[P] Random Forest