Forecasting Sales Trends Using Time Series Analysis: A Comparative Study Of Traditional And Machine Learning Models

Abstract

This paper aims to analyze the merits and demerits of using traditional time series models and the more sophisticated machine learning approaches in sales forecasting with the view of determining their effectiveness in different conditions. These models include the ARIMA and Exponential smoothing models which are popular because they are easy to understand and easy to compute and as such are useful in the short term forecasting of data that has a clear seasonal pattern. But these models are not very effective with non-linear and complex data structures. On the other hand, XGBoost, LightGBM, and DeepAR models are found to outperform other models in terms of accuracy for high dimensional and highly volatile data. These models are complex and computationally intensive, and the interpretability of these models is also relatively low; however, they offer better flexibility and improved prediction for dynamic forecasting. The use of hybrid models which incorporates both the conventional and the machine learning models is identified to be as being effective for organizations that want to achieve high prediction, model transparency and computational speed. This study offers implications for model selection based on data complexity, forecast horizon, and business needs to enrich the literature on data-driven decision making in sales forecasting.