1 - 3 of 3
ItemAugmenting Audit and Control: a Blockchain Based Control Framework (BBCF)( 2021)Audit and control have become key elements of sound corporate governance. While the Three Lines Model (TLM) provides an organizational structure to execute risk and control duties, research and practice show that this model also has limits even when integrated within proper Enterprise Risk Management (ERM) and Internal Control (IC) frameworks. Such control weaknesses could be addressed by leveraging properties of distribution, transparency, and immutability of blockchain technology. To this end, this paper proposes a conceptual control framework based on blockchain technology to augment common control practice with more trustworthy and accountable blockchain based control patterns. The design of the resulting Blockchain Based Control Framework (BBCF) and its prototype are presented and discussed in terms of potential impact in the context of the identified limits and in particular with respect to COSO, the TLM and risks in general. The contribution intends to serve both as a starting point for discussing the evolution of audit and control practice based on blockchain technology, as well as an initial actionable prototype for experimentation and further development.
ItemCan Artificial Intelligence Detect Biased Client Statements to Improve the Auditor-Client Inquiry Process?( 2021)Prior research has identified extensive limitations in the auditor-client inquiry process. For example, clients can craft inquiry responses to persuade the auditor to accept the client’s aggressive accounting position and research shows that auditors are susceptible to such persuasion attempts. Additionally, research shows that auditors have difficulty identifying deception in client responses and are faced with challenges related to the sheer number of transactions present in modern global enterprises. To address these concerns, we are developing an innovative automated inquiry system that relies on natural language processing and machine learning to evaluate client responses to automated inquiry. In this paper, we follow a design science approach to develop and test the key artifact that learns from and evaluates a set of communication data generated by students who took the roles of auditors and clients. These participants were paired in dyads who communicated over email to discuss a potential inventory obsolescence issue. The clients were randomly assigned to either an aggressive reporter condition (i.e., they aim to report income as high as possible) or an accurate reporter condition (i.e., they aim to report income as accurately as possible). Using a subset of the data, the system learns to identify systematic differences between aggressive and accurate reporters and then evaluates the remaining data to identify whether the clients are aggressive versus accurate reporters. We find that the system properly classifies the clients at a rate greater than chance, demonstrating the feasibility of the technology. Additionally, the system’s learning approach outperforms text analysis using LIWC software which ascribes advance meaning to the words. As audit firms invest heavily in automation, this study has important practical contribution to enhance auditors’ ability to audit enormous, highly complex global companies.
ItemNeural Network Translated into Bag-of-Words: Lexicon of Attentions( 2021)We present a framework that translates trained neural network's decision making process to a lexicon. First, we train an interpretable neural network, hierarchical neural network (HAN) that predicts cumulative abnormal returns (CAR) with analyst reports. Second, we relate the trained attentions with words and compile the analytical results as a parsimonious lexicon. The attention-based lexicon reflects contextual information, and through out-of-sample experiments, we show that it outperforms as well as complements with Loughran and McDonald (LM) lexicon, and also offers a smart weighting scheme that dominates existing word weighting methods. Additional experiments confirm the advantage of the proposed lexicon for earnings call transcripts and generalize its usefulness beyond the original training corpus. Our proposed framework materializes contextual information in financial texts and allows bag-of-words models to incorporate it, and thus it provides subsequent users with a way of exploring contextual information in an interpretable manner.