Comparing the Performance of XGBoost and Random Forest Models for Predicting Company Cash Position

Abstract

This study explores the use of XGBoost and Random Forest machine learning algorithms for predicting cash flow using transaction data from small and medium-sized enterprises (SMEs). The research aimed to identify performance differences between the two algorithms and assess their feasibility for practical use by accountants in their daily operations. While both algorithms showed potential, the Random Forest model marginally outperformed XGBoost, but the performance varied depending on the training data used. Despite XGBoost exhibiting a faster model training process, neither model yielded predictions reliable enough for practical use by accountants, with an average error rate of approximately 101.94% of the target variable's average magnitude. The complex nature of company's transactions and the limitations in the dataset used could have contributed to this low performance. This research contributes new insights to the domain of cash flow prediction, highlighting the need for more accurate and reliable machine learning models for this purpose and suggesting a potential path for further research to explore other models and incorporate additional company-related features. However, a more critical aspect to address would be the enhancement of data quality and the identification of clearer patterns within the dataset, as these factors significantly influence the predictive performance. These findings can guide future investigations and efforts in not only improving cash flow predictions for SME, but also advancing the broader field of time series forecasting.