Retail Sales Forecasting using Artificial Intelligence (AI)

Generating more accurate sales prediction outcomes can be complex and challenging due to the nature of the data, in particular, higher dimensionality, nonlinearity, trends, seasonality, and noise. Sales forecasting has been transformed by integrating the advanced analytical capabilities of Artificial Intelligence (AI) and machine learning (ML) methods. However, often referred to as “black box” models, they are challenging to interpret and explain how these models make predictions. Hence, interpretability and explainability are essential for gaining insights into sales forecasting models and building trustworthiness in models’ decisions. Considering this phenomenon, this research proposes two Artificial Intelligence (AI) empirical modalities to predict retail sales to produce more accurate outcomes than existing methods and interpret the model’s prediction outcomes using eXplainable Artificial Intelligence (XAI) techniques to assist managerial decision-making.

The primary objective of modality one is to introduce an Artificial Intelligence (AI) paradigm to discern an efficacious AI model that can be used to enhance the accuracy of retail sales forecasting. For this scenario, five forecasting models were employed, including Autoregressive Integrated Moving Average (ARIMA), Elman Recurrent Neural Networks (ERNN), Long Short-Term Memory (LSTM), and Multi-Layered Perceptron Neural Networks (MLP). The ERNN model outperformed the other models with a Root Means Squared Error (RMSE) test of 2.2360. The result indicates that ERNN generates an improvement of 12.56% in prediction accuracy compared to existing methods using the same dataset utilised in this study.

Owning to the inherent opacity of AI models, this research endeavours to integrate Machine Learning (ML) models with eXplainable Artificial Intelligence (XAI) techniques as a succeeding modality introduced in this study. This proposed method employed eight ML techniques in the prediction process, including Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF), Extra Tree (ET), Naïve Bayes (NB), Logistic Regression (LR), and eXtreme Gradient Boosting (XGBoost). Measures such as accuracy, sensitivity, specificity, and F1 score were adopted to evaluate the efficacy of the employed models. Among the models, XGBoost demonstrated superior performance, exhibiting an impressive accuracy of 95.814% compared to the other models. The approach employed three XAI techniques in the interpretation process: SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and SHAPASH. By leveraging XAI methodologies, this research unveils actionable insights concerning the significance of influential features. It emphasises their utility in enhancing managerial decision-making processes in the context of retail sales forecasting.

By highlighting the capabilities of ML techniques in large-scale retail sales prediction, this study opens avenues for further exploration and advancements in the field. It provides a foundation for future research to refine and expand upon the proposed data-driven approach, integrating additional data sources, refining algorithms, and considering contextual factors. This can drive untapped knowledge in the context of retail sales forecasting, reshape best practices, and guide businesses towards more data-driven decision-making.