Comparison of Statistical and Machine Models for Predicting Cash Holdings and Providing the Optimal Model

The current paper has investigated the comparison of the accuracy of machine learning and statistical models in predicting cash holdings using a set of financial and economic variables. Research methodology can be divided into three stages: selection of data set and variables, modeling and estimation. The statistical sample of the current research is the Tehran Stock Exchange, where the data of 173 companies have been analyzed during the period of 2010-2021. The results indicate the high accuracy of the symbolic regression model using the genetic algorithm with an accuracy factor of 71% in this field. After that, Gradient Boosted Trees, MARS regression, neural network and XGboost models were evaluated as the most accurate models for prediction. Finally, the KNN model showed the weakest prediction accuracy. Also, although the statistical models showed low prediction accuracy, they obtained a higher accuracy coefficient from some machine learning models. Also, the results showed that the use of Lasso regression improves the accuracy of statistical models and some machine learning models. This research can add new angles of cash retention forecasting techniques in financial studies, which have not been investigated in financial literature so far.