Forecasting business exceptions in robotic process automation with machine learning

Abstract

Business exceptions interrupt robotic process automation (RPA) workflows and oblige costly human intervention. This paper explores the application of machine learning (ML) time series forecasting techniques to predict business exceptions in RPA. Using RPA robot logs from a financial service company, we employ ARIMA, SARIMAX, and Prophet statistical models, comparing their performance with ML models such as XGBoost and LightGBM. Furthermore, we explore hybrid approaches that combine the strengths of statistical models with ML techniques, specifically integrating Prophet with XGBoost and LightGBM. Our findings reveal that a hybrid LightGBM model substantially outperforms traditional methods, achieving a 40% reduction in the weighted absolute percentage error (WAPE) when compared to the top-performing statistical model. These results suggest the potential of ML forecasting in optimizing RPA operations through the analysis of log-generated data.