Financial Statement Analysis with Large Language Models

We investigate whether an LLM can successfully perform financial statement analysis in a way similar to a professional human analyst. We provide standardized and anonymous financial statements to GPT4 and instruct the model to analyze them to determine the direction of future earnings. Even without any narrative or industry-specific information, the LLM outperforms financial analysts in its ability to predict earnings changes. The LLM exhibits a relative advantage over human analysts in situations when the analysts tend to struggle. Furthermore, we find that the prediction accuracy of the LLM is on par with the performance of a narrowly trained state-of-the-art ML model. LLM prediction does not stem from its training memory. Instead, we find that the LLM generates useful narrative insights about a company's future performance. Lastly, our trading strategies based on GPT's predictions yield a higher Sharpe ratio and alphas than strategies based on other models. Taken together, our results suggest that LLMs may take a central role in decision-making.

DEBATE: Devil's Advocate-Based Assessment and Text Evaluation

As natural language generation (NLG) models have become prevalent, systematically assessing the quality of machine-generated texts has become increasingly important. Recent studies introduce LLM-based evaluators that operate as reference-free metrics, demonstrating their capability to adeptly handle novel tasks. However, these models generally rely on a single-agent approach, which, we argue, introduces an inherent limit to their performance. This is because there exist biases in LLM agent's responses, including preferences for certain text structure or content. In this work, we propose DEBATE, an NLG evaluation framework based on multi-agent scoring system augmented with a concept of Devil's Advocate. Within the framework, one agent is instructed to criticize other agents' arguments, potentially resolving the bias in LLM agent's answers. DEBATE substantially outperforms the previous state-of-the-art methods in two meta-evaluation benchmarks in NLG evaluation, SummEval and TopicalChat. We also show that the extensiveness of debates among agents and the persona of an agent can influence the performance of evaluators.

Corporate Bankruptcy Prediction with Domain-Adapted BERT

This study performs BERT-based analysis, which is a representative contextualized language model, on corporate disclosure data to predict impending bankruptcies. Prior literature on bankruptcy prediction mainly focuses on developing more sophisticated prediction methodologies with financial variables. However, in our study, we focus on improving the quality of input dataset. Specifically, we employ BERT model to perform sentiment analysis on MD&A disclosures. We show that BERT outperforms dictionary-based predictions and Word2Vec-based predictions in terms of adjusted R-square in logistic regression, k-nearest neighbor (kNN-5), and linear kernel support vector machine (SVM). Further, instead of pre-training the BERT model from scratch, we apply self-learning with confidence-based filtering to corporate disclosure data (10-K). We achieve the accuracy rate of 91.56% and demonstrate that the domain adaptation procedure brings a significant improvement in prediction accuracy.

Detecting Rumor Veracity with Only Textual Information by Double-Channel Structure

Kyle (1985) proposes two types of rumors: informed rumors which are based on some private information and uninformed rumors which are not based on any information (i.e. bluffing). Also, prior studies find that when people have credible source of information, they are likely to use a more confident textual tone in their spreading of rumors. Motivated by these theoretical findings, we propose a double-channel structure to determine the ex-ante veracity of rumors on social media. Our ultimate goal is to classify each rumor into true, false, or unverifiable category. We first assign each text into either certain (informed rumor) or uncertain (uninformed rumor) category. Then, we apply lie detection algorithm to informed rumors and thread-reply agreement detection algorithm to uninformed rumors. Using the dataset of SemEval 2019 Task 7, which requires ex-ante threefold classification (true, false, or unverifiable) of social media rumors, our model yields a macro-F1 score of 0.4027, outperforming all the baseline models and the second-place winner (Gorrell et al., 2019). Furthermore, we empirically validate that the double-channel structure outperforms single-channel structures which use either lie detection or agreement detection algorithm to all posts.

From Transcripts to Insights: Uncovering Corporate Risks Using Generative AI

We explore the value of generative AI tools, such as ChatGPT, in helping investors uncover dimensions of corporate risk. We develop and validate firm-level measures of risk exposure to political, climate, and AI-related risks. Using the GPT 3.5 model to generate risk summaries and assessments from the context provided by earnings call transcripts, we show that GPT-based measures possess significant information content and outperform the existing risk measures in predicting (abnormal) firm-level volatility and firms' choices such as investment and innovation. Importantly, information in risk assessments dominates that in risk summaries, establishing the value of general AI knowledge. We also find that generative AI is effective at detecting emerging risks, such as AI risk, which has soared in recent quarters. Our measures perform well both within and outside the GPT's training window and are priced in equity markets. Taken together, an AI-based approach to risk measurement provides useful insights to users of corporate disclosures at a low cost.