Does a company has bright future? Predicting financial risk from revenue reports

This paper investigates predicting the financial risk of publicly-traded corporations using their revenue reports. Unlike many existing algorithms where a prediction model is learnt using real-valued ground truth risks, we propose to solve the prediction as a learning-to-rank problem with pairwise constraints (e.g., company A is financially more stable than company B). To further increase the flexibility of our approach, we solve the pairwise learning formulation in its dual format, which makes our model nonlinear and thereby can be applied to complex prediction tasks. The advantage of using pairwise supervision is not just limited to the easier acquisition of training data, it also motivates new problem settings. We explore one such setting - the prediction model can actively ask humans informative questions so as to improve the prediction accuracy. Our work aims to address three limitations of existing works: (i) Pointwise supervision - we adopt pairwise supervision which reduces the cost of collecting training samples; (ii) Linearity - we kernelize the formulation to make it nonlinear which would broaden its applicability; (iii) Training data bottleneck - the proposed model can actively involve humans into the learning loop, such that when the initial training samples does not carry enough knowledge, additional examples can be added to learn a better prediction model. Using the proposed efficient optimization method, we evaluate our approach on real text files (annual revenue reports) and compare with state-of-the-art methods. The superior empirical result demonstrates the performance of our proposed approach, and validates the effectiveness of our active knowledge injection in the context of human-machine interaction.