IOV LABSIOV LABS has built a forecasting model for the 2026 Korean local elections (June 3). This post is about the method, with a full pre-registered write-up: the forecast is committed before the result, and after polls close we grade every prediction against reality together with an honest scorecard of how close it was.
How it works
The model is two efforts in one. First, an AI persona electorate: thousands of synthetic voters, grounded in public demographics, are asked across several language models how they would vote. Second, a 538-style poll + fundamentals model: published polls are de-biased (notably for the large gap between phone and automated surveys), blended with each region's structural lean, and run through a correlated Monte-Carlo simulation for win probabilities and vote-share intervals.
The one honest finding we can share now
The AI-persona half did not beat the polls even after calibration it leaned toward the incumbent while the polls pointed the other way. The lesson is the opposite of hype: large language models, asked to role-play voters, add noise and bias rather than signal. Polls remain the thing to trust; the model's real job is to de-bias them and quantify uncertainty honestly. We keep the persona experiment in the open precisely because it failed.
A forecast is only worth something if you publish it before the result and score it after. That is the whole point.
What happens next
The full forecast predicted vote share, vote counts, win probability and 90% intervals for all 16 races is committed and written up as a pre-registered paper. After the polls close at 18:00 on June 3, we grade it: vote-share error, winner accuracy, and a Brier score for the probabilities. Then you can judge the model the only fair way against reality. Stay tuned.