Abstract

Whether deep learning architectures offer a meaningful accuracy advantage over traditional machine learning methods for retail sales prediction remains an actively debated question, with published comparative studies reporting inconsistent conclusions depending on dataset size, feature structure, and task formulation. This study empirically compares five algorithms on a 5,000-order e-commerce transactions dataset: three traditional machine learning methods (Linear Regression, Support Vector Regression, and Random Forest), one modern ensemble boosting method (Histogram-based Gradient Boosting, reported as the XGBoost-equivalent since the dedicated XGBoost library was unavailable in the study's offline computational environment), and one neural-network deep-learning representative (a Multilayer Perceptron). The two sequence-specialised deep-learning architectures originally specified for this study: Long Short-Term Memory (LSTM) and Bidirectional Gated Recurrent Unit (Bi-GRU) networks require TensorFlow or PyTorch, neither of which was installed nor installable in the execution environment (confirmed via direct connectivity testing), and this study explicitly declines to fabricate results under those algorithm names; instead, Section 2 synthesises what the peer-reviewed literature reports about their expected relative performance, clearly distinguished throughout from this study's own empirical findings. Among the five models actually trained and evaluated, Histogram-based Gradient Boosting achieved the strongest performance (RMSE = $21.19, R² = 0.9993), followed by Random Forest (R² = 0.9985) and Support Vector Regression (R² = 0.9857), while the Multilayer Perceptron, this study's deep-learning representative — substantially underperformed all three traditional/ensemble methods (R² = 0.9676) and even trailed Linear Regression on several test folds' worth of the tail distribution, despite requiring roughly 100 times Linear Regression's training time. This pattern, traditional and ensemble methods outperforming a neural network on a moderately sized, low-noise tabular dataset directly replicates findings reported in several published comparative studies reviewed in Section 2, reinforcing the conclusion that architectural sophistication does not guarantee superior accuracy, particularly for structured, moderately sized retail transaction data. The study concludes with a decision framework for practitioners choosing between traditional ML and deep learning for retail sales prediction, and specifies the exact steps required to complete the LSTM/Bi-GRU comparison once appropriate deep-learning infrastructure becomes available.

Keywords