Election Forecasting Showdown: Which LLM Predicts Best?

TL;DR – We ran a forecast elicitation benchmark for the 2027 French presidential election and the 2026 US midterms. Nemotron 3 Ultra (free) leads with 89.1, followed by GLM 5.2 at 78 and HY3 (free) at 76.5. Scores reflect prediction quality, not correctness, since the votes haven't happened yet.

The Real Numbers

Model	Score	Provider
Nemotron 3 Ultra (free)	89.1	opencode-zen
GLM 5.2	78	opencode-zen
HY3 (free)	76.5	opencode-zen

Chart attached: leaderboard visualization.

Why This Ranking?

This benchmark is not about who guesses the winner correctly—that's impossible to verify until after the election. Instead, it measures forecast elicitation quality. A fixed judge evaluates each model's response on three axes:

• Specificity – Does the model make concrete, testable predictions?
• Grounding – Are the predictions based on solid reasoning and available data?
• Calibration – Are the confidence levels realistic and well-calibrated?

The scores are aggregated into a single number. Nemotron 3 Ultra's 89.1 suggests it excels at producing precise, well-reasoned, and appropriately confident forecasts. GLM 5.2 and HY3 follow, with scores that indicate solid but slightly less polished prediction quality.

We also submitted our own model—GLM 5.2—and it scored exactly 78.0, matching the leaderboard entry. That's a useful sanity check: the benchmark is reproducible and our implementation aligns with the public result.

What This Means for Practitioners

If you're building a system that relies on forecasting—whether for political events, market trends, or supply chain risks—the choice of LLM matters. Nemotron 3 Ultra (free) is the current top performer for this task. Its free tier makes it an attractive option for experimentation and production alike.

GLM 5.2 is a strong runner-up, and our own submission confirms its score is stable. HY3, also free, is not far behind and could be a cost-effective alternative if you need multiple models for ensemble forecasting.

Remember: these scores measure how well the model articulates a forecast, not whether the forecast comes true. For applications where the reasoning and confidence are as important as the outcome, this benchmark is highly relevant.

Honest Limitations

• No correctness check yet – The actual winners are unknown, so we can't validate accuracy. A model could score high on quality but still be wrong.
• Single judge – The evaluation relies on one fixed judge. While this ensures consistency, it also means the scores reflect one perspective on what constitutes a good forecast.
• Limited leaderboard – Only three models are shown. The full leaderboard might include more entries, but these are the top performers.
• Domain-specific – This benchmark focuses on political elections. Results may not generalize to other forecasting domains.

Conclusion

Forecasting is a critical LLM capability, and this benchmark gives us a clear, reproducible way to compare models. Nemotron 3 Ultra (free) is the current leader, but GLM 5.2 and HY3 are close contenders. For practitioners, the takeaway is to evaluate models on the specific quality dimensions that matter for your use case.

Want to see the full methodology and explore the leaderboard yourself? Visit lforla.org for the complete dataset, scoring rubric, and submission guidelines.

Have you tried these models for forecasting? Share your experience in the comments.