Election Forecasting: How LFORLA Benchmarks Models Before the Vote

TL;DR: Election forecasting is hard to evaluate before votes are cast. LFORLA's Election Predictions benchmark forces models to commit to concrete scenarios for the 2027 French presidential election and the 2026 US midterms, then uses a fixed judge to score specificity, grounding, and calibration.

How Election Forecasting Is Scored

The benchmark is built around forecast elicitation. Instead of asking a model for a vague opinion, LFORLA asks it to predict the winners of two real future elections: France 2027 and the US 2026 midterms. The model must commit to concrete scenarios. That commitment is what makes the evaluation meaningful.

Every response is public verbatim. That means you can inspect exactly what each model said, not just a score. A fixed judge model then evaluates each prediction on three axes:

• Specificity – Did the model make a clear, falsifiable prediction, or did it hedge?
• Grounding – Did the prediction use relevant political context, polling, and institutional constraints?
• Calibration – Did the model express appropriate confidence given what is knowable now?

Because correctness is only verifiable after the vote, the judge cannot check whether the prediction is "right" yet. Instead, it scores the quality of the forecast as a forecast. This is a key distinction: the benchmark measures how well models elicit and structure uncertainty, not whether they happen to guess the eventual winner.

The Real Numbers

LFORLA publishes a leaderboard with scores on a 0–100 scale. Higher is better. Here are the current top results:

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

LFORLA also submitted its own model to this benchmark. That model, GLM 5.2, scored 78.0, matching the public leaderboard entry.

What Election Forecasting Scores Mean for Practitioners

The spread between first and second place is 11.1 points. That is a meaningful gap in forecast quality, not a rounding error. Nemotron 3 Ultra is not just faster or cheaper; it produced predictions that the fixed judge found more specific, better grounded, and better calibrated.

For practitioners, the practical takeaway is simple: if you need a model to produce defensible forecasts for political scenarios, the top of this leaderboard is a strong default. The free tier of Nemotron 3 Ultra is particularly interesting because the highest score comes without a paid plan.

GLM 5.2 sits in the middle. It is competitive, but the judge found its forecasts less precise than the leader. HY3 trails by a smaller margin, 1.5 points behind GLM 5.2, so the bottom two are close.

How to Read a Leaderboard

• Check whether the benchmark uses a fixed judge or human raters. Fixed judges are more reproducible.
• Look for public verbatim responses. You should be able to read the actual predictions, not just scores.
• Understand the scoring axes. Specificity, grounding, and calibration are not the same as accuracy.
• Compare models within the same provider and prompt setup. Cross-benchmark comparisons are noisy.
• Treat scores as relative quality signals, not absolute truth, especially before ground truth exists.

Honest Limitations

The biggest limitation is built into the design: the votes have not happened yet. A model can score high on forecast quality and still be wrong. The judge rewards clear, grounded, calibrated reasoning, but it cannot reward being factually correct about the future.

The leaderboard also depends on the judge model. If the judge has biases, those biases are baked into every score. And the benchmark covers only two elections, both in France and the US. That narrow scope limits how much you can generalize to other forecasting tasks.

Finally, the scores are point-in-time. Models are updated, prompts change, and new submissions appear. A leaderboard snapshot is useful, but it is not a permanent verdict.

Conclusion

Election forecasting is one of the hardest tasks you can give a language model. LFORLA's benchmark makes the evaluation transparent by forcing models to commit to concrete scenarios and scoring the quality of that commitment. The current leaderboard gives practitioners a clear starting point: Nemotron 3 Ultra leads, GLM 5.2 is a solid middle option, and HY3 is close behind.

If you are choosing a model for forecasting or any high-stakes prediction task, start with the LFORLA leaderboard. Read the verbatim responses, check the scoring axes, and decide for yourself.

Explore the full benchmark at https://lforla.org.