Election Prediction Benchmarks: Which AI Model Forecasts Best?

TL;DR – We submitted our own model to a new benchmark for election predictions. Our GLM 5.2 scored 78.0, placing second behind Nemotron 3 Ultra (free) at 89.1. The benchmark measures prediction quality, not correctness, using a fixed judge. Check out the full results at lforla.org.

The Real Numbers

The benchmark, called Election Predictions (FR 2027 · US 2026), evaluates how well models forecast the winners of the 2027 French presidential election and the 2026 US midterms. Responses are public verbatim, and a fixed judge scores prediction quality on specificity, grounding, and calibration. Since correctness is only verifiable after the vote, these scores reflect how well models articulate forecasts, not whether they are right.

Here are the top three scores from the leaderboard:

Model	Score	Provider
Nemotron 3 Ultra (free)	89.1	opencode-zen
GLM 5.2	78.0	opencode-zen
HY3 (free)	76.5	opencode-zen

We submitted GLM 5.2 ourselves, and it scored 78.0, matching the second-place entry exactly.

What This Benchmark Measures

The benchmark is designed to test a model's ability to make specific, grounded, and calibrated predictions about political events. Specificity means the model gives concrete outcomes rather than vague statements. Grounding means the prediction is based on relevant evidence or reasoning. Calibration means the model's confidence levels align with actual probabilities. A fixed judge model evaluates each response against these criteria, producing a score from 0 to 100.

Because the elections haven't happened yet, the benchmark cannot measure actual accuracy. Instead, it measures how well models can articulate forecasts that are logically sound and well-reasoned. This is a useful proxy for real-world forecasting ability, but it's not the same as being right.

Why the Ranking Looks This Way

Nemotron 3 Ultra (free) leads by a significant margin – 11.1 points over GLM 5.2. This suggests that free models can be highly competitive in this domain. The gap between GLM 5.2 and HY3 (free) is small, only 1.5 points, indicating a tight race for second place.

The scores likely reflect differences in how each model handles uncertainty and expresses predictions. Nemotron 3 Ultra may have been optimized for reasoning or calibration, giving it an edge. GLM 5.2 and HY3 are close, suggesting similar capabilities. However, we don't have access to the individual responses or the judge's detailed scoring, so we can only speculate on the exact reasons.

What This Means for Practitioners

If you're building applications that require election forecasting or similar prediction tasks, these scores offer a quick comparison. The leaderboard shows that free models can outperform paid ones – Nemotron 3 Ultra is free and tops the chart. This is valuable for budget-conscious teams.

However, keep in mind that this benchmark is narrow. It covers only two elections, and the judge is a single model. The scores are a starting point, not a definitive ranking. You should test models on your own use cases and evaluate them on the specific outcomes you care about.

Honest Limitations

• Limited scope: The benchmark only covers two elections, so it doesn't generalize to all forecasting tasks.
• Single judge: The judge model may have biases or blind spots. A different judge could produce different scores.
• No correctness yet: The actual election outcomes are unknown, so we can't validate the predictions.
• Provider concentration: All top models come from opencode-zen, which limits diversity. Other providers might perform differently.

These limitations mean the scores should be interpreted with caution. They are a snapshot of prediction quality, not a guarantee of future performance.

Conclusion

We're excited to see a public benchmark for election predictions. It's a challenging task that requires nuanced reasoning and calibration. Our GLM 5.2 performed well, but there's room for improvement. We encourage you to explore the full benchmark at lforla.org, where you can see the complete leaderboard, submit your own model, and contribute to a more transparent evaluation ecosystem.

CTA: Visit lforla.org to dive into the data and join the benchmark.