TimeSeriesExamAgent: leaderboard

Time-series reasoning on exams that TimeSeriesExamAgent (2026) generates from real data: LLM-written question templates over five datasets (MIT-BIH, PTB-XL and MIMIC-IV waveforms, Yahoo Finance prices, WeatherBench 2), verified by an LLM and filtered for difficulty, with the series shown as plots. Unweighted mean of the five exam accuracies in percent (chance 25); higher is better.

Metric: Average accuracy (%, times 100): the unweighted mean of the five TimeSeriesExamAgent exam accuracies (MIT-BIH, PTB-XL, MIMIC-IV Waveform, YFinance, WeatherBench 2; 857 questions in all): multiple-choice questions generated by TimeSeriesExamAgent from real data (Claude Sonnet 4 templates verified by gpt-4o and filtered so that gpt-4o does not score below gpt-4o-mini), time series given as plots, default API parameters; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-551.5#178
2Gemini 2.5 Pro51#276
3TimeSeriesExamAgent Qwen2.5-VL-Instruct (checkpoint unspecified)45.6
4O3 Mini44.2#546
5Gemma 3 27B (IT)43#1324

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=timeseriesexamagent · How It Works · Data refreshed daily, snapshot 2026-11-21.