Replication Dataset for "An Intelligent Decision-Support Framework for Stock Ranking and Portfolio Construction"

This repository contains the processed datasets used in the study “An Intelligent Decision-Support Framework for Stock Ranking and Portfolio Construction”. The data were constructed to support the training, validation, and evaluation of a machine learning framework for cross-sectional stock return ranking across multiple international equity markets. The dataset covers four markets: the United States, the United Kingdom, Germany, and Brazil. For each market, the investment universe consists of the constituents of the main national equity index as of December 2025, as described in the paper. Historical daily price data were obtained from Yahoo Finance and used to compute return-based features. The data include: Stock-level return features: cumulative returns over rolling windows of 1 to 12 months; Market-level features: corresponding return histories of benchmark equity indices; Fixed income features (where available): returns of short- and long-duration bond ETFs used as proxies for interest-rate conditions; Target variables: normalized cross-sectional ranking targets derived from future returns using a probit transformation of percentile ranks. The datasets were preprocessed to ensure consistency across markets, including alignment of trading days, handling of missing values, and transformation of raw prices into return-based features. These processed CSV files correspond directly to the inputs used for training the CNN–LSTM models and for constructing the TWIN ranking scores described in the paper. The raw data underlying these datasets are publicly available from Yahoo Finance. Given the dynamic nature of financial data sources, minor differences may arise if the data are downloaded at a different time. However, the structure of the dataset and the feature construction procedures described in the manuscript enable full replication of the analysis. This repository is intended to facilitate transparency and reproducibility of the empirical results. Code used for model training and portfolio construction can be provided by the authors upon reasonable request.