MD&A-ESP: Data and code for the revised manuscript
Description
This dataset provides the analytical data, out-of-sample predictions, statistical results, figure-source data and code accompanying “What Matters in MD&A Text: A Modular Decision Support System for Attributing Predictive Content Across Sentiment, Structure, and Fundamentals”. It corresponds to the frozen Revision V2 experiment used in the October 2026 revised manuscript. The study examines annual and semi-annual reports of Chinese A-share firms over input report years 2016–2024 and predicts next-fiscal-year return on equity (ROE) and return on assets (ROA). The accepted union panel contains 20,450 firm-report observations. Target-specific samples contain 19,983 ROE observations and 20,432 ROA observations. Historical rolling out-of-sample evaluation covers report years 2022–2024, with 9,366 ROE and 9,616 ROA observations. The package contains 1,908 formal prediction runs comprising 6,036,276 prediction rows, sample and variable records, model and information-layer attribution results, statistical comparisons, robustness analyses, computational-cost tables, and the data used for the manuscript figures. Six model families compare financial fundamentals, disclosure quantity and sentiment in matched F, F+Q, F+S and F+Q+S configurations. RandomForest, XGBoost, LightGBM and MLP use five fixed seeds; OLS and ElasticNet are deterministic. English documentation describes the variables, original sources, external resources, environment and reproduction steps. Code is provided for data processing, feature extraction, model training, performance evaluation, statistical analysis and figure generation. The three archives form one folder structure and should be extracted into the same directory. Keywords Management discussion and analysis; Corporate profitability; Financial text analysis; Disclosure quantity; Sentiment analysis; Information-layer attribution; Out-of-sample prediction; Machine learning; Reproducible research Version update note Updated to the frozen Revision V2 materials used in the revised manuscript: target-specific analytical samples, 2022–2024 historical rolling evaluation, five stochastic seeds, matched F/Q/S comparisons, formal prediction files, attribution and statistical results, robustness analyses, manuscript figure data, reproduction code and English documentation. This version replaces the earlier single-2024-holdout package.
Files
Steps to reproduce
Extract all three archives into a common directory. Install environment/requirements-postprocess.txt. Run python code/reproduce_results.py --output ../R1_results to reconstruct performance metrics and information-layer contributions from the supplied predictions, and python code/replot_manuscript_figures.py --output ../R1_figures to generate the manuscript figure topics. The reproduction guide describes additional statistical-analysis and training commands and their inputs. File descriptions 01_documentation_code.zip: English documentation, data dictionary, reproduction guide, Python code, frozen settings, dependency specifications and file manifest. 02_analysis_predictions.zip: Target-specific sample records, numerical research features, source links and all formal out-of-sample prediction files. 03_results_figures.zip: Model and statistical results, robustness tables, selected settings and run records, figure-source data, generated plots and final manuscript TIFFs.
Institutions
- China University of Geosciences (Beijing)Beijing, Beijing