An Analytical Framework & Implementation for Firm's Credibility Assessment

In recent years, a number of market research firms (e.g., Muddy Waters Research and Citron Research) have dedicated resources to investigate possible financial fraud involved with Chinese companies that went public in North American capital market. The market response has been extremely negative. However, it has been pointed out that the evidence and research methods presented by those market research firms are questionable. While focusing on statistical analysis of structured financial and market activity data, conventional auditing practices largely ignore textual information, including but not limited to news reports and various social media contents. Identifying and detecting “whiskers of useful information” from those unstructured online content is a huge challenge, yet holds great opportunities if properly done.Nevertheless, the volume and variety of potentially relevant business information is huge, causing cognitive load on both shareholders and data analysts. To the best of our knowledge, there is no systematic and effective measurement method for the detection of deceptive business information and for the verification of market research firms’ reports. Therefore, in the proposed project, the general research question is: how can we protect the integrity of the financial market by (1) assessing the credibility of financial fraud claims, (2) designing an analytic framework that can automate the fraud discovery process, and (3) accurately detecting and predicting financial frauds. We propose a conceptual research framework based on both primary and secondary research, aiming at predicting the fraud risk of a firm. A prototype system will be developed to implement the framework. Our information sources include financial statements, SAIC files (include audit report), SEC filings, as well as press releases, social network data, private VC company data, online news and research, etc. These unstructured data along with the structured financial information will be fed into the prototype system. Relationship of changes in the data from year to year will also be considered in the credibility assessment. The core enabling techniques to be used in the system include classification, latent text mining, linguistic analysis, outlier detection, and time series analysis.This project contributes to the existing literature by proposing an analytic framework that can automate the credibility assessment process. The proposed framework integrates important cues/signals extracted from various perspectives and performs accurate predictive modelling for business fraud detections. We build a prototype of the framework and demonstrate that it is both effective and efficient. Applications of the prototype system are also discussed in the project. These include improved decision making for the securities & exchange commissions, accounting professionals, individuals, and venture capital firms that seek to identify deceptive companies.