Classification of Sales Time Series Through Trend Measurement: Forecasting Methods Comparison

Abstract

Time series forecasting is a crucial tool that utilizes historical data to project future events based on previous patterns. These projections provide an early glimpse into the future and serve as a foundation for critical decision-making. Therefore, the ability to perform accurate forecasting is essential for effective planning, optimizing inventory management in sales, and guiding investment decisions in finance. This chapter presents a methodology for calculating the trend. The methodology was tested using 300 randomly selected time series from the M5 competition to select two subsets of time series based on the identified trend. For testing we chose a classical and two machine learning-based forecasting methods. We generate forecasts for the two early mentioned subsets and the original set for evaluation using sMAPE. According to Friedman and Wilcoxon statistical tests, the LightGBM method was consistently the most accurate, outperforming ARIMA, and LSTM neural network in seven of nine of the Friedman rankings while producing a close to zero p-value.