AI-powered forecasting boosts retail efficiency: LSTM outshines traditional models

Demand forecasting remains one of the most complex challenges in retail management. As consumer behavior evolves rapidly, traditional statistical models have struggled to interpret nonlinear, dynamic patterns in sales data. The authors note that artificial intelligence, particularly machine learning and deep learning, has emerged as a transformative force capable of recognizing patterns and learning from historical datasets.

In a new comparative analysis of artificial intelligence applications in retail, researchers have revealed that advanced deep learning models can dramatically enhance the accuracy of demand forecasting. Their study concludes that Long Short-Term Memory (LSTM) networks outperform traditional and ensemble models in predicting consumer demand fluctuations, offering a decisive edge for inventory and supply chain management.

Published in Engineering Proceedings under the title “Optimization of Forecasting Performance in the Retail Sector Using Artificial Intelligence” (2025, Vol. 112, No. 37), the research provides empirical evidence that AI-powered forecasting tools can significantly improve operational efficiency and help retailers make data-driven decisions. Conducted using a publicly available dataset from Kaggle, the study compares the performance of six widely used algorithms, Linear Regression, Decision Tree, Random Forest, XGBoost, Prophet, and LSTM, under identical conditions to assess which approach best predicts retail product demand.

The growing role of AI in retail forecasting

Demand forecasting remains one of the most complex challenges in retail management. As consumer behavior evolves rapidly, traditional statistical models have struggled to interpret nonlinear, dynamic patterns in sales data. The authors note that artificial intelligence, particularly machine learning and deep learning, has emerged as a transformative force capable of recognizing patterns and learning from historical datasets.

According to the study, forecasting accuracy has a direct impact on inventory optimization, waste reduction, and supply chain responsiveness. By applying six models to more than 100,000 sales records, the researchers sought to identify the most efficient algorithmic approach to predict high and low demand categories. The dataset included product codes, warehouse IDs, and daily order volumes, variables essential to capturing the temporal dependencies that shape demand cycles.

A meticulous preprocessing phase involved cleaning the dataset, encoding categorical variables, and removing inconsistencies. The research team also introduced a binary classification scheme based on the average demand threshold of 5248.1 units, dividing predictions into low- and high-demand categories. This binary transformation reflects a realistic business decision-making process, where managers often act on whether demand surpasses a critical limit rather than focusing solely on exact figures.

By designing this unified framework, the authors established a fair ground for model comparison across both regression and classification performance metrics such as accuracy, precision, recall, and F1-score. The evaluation was conducted using Python 3.12, with implementations in Scikit-learn, TensorFlow/Keras, XGBoost, and Prophet libraries.

Comparing models: From regression to deep learning

The study’s comparative results reveal a clear performance hierarchy among the models. Traditional Linear Regression performed the weakest, achieving an accuracy of only 60.67 percent and an F1-score of 39.67 percent, underscoring its limitations in handling nonlinear data. XGBoost, a popular gradient boosting algorithm, improved performance with an accuracy of 83.21 percent but suffered from lower precision levels.

In contrast, ensemble models such as Random Forest and Decision Tree delivered more reliable results. The Random Forest model achieved 91.99 percent accuracy with the lowest root mean square error (RMSE) of 21,527.8, indicating strong generalization capabilities for structured datasets. The Decision Tree, known for its interpretability, scored 93.05 percent accuracy and balanced recall and F1 measures, making it a strong classical alternative.

However, it was the LSTM model that achieved a decisive lead. With an accuracy of 92.31 percent, precision of 92.31 percent, recall of 100 percent, and an F1-score of 96 percent, it demonstrated unmatched ability to model temporal dependencies inherent in retail data. The LSTM’s sequential learning architecture allowed it to interpret time-based signals and long-term relationships, which other models could not capture effectively.

The Prophet model, developed by Facebook for time-series forecasting, closely followed with accuracy of 85.71 percent and an equally high recall and F1-score. While Prophet excelled at identifying seasonal patterns and long-term trends, its forecasts were slightly less adaptive to sudden demand spikes compared to LSTM.

The authors highlight that ensemble and deep learning approaches, particularly LSTM, offer the most robust solutions for dynamic, data-intensive retail environments. The integration of these models enables retailers to anticipate customer needs, optimize stock levels, and avoid costly imbalances between supply and demand.

Implications for the future of retail decision-making

The study not only compares algorithms but also underscores the real-world implications of AI adoption in retail. Accurate forecasting directly influences key operational outcomes, preventing stockouts, minimizing overstocking, and aligning procurement cycles with consumer behavior. The authors argue that implementing advanced AI forecasting systems allows companies to shift from reactive to proactive management, achieving greater resilience against market volatility.

Beyond immediate operational benefits, AI-driven forecasting contributes to sustainability by reducing waste and optimizing resource allocation. The automation of forecasting also supports better pricing strategies, workforce planning, and logistics scheduling. As the researchers explain, aligning predictive models with business processes enables data-driven decision-making that enhances profitability and customer satisfaction simultaneously.

The findings reinforce the increasing necessity for time-aware models in retail. LSTM’s ability to process sequential data makes it ideal for industries where consumer demand varies across days, weeks, and seasons. The results also validate Prophet’s interpretability as a supporting tool for understanding trends and seasonality, attributes valuable for long-term strategic planning.

While the study confirms the superiority of deep learning, it also recognizes that computational efficiency remains a factor in model selection. LSTM’s longer training time (20.6 seconds) and moderate memory usage contrast with the faster, lighter XGBoost and Random Forest models, which may still be preferred when real-time forecasting speed is critical. The authors emphasize that selecting a model depends on balancing accuracy, interpretability, and computational demands.

Toward smarter forecasting systems

The authors recommend that future research should explore hybrid models combining sequential and non-sequential techniques to further enhance robustness. Integrating external variables, such as weather conditions, holidays, and marketing activities, could also refine predictions and expand the models’ practical applicability. The inclusion of more advanced feature engineering and continuous learning pipelines is expected to make forecasting systems even more responsive and precise.

AI adoption in retail is not merely a technological upgrade but a strategic transformation. By enabling retailers to anticipate demand with precision, AI serves as a critical tool for maintaining competitiveness in an increasingly data-driven marketplace.