On Monday, October 27, 2025, at 16:00, Mr. Konstantinos Panoutsakos of the
postgraduate program “Data Science and Information Technologies”, track on “Big Data and Artificial Intelligence”, will present his MSc thesis
titled:
“From Audit Opinion Analysis to Stock Market Signals: Bankruptcy Prediction and Short Selling Strategies”
Abstract
This thesis examines the potential predictive value of audit opinions for forecasting corporate bankruptcy and subsequent stock price movements within the U.S. market. Most prior work on bankruptcy prediction relies on comprehensive financial disclosures (such as Item 10 filings) or combines these with audit opinion data. This research takes a distinct approach, isolating audit opinions as the sole predictive source due to their ready accessibility, interpretability, and computational efficiency for constrained modeling environments.
The study constructed its working dataset from a subset of the ECL data, focusing on U.S. stocks from 1998 to 2021. Critically, each stock required both an available audit opinion and a verifiable bankruptcy outcome (specifically, whether bankruptcy occurred within one year of the opinion’s issuance).
To address the severe class imbalance—where bankruptcies constituted less than 1% of the original sample—the minority class was augmented using lightweight large language models (LLMs), each having fewer than eight billion parameters. This technique effectively expanded the data, rebalancing the class distribution to approximately 86%–14%, thereby providing a more stable foundation for model training. Four machine learning models were then created: three stacked ensemble classifiers and a logistic regression baseline.
Moving beyond traditional bankruptcy forecasting, the resulting probabilities were employed as trading signals to initiate short positions immediately following the release of an audit opinion Interestingly, the model with the weakest F1 score for bankruptcy prediction produced the highest trading returns. This key finding suggests that audit opinions indicating significant financial uncertainty—even in cases that do not result in formal bankruptcy—can precede substantial declines in stock performance.
In sum, this research connects financial auditing, natural language processing, and quantitative finance, indicating that audit opinions, when processed via lightweight LLM-assisted augmentation and classical ensemble learning, may offer valuable predictive insights for both bankruptcy forecasting and anticipating adverse stock price movements.
EXAMINATION COMMITTEE:
Dr Perantonis Stavros, Head of CIL Lab, Institute of Informatics & Telecommunications (IIT), National Centre for Scientific Research “Demokritos”
Dr Zavitsanos Ilias, Research Scientist (C), National Centre for Scientific Research “Demokritos”
Assistant Prof. Panagiotis Stamatopoulos, Department of Informatics and Telecommunications of the University of Athens
Topic: DSIT: Master’s Thesis Presentation, Konstantinos Panoutsakos
Time: Oct 27, 2025 04:00 PM Athens
Join Zoom Meeting
https://us06web.zoom.us/j/81131108334?pwd=1UcoAO6e3AOsr7EjDtXsmFKLlrpEAm.1
Meeting ID: 811 3110 8334
Passcode: 382966
Leave A Comment