benroodman/HillStreetSample

HillStreet是一个大规模、纵向的数据集和多模态动态图，用于形式化国会山与华尔街的交集。它涵盖了13.5年的强制性STOCK法案披露数据（2012年7月至2025年12月），将国会交易生态系统统一到一个单一的、适合机器学习的环境中。数据集代表了1,137位立法者与6,825家公司之间的关系，通过将国会交易建模为动态二分图，使研究人员能够将交易信号验证视为边分类任务。节点包括立法者（按会议特定）和上市公司；目标边为个股交易；结构边包括游说记录、竞选财务（PAC/527）捐款以及地理/工业-选区对齐。数据集分为预构建的图对象（用于深度学习）和通过Hugging Face配置访问的关系型表格数据库。图对象包括年度PyTorch Geometric Temporal文件，节点特征包括滚动的DW-NOMINATE分数、SEC财务事实和人口普查区就业统计数据。关系表包括处理后的边表（如游说、竞选财务、地理工业事件）和原始源表（如竞选财务、立法者数据、公司行业数据）。特征工程使用有符号对数缩放来稳定图训练中的方差。数据集旨在用于交易信号验证和图表示学习，但仅用于研究目的，不适用于内幕交易的法律判定或实时自动化交易。HillStreet is a large-scale, longitudinal dataset and multimodal dynamic graph formalizing the intersection of Capitol Hill and Wall Street. It spans 13.5 years of mandatory STOCK Act disclosures (July 2012–December 2025), unifying the congressional trading ecosystem into a single, machine-learning-ready framework. The dataset represents the relationship between 1,137 legislators and 6,825 companies. By framing congressional trading as a dynamic bipartite graph, HillStreet allows researchers to treat trade signal validation as an edge classification task. Nodes include legislators (session-specific) and publicly traded companies; target edges are individual stock trades; structural edges include lobbying records, campaign finance (PAC/527) contributions, and geographical/industrial-constituency alignments. The dataset is divided into pre-built graph objects for deep learning and a relational tabular database accessed via Hugging Face configurations. Graph objects include annual PyTorch Geometric Temporal files, with node features such as rolling DW-NOMINATE scores, SEC fiscal facts, and Census district employment statistics. Relational tables include processed edge tables (e.g., lobbying, campaign finance, geographical industry events) and raw source tables (e.g., campaign finance, legislator data, corporate industry data). Feature engineering uses signed log-scaling to stabilize variance in graph training. The dataset is intended for trade signal validation and graph representation learning, but is for research purposes only, not for legal determinations of insider trading or real-time automated trading.