MARS-Q: Reproducibility Metadata, Data Sources, Walk-Forward Partitions, and Model Configuration

This repository provides the reproducibility metadata and supporting research materials associated with the article “Regime-Aware Multimodal Fusion for Stock Forecasting: A Walk-Forward Evaluation with Realistic Trading Frictions.” The study introduces MARS-Q, a regime-aware multimodal framework for stock-return forecasting that integrates heterogeneous financial information through modality-specific encoders, cross-modal Transformer fusion, volatility-conditioned regime adaptation, concept-drift monitoring, and walk-forward evaluation under realistic transaction costs. The deposited materials are intended to support transparency, methodological traceability, and reproducibility while respecting the licensing restrictions that apply to several proprietary data sources used in the study. The repository does not redistribute licensed raw data from Reuters, Bloomberg, Refinitiv StreetEvents, OptionMetrics, or CRSP. Instead, it documents the public data sources, data-access conditions, model configuration, temporal partitions, feature structure, alignment procedures, and other reproducibility information that can be legally shared. The repository includes the following components: A structured inventory of the publicly accessible data sources used in the study, including SEC EDGAR, Federal Reserve Economic Data (FRED), and Cboe Global Markets, with specific access links. Documentation of licensed data providers and the associated redistribution restrictions. A summary of the asset universe used in the analysis, including the S&P 500, Nasdaq-100, FTSE 100, Nikkei 225, and MSCI Emerging Markets universes. The 19 fixed-length rolling walk-forward partitions used for out-of-sample evaluation between April 2020 and December 2024. The training, validation, and testing date ranges associated with each walk-forward iteration. The six modality streams incorporated into the MARS-Q framework, including social-media information, financial news, SEC filings, earnings-call text and audio, macroeconomic variables, and implied-volatility and technical information. Documentation of the feature-processing architecture, including modality-specific encoders and missing-modality handling procedures. Point-in-time alignment rules and temporal-integrity controls designed to reduce look-ahead bias and information leakage. The principal model and training hyperparameters, including embedding dimension, fusion depth, attention heads, dropout, optimization settings, learning rate, batch size, early-stopping configuration, Page-Hinkley drift-detection threshold, retraining constraints, and random seeds. The computational environment reported in the manuscript, including the main software libraries and hardware configuration. A reproducibility manifest distinguishing between materials included in the repository and materials that cannot be redistributed because of licensing restrictions. The walk-forward evaluation follows a fixed-length rolling design. Each iteration uses the most recent 24 months of observations for training, the subsequent three months for validation, and the following non-overlapping three-month period for out-of-sample testing. The complete structure advances in three-month increments, producing 19 feasible out-of-sample evaluation windows from 1 April 2020 through 31 December 2024. The study uses a combination of publicly accessible and licensed financial data. Public SEC filings, including Forms 10-K, 10-Q, and 8-K, are accessible through the U.S. Securities and Exchange Commission EDGAR database. Macroeconomic and financial variables are obtained from Federal Reserve Economic Data, including the CBOE Volatility Index and the 10-Year Treasury Constant Maturity Rate. Historical VIX information is additionally documented through Cboe Global Markets. Proprietary financial news, earnings-call, implied-volatility, and security-level data were obtained through licensed services and are therefore not redistributed in this repository. The purpose of this deposit is not to reproduce restricted proprietary datasets, but to provide sufficient methodological and computational documentation to allow researchers to understand, audit, and, where legally and technically possible, reconstruct the study workflow using equivalent authorized data sources. The repository should therefore be interpreted as a reproducibility and data-availability resource accompanying the published manuscript. Researchers using these materials should cite both the associated article and this Zenodo record. Any reuse of third-party data remains subject to the original providers’ licensing, access, and redistribution conditions.