A unified framework for financial commentary prediction

Abstract

Companies generate operational reports to measure business performance and evaluate discrepancies between actual outcomes and forecasts. Analysts comment on these reports to explain the causes of deviations. In this paper, we propose a machine learning-based framework to predict the commentaries from the operational data generated by a company. We use time series classification to predict labels for the existing commentaries, and compare various machine learning models for the prediction task including XGBoost, long short term memory networks and fully convolutional networks (FCN). Classification models are trained on three datasets and their performance is evaluated in terms of accuracy and F1-score. We consider AI interpretability as an additional component in our framework to better explain the predictions to the decision makers. Our numerical study shows that FCN architecture provides higher classification performance, and Class Activation Maps and SHAP interpretability methods provide intuitive explanations for the model predictions. We find that the proposed framework that is enabled by machine learning-based methods offers new avenues to leverage management information systems for providing insights to the managers on key financial issues including sales forecasting and inventory management.