Financial TimeSeries Reasoning Benchmarks at Scale

We introduce TimeSeriesExamAgent, a scalable and domain-agnostic framework for automatically generating and validating time series reasoning benchmarks. Existing benchmarks lack scalability, are limited to a few specific domains, while building them remains labor intensive. Automated solutions for benchmark creation have been proposed, but these typically rely on a single-step generation process without verification, leading to lower-quality exams. Our framework addresses these limitations by enabling stakeholders—such as financial institutions with highly confidential data—to easily construct high-quality, domain-specific benchmarks from their own private datasets. A domain expert provides a dataset, a natural language description, and a simple data-loading method. The agent then orchestrates the generation pipeline, including creating question templates, robustness verification from multiple perspectives, and iterative refinement. We demonstrate the framework on financial datasets and evaluate multiple state-of-the-art language models on the generated benchmarks. Empirically, we demonstrate that the framework produces domain-agnostic benchmarks whose diversity matches human-generated counterparts, and our evaluation of several Large Language Models shows that accuracy remains limited, underscoring open challenges in time-series reasoning.