Dr. Xiao Fang, University of Delaware

Title: A Deep Learning Approach to Industry Classification

Abstract: Business researchers and practitioners often attempt to divide firms into more homogeneous groups that bear similar characteristics. In numerous instances, industry classification systems (ICSs) have been selected to achieve this objective. Traditional expert-driven ICSs, such as Standard Industrial Classification (SIC) and Global Industry Classification Standard (GICS), have several issues including limited granularity and high cost of maintenance. An alternative algorithm-driven approach to ICS infers similarities between firms from their financial documents (e.g., annual 10-K reports). While the algorithm-driven approach is promising in circumventing the limitations faced by the expert-driven approach, one needs to solve the challenging problem of text representation, that is, how to accurately represent textual documents with numerical features. In this paper, we propose a novel text representation model based on deep learning. The proposed model is capable of handling arbitrary long documents with heterogeneous concepts. It balances the trade-off between memorizing local details and capturing global concepts by introducing a multiplicative mechanism to integrate information from contexts of different scopes. We further develop an ICS by applying the proposed text representation model to Item 1 and 1A sections of firms’ annual 10K reports. We demonstrate the effectiveness of the proposed text representation model by benchmarking the developed ICS against state-of-the-art expert-driven and algorithm-driven ICSs.

Learn more about Deep Analytics

Read more about Dr. Fang