evaluation result of Prophet (high error RMSE value of 43991.04, MAE value of 38754.39, MAPE of 27.4%, and R 2 of − 6.5); ARIMA (RMSE value of 68.35, MAE value of 60.36, MAPE of 49.0%, and R 2 of − 0.54), Random Forest (RMSE value of 3.66, MAE value of 1.90, MAPE of 1.54, and R 2 of 0.99). Our hybrid model was also compared with other state-of-the-art ensemble models which equally showed improve sales forecasting performance with a prediction accuracy of 0.40% more than the second-best performing model.