Deep reinforcement learning model audits and improves listed companies’ disclosure quality

Researchers in China have built an intelligent audit model that uses deep reinforcement learning to monitor and improve the quality of accounting information disclosure by listed companies, reporting average accuracies of 88.67 percent in identifying disclosure quality grades and 86.50 percent in detecting violations. The work, published as an open access research article in Discover Artificial Intelligence, addresses a persistent weakness in current oversight: static quantitative screening models that cannot adapt as market conditions, regulatory policies and company circumstances change.

The study focuses on environmental accounting information disclosure, the environmental responsibility reporting that listed companies provide to shareholders, investors, regulators, the public and environmental organizations. Drawing on stakeholder theory and signaling theory, the authors argue that firms with strong environmental management and low environmental risk tend to disclose detailed environmental accounting information to signal sustainable development, attract green investment and reduce financing costs. The framework for what should be disclosed follows the environmental, social and governance guidelines of the Hong Kong Stock Exchange, which cover four general disclosure indicators and thirteen key performance indicators under an “explain if you do not comply” clause.

At the heart of the system is a dual recording mechanism for operation and maintenance behavior logs. When users connect to servers through character transmission protocols such as Secure Shell, using tools like PUTTY or XShell, every keystroke and every server response travels in both directions. The audit system records the instructions input by the user and, separately, the data transmitted by the server, which contains those instructions superimposed with execution feedback. If the full logs and compact logs disagree, the model flags the discrepancy as a suspicious signal and triggers a high-precision audit. The authors describe this dual recording as the data foundation for dynamic auditing and a source of robustness, with results stored both in files and in a dedicated audit results table in a database.

The auditing procedure is formulated as a Markov Decision Process, the standard mathematical scaffold for sequential decision-making. The state space has 27 dimensions: ten quantified disclosure influencing factors, six dimensions of operation and maintenance log features, one quantified disclosure quality grade, and ten interaction terms between the influencing factors and the log features. The action space contains eight executable audit actions in three groups: quality grading judgments, violation identification, and dynamic optimization and rectification suggestions. The ten influencing factors include corporate asset size, ownership concentration, executive shareholding ratio, internal control effectiveness, environmental penalty records, green research and development investment, ESG rating, industrial pollution attribute, mandatory policy disclosure requirements, and financing constraint level. The six log dimensions cover audit operation instruction sequences, server-side data feedback, operation timestamps, anomaly marking records, verification field types, and operator permission labels.

Deep Q Network, or DQN, serves as the core algorithm. The approach, proposed by Google’s DeepMind team, combines the feature extraction ability of deep learning with the decision optimization ability of reinforcement learning. DQN pairs a dual playback mechanism with a target network, which the authors say increases training stability and convergence speed while preventing overfitting or oscillation. The reward function blends three weighted components: audit accuracy, the change in disclosure quality after an audit action, and audit cost, with weights set at 0.55, 0.30 and 0.15 respectively. A discount factor of 0.92 balances immediate audit gains against long-term returns from disclosure quality improvement, a value chosen after comparative experiments. The initial learning rate is 0.0001 with adaptive decay, the experience replay buffer holds 10,000 samples, the target network updates every 200 steps, and the batch size is 64.

The empirical work uses annual reports from 2018 to 2022 for 200 Shanghai and Shenzhen A-share listed companies, together with third-party data from the Wind database. To avoid distortion from a single sector, the sample spans both heavy-pollution and non-heavy-pollution industries, with 25 companies chosen from each industry. Samples with more than 20 percent missing data were removed, and remaining gaps were handled by multiple imputation, though the authors caution that the uncertainty introduced by imputation may harm the model’s generalization. The data were split into 70 percent training, 20 percent validation and 10 percent test sets, and features were normalized to the interval from zero to one using Min-Max standardization. Violation labels were assigned by a three-person professional labeling team working with regulatory punishment records, with labeling consistency verified by a Kappa value of at least 0.85. Experiments ran on a Windows Server 2022 machine with an Intel Core i9-13900K processor and an NVIDIA RTX 4090 graphics card, using Python 3.9.10, TensorFlow 2.10.0 and OpenAI Gym 0.26.2.

