Automated Selection of Time Series Forecasting Models for Financial Accounting Data: Synthetic Data Application

Abstract

The continuous forecasting of anticipated trends in company accounting helps to prepare for possible challenges and investment possibilities. The forecasting of performance indicators for each partner and metric in a company requires a high volume of resources, as each forecasting model requires some supervised adjustment. To solve the challenge of manual work in this task, AutoML solutions can be used. In this study, we propose an automated forecasting model selection option. The time series data are summarized by 21 statistical features and different experiments are performed to find the best-suited forecasting method. Different classifier models are tested with the dataset, estimating the impact of data sampling and synthetic data generation. The results indicate that the undersampling approach is more suitable, as it helps in balancing the classes. Random forest methods usually show the best performance, achieving about 74% accuracy. The usage of a synthetic data-based dataset for model training reduced the accuracy by almost 20%, while the integration of synthetic and real time series data allowed us to achieve balance between both classes. This confirms that synthetic time series data generation might increase the accuracy of forecasting method selection, but it should be used in combination with real data.