IMPLEMENTING AND COMPARING FORECASTING ALGORITHMS FOR SALES DATA USING PYTHON: FACEBOOK PROPHET, SARIMA AND HOLT-WINTERS

Abstract

The primary objective was to compare the performance of Facebook Prophet, a popular ML algorithm, against established statistical models like SARIMA and Holt-Winters. This study focused on analyzing time-series sales data, acknowledging the significant influence of seasonality on sales patterns. A Python script has been used as a central tool, enabling the evaluation of various forecasting algorithms. This compared Facebook Prophet, SARIMA, and Holt-Winters based on their accuracy, robustness, and applicability to the data. This involved implementing each algorithm in the script and analyzing their performance metrics, such as R-squared scores. The analysis revealed that Facebook Prophet emerged as the most accurate model for the dataset and chosen error metrics. It obtained an R-squared score of 0.94 for fitted data (the best result achieved by using simulated annealing for fine tuning), demonstrating its strong ability to capture the underlying patterns in the sales data. While Holt-Winters performed competitively with an R-squared value of 0.87 (used simulated Annealing in fine tuning) and 0.87 (used simulated annealing in fine tuning) on the training set, it is important to note that these results might vary depending on specific data and evaluation criteria. Interestingly, the remaining algorithms showed relatively minor performance differences, with SARIMA lagging behind at an R-squared score of a 0.61 (the best result by using simulated annealing for fine tuning). This highlights the crucial role of choosing the appropriate algorithm based on individual needs and data characteristics. The second part of the thesis delves into the technical aspects of the Python scripts. We meticulously detail the data analysis and preprocessing steps, unveiling the inner workings of each script with descriptive comments, visuals, and step-by-step execution guides. Witnessing the results unfold, you'll understand how these scripts compare and rank the algorithms based on performance metrics like R-squared. The research underscores the potential of ML in sales forecasting, particularly Facebook Prophet's ability to deliver accurate predictions. Furthermore, the developed Python script offers a versatile tool for running and comparing multiple forecasting models simultaneously. This readily implementable solution worked with businesses across various sectors, from corporations managing inventory to logistics providers anticipating client needs, to leverage historical data and optimize their sales forecasting processes, ultimately contributing to improved efficiency and success.