Using account level credit-card data from six major commercial banks from January 2009 to December 2013, we apply machine-learning techniques to combined consumer-tradeline, credit-bureau, and macroeconomic variables to predict delinquency. In addition to providing accurate measures of loss probabilities and credit risk, our models can also be used to analyze and compare risk management practices and the drivers of delinquency across the banks. We find substantial heterogeneity in risk factors, sensitivities, and predictability of delinquency across banks, implying that no single model applies to all six institutions. We measure the efficacy of a bank's risk-management process by the percentage of delinquent accounts that a bank manages effectively, and find that efficacy also varies widely across institutions. These results suggest the need for a more customized approached to the supervision and regulation of financial institutions, in which capital ratios, loss reserves, and other parameters are specified individually for each institution according to its credit-risk model exposures and forecasts

This paper examines the information content of the forward‐looking statements (FLS) in the Management Discussion and Analysis section (MD&A) of 10‐K and 10‐Q filings using a Naïve Bayesian machine learning algorithm. I find that firms with better current performance, lower accruals, smaller size, lower market‐to‐book ratio, less return volatility, lower MD&A Fog index, and longer history tend to have more positive FLSs. The average tone of the FLS is positively associated with future earnings even after controlling for other determinants of future performance. The results also show that, despite increased regulations aimed at strengthening MD&A disclosures, there is no systematic change in the information content of MD&As over time. In addition, the tone in MD&As seems to mitigate the mispricing of accruals. When managers “warn” about the future performance implications of accruals (i.e., the MD&A tone is positive (negative) when accruals are negative (positive)), accruals are not associated with future returns. The tone measures based on three commonly used dictionaries (Diction, General Inquirer, and the Linguistic Inquiry and Word Count) do not positively predict future performance. This result suggests that these dictionaries might not work well for analyzing corporate filings

Many countries have converged their domestic auditing standards with International Standards on Auditing (ISA). This study provides global empirical evidence on first-order determinants of audit quality by examining whether and how convergence affects audit quality through utilizing data on 41 jurisdictions and using a staggered difference-in-differences approach. We find that ISA convergence leads to higher audit quality on average. The positive effect is stronger for clients of domestic audit firms, in jurisdictions with stronger enforcement, and when the ISA convergence level is higher. Insights from textual features suggest that changes in principle-orientation, comparability, readability, and size (or length) of auditing standards are positively related to audit quality. Exploratory analyses of textual content using machine learning reveal that the emphases of ISA on going-concern assessment and legal compliance, fraud risk assessment and internal control evaluation, and related-party transactions and subsequent events contribute to enhanced audit quality

We examine firm disclosure choice during the initial public offering (IPO) roadshow presentation to understand the informativeness of a management presentation designed to attract investors. Although firms submit a comprehensive registration filing during the IPO, managers also prepare a roadshow presentation, which is shorter and typically allows managers more autonomy to select the information released and how it is discussed. We find that IPO roadshows have significantly more positive, less negative, and less uncertain language than the SEC filing. Using machine learning to classify roadshow sentences into five major topics from the registration statement, we find that roadshows differ in both the topics selected and the language used within each topic. We then examine the predictive ability of the roadshow language, finding that roadshow language predicts future accounting performance, whereas filing language does not. These results highlight the informational role of management presentations, despite the flexibility they grant managers

Companies, including public accounting firms, are integrating systems with advanced algorithms into decision-making processes to assist with developing and evaluating complex estimates. However, individuals may hesitate to rely on algorithmic output, particularly under conditions of uncertainty. We conduct two experiments examining whether and how a system’s ability to adapt—an emerging feature of machine learning—interacts with uncertainty to influence accounting professionals’ reliance on algorithmic advice. In Experiment 1, we find that auditors are more willing to rely on advice from learning algorithms than static algorithms when estimation uncertainty is relatively high. Experiment 2 replicates this result in a general accounting context where preparers develop their own estimates. Our findings demonstrate that accounting professionals’ reliance on algorithms is contextually dependent, and highlights algorithm adaptability as an important technological feature that can promote advice utilization, particularly when adaptability is likely important to the judgment context (e.g., when estimation uncertainty is high

