Data Analytics Approach for Enhanced Sales Forecasting (DAAESF): Feature Selection and Classifier Integration Analysis

Accurately forecasting sales has significant ramifications for producers, distributors, and investors. Sales forecasting accuracy enables businesses to enhance their manufacturing, distribution, and promotional activities. The current research intends to examine the implications of feature selection methods on enhancing the accuracy of seasonal sales forecasts. The author evaluates different feature selection methods in combination with predictive models, aiming to determine their impact on the effectiveness of predictions. Within this research, a diverse set of eight classifiers has been utilized: namely Naïve Bayes, Logistic Regression, Neural Network, Random Forest, J48, IBK, SVM, and K Star. Alongside these classifiers, four distinct feature selection techniques-namely Gainratio, Infogain, Relief, and CFS have also been employed. The effectiveness of these strategies was evaluated individually as well as collaboratively. The outcome of the proposed novel methodology DAAESF engendered a notable advancement in accuracy rates. Combining feature selection techniques with Neural Network led to a 32% accuracy enhancement compared to other classifiers for cement sales prediction, while Naïve Bayes experienced a decline in performance from 19.55% to 32.55% due to its distinct functions. Additionally, feature selection notably improved prediction accuracy across classifiers, with Neural Network achieving up to 22.68% improvement using CFS, and SVM showing gains of 26.3% with Infogain, highlighting the critical role of feature selection in model optimization. Naïve Bayes and J48 exhibited mixed results across datasets and feature selection methods. Additionally, to substantiate the robustness and validity of the observed outcomes, the Friedman test was judiciously applied.