Learning to rank firms with annual reports

Abstract

The textual content of company annual reports has proven to contain predictive indicators for the company future performance. This paper addresses the general research question of evaluating the effectiveness of applying machine learning and text mining techniques to building predictive models with annual reports. More specifically, we focus on these two questions: 1) can the advantages of the ranking algorithm help achieve better predictive performance with annual reports? and 2) can we integrate meta semantic features to help support our prediction? We compare models built with different ranking algorithms and document models. We evaluate our models with a simulated investment portfolio. Our results show significantly positive average returns over 5 years with a power law trend as we increase the ranking threshold. Adding meta features to document model has shown to improve ranking performance. The SVR & Meta-augemented model outperforms the others and provides potential for explaining the textual factors behind the prediction.