MacroLens Benchmark Released for Contextual Financial AI Reasoning

A new benchmark dataset, MacroLens, has been introduced to address the complexities of financial decision-making for AI models. Traditional time-series evaluation often fails to account for the contextual nature of finance, which requires integrating diverse data types like price history, accounting figures, macroeconomic conditions, and contemporaneous text. MacroLens overcomes these limitations by providing a comprehensive dataset that correctly gates text by publication date, accounts for reporting lags in fundamentals, and manages macroeconomic regime leakage. The benchmark focuses on 4,416 U.S. small- and micro-cap equities from 2021-2026, offering a rich panel of data including prices, 46.8 million XBRL accounting facts, 53 macroeconomic series, nearly 300,000 SEC filings, and over 200,000 news articles. It also features a unique scenario layer with 1,130 automatically detected macroeconomic events. MacroLens includes seven distinct tasks, ranging from contextual forecasting and valuation to scenario-conditioned returns, and has been used to evaluate 19 different AI methods, from naive heuristics to advanced LLM-based time-series models.