In the disclosure quality identification task, the model achieved an average accuracy of 88.67 percent and an F1 score of 87.92 percent, grading environmental accounting information as excellent, good, qualified or unqualified. In violation detection, it reached 86.50 percent accuracy and an F1 score of 85.73 percent, identifying false disclosure, omitted disclosure, delayed disclosure and other violations. Against four comparison algorithms, the margins were substantial: accuracy and F1 in quality identification improved by 13.34 and 13.07 percentage points over K-Nearest Neighbors and by 10.17 and 10.24 points over Bidirectional Long Short-Term Memory. The authors attribute KNN’s weaker showing, with F1 scores below 84 percent on both tasks, to its inability as a lazy learning algorithm to capture temporal trends, while Bi-LSTM extracts sequential features but cannot adjust policies based on action outcomes. Policy gradient methods such as PPO and A2C suit continuous action spaces, but this study’s discrete, low-dimensional action space produced elevated training variance and higher hyperparameter sensitivity for them.

Beyond classification, the model generates rectification suggestions. In a test of 50 samples with unqualified disclosure quality or potential violations, 20 samples with illegal behaviors were revised following the model’s optimization advice, lifting the quantitative disclosure quality value from 0.58 to 0.82, an average improvement rate of 41.38 percent, with all samples reaching qualified level or above. In a quality improvement scenario, 30 samples that were already qualified but improvable rose from 0.71 to 0.88 after adjustment, an average improvement rate of 23.94 percent, moving them closer to the excellent level. Ablation experiments showed that removing the target network mechanism cut precision by 4.00 to 4.70 percent and raised the standard deviation of stability by 1.38 percent, making it the most influential module, because it anchors the target Q value calculation and prevents divergence during training. In sample disturbance tests, a 10 percent perturbation of test-set features reduced quality identification accuracy by only 1.13 percent and violation detection accuracy by 1.14 percent, with all index fluctuations below 1.2 percent; even at 30 percent disturbance, accuracy held at 85.12 percent for quality identification and 83.17 percent for violation detection.

The authors position the work as the first application of a DQN deep reinforcement learning framework to the dynamic audit of environmental accounting information disclosure, offering listed companies a tool to maximize disclosure quality and regulators an intelligent auditing instrument. They also note the study’s limitations: model performance depends heavily on data quality, and the current research covers only environmental disclosure. Future work, they write, could extend toward a generic audit framework spanning domains and markets, broaden the model’s application scope, and strengthen data governance and cleaning technologies.

For capital markets, the implications are practical. China’s A-share market exceeded 100 trillion yuan in total value in 2023, and poor disclosure quality can lead investors to make wrong choices and destabilize the market. Manual audit processes were slow, error-prone and costly, and regulators struggled to find anomalies in time, eroding transparency and trust. A system that learns optimal audit strategies through continuous interaction with the auditing environment, updates in real time as conditions shift, and withstands moderate data disturbance offers a technical path toward keeping pace with that scale, though its real-world value will depend on the quality of the data it is fed.

Subject of Research: A deep reinforcement learning model for dynamically auditing and optimizing the environmental accounting information disclosure quality of listed companies

Article Title: Dynamic control and optimization model of accounting information disclosure quality of listed companies based on deep reinforcement learning

Article References: Hu, T. (2026). Dynamic control and optimization model of accounting information disclosure quality of listed companies based on deep reinforcement learning. Discover Artificial Intelligence, 6(1), Article 1204. https://doi.org/10.1007/s44163-026-02295-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02295-9

Keywords: deep reinforcement learning, DQN, accounting information disclosure, listed companies, intelligent audit, environmental accounting, disclosure quality, violation detection, Markov Decision Process, A-share market, target network, ESG disclosure