The relationships between accounting earnings and stock prices, as well as between unexpected earnings and returns, have received substantial attention in the empirical literature. Several theoretical models predict the shapes of these relationships. However, a comprehensive empirical description that could be used to evaluate these predictions is lacking. By integrating recent advances in statistics and machine learning with findings in the accounting literature, we develop an empirical method to identify the relationships, which is consistent with the firm-specific and nonlinear features of the theoretical models. Our approach provides a clear description of stylized and robust patterns in the relationships that are relevant to distinguish between existing models and to aid future theory development. The findings are consistent with recently proposed dynamic option models for both the earnings-price and unexpected earnings-returns relationships. Data availability: Data are available from the public sources cited in the text. A summary of the R code used in the article is available in Starica and Marton (2024

This study uses machine learning models to forecast future material misstatements. Using raw financial data, audit variables, qualitative features, and an efficient algorithm, we design a dynamic model that continuously updates with new information. Our model outperforms the benchmarks for both one-year-ahead and two-year-ahead predictions in terms of out-of-sample predictive power and economic impact on net income. Using Explainable Artificial Intelligence, we identify key predictive features, including comprehensive income, foreign firm status, and accrued interest and penalties from unrecognized tax benefits. Results show that investors achieve better outcomes using a proactive investment strategy based on our prediction models than reactive detection models. Furthermore, our prediction model can help managers prevent internal control weaknesses, assist auditors in assessing misstatement risks in advance, and enable regulators to allocate inspection resources proactively. Our study advances the literature by moving beyond the detection of past material misstatements to the forecasting of future misstatements. Data Availability: Publicly available

To mitigate information asymmetry about borrowers in developing economies, digital lenders use machine-learning algorithms and nontraditional data from borrowers’ mobile devices. Consequently, digital lenders have managed to expand access to credit for millions of individuals lacking a prior credit history. However, short-term, high-interest digital loans have raised concerns about predatory lending practices. To examine how digital credit influences borrowers’ financial well-being, we use proprietary data from a digital lender in Kenya that randomly approves loan applications that would have otherwise been rejected based on the borrower’s credit profile. We find that access to digital credit improves borrowers’ financial well-being across various mobile-phone-based well-being measures, including monetary transactions and balances, mobility, and social networks as well as borrowers’ self-reported income and employment. We further show that this positive impact is more pronounced when borrowers have limited access to credit, take loans for business purposes, and obtain more credit

This study evaluates alternative measures of the tone of financial narrative. We present evidence that word-frequency tone measures based on domain-specific wordlists—compared to general wordlists—better predict the market reaction to earnings announcements, have greater statistical power in short-window event studies, and exhibit more economically consistent post-announcement drift. Further, inverse document frequency weighting, advocated in Loughran and McDonald (2011), provides little improvement to the alternative approach of equal weighting. We also provide evidence that word-frequency tone measures are as powerful as the Naïve Bayesian machine-learning tone measure from Li (2010) in a regression of future earnings on MD&A tone. Overall, although more complex techniques are potentially advantageous in certain contexts, equal-weighted, domain-specific, word-frequency tone measures are generally just as powerful in the context of financial disclosure and capital markets. Such measures are also more intuitive, easier to implement, and, importantly, far more amenable to replication

How effective are physicians at diagnosing heart attacks? To answer this question, we contrast physician testing decisions with a machine learning model of risk. When the two deviate, we use actual health outcome data to judge whether the algorithm or the physician was right. We find physicians over-test: tests that are predictably useless are still performed. At the same time, physicians also under-test: many predicted high-risk patients are untested and then suffer adverse health events (including death) at high rates. A natural experiment using shift-to-shift testing variation confirms these findings: increasing testing improves health and reduces mortality, but only for patients flagged as high-risk by the algorithm. The simultaneous existence of over- and under-testing cannot easily be explained by incentives alone, and instead suggests errors. We provide suggestive evidence on the psychology behind these errors:(i) physicians use too simple a model of risk, suggesting bounded rationality; (ii) they over-weight salient information; and (iii) they over-weight symptoms that are representative or stereotypical of heart attack. Together, these results suggest the need for health care models and policies to incorporate not just physician incentives, but also physician mistakes