This study examines the influence of preprocessing techniques in relation to time series forecasting using the example of demand prediction. By conducting 80 different experiments with various preprocessing methods like transformation (Log, Box-Cox, Yeo-Johnson), detrending (differencing), and scaling (standardization, min-max and robust scaling), insights about the behavior and correlation between the preprocessing step and a forecasting task are gathered. Using the M5 dataset with real-world sales data from Walmart and two different neural network types, one Multi-Layer-Perceptron and one Long Short-Term Memory network, the impact of preprocessing on a model’s performance can be shown. The results emphasize the crucial role of preprocessing in improving model accuracy and form the basis for future research on automatic preprocessing selection based on time series characteristics.
Grimm et al. (Thu,) studied this question.