Frontier AI models outperform analysts on earnings prediction
4 hours ago
- Frontier AI models using Samaya's finance harness outperform human experts in predicting earnings surprises for the first time.
- The prediction environment provides point-in-time financial data parity with human analysts and tests metrics like prediction error, surprise correlation, and hit rate.
- GPT-6 Astra performed best across most metrics, while Claude Fable 5.1 and Opus 5.5 excelled in surprise correlation.
- Ablation studies show that access to live data is the most important factor, and expert guidance improves reasoning and search depth.
- Top models maintain performance even without consensus by reconstructing it from public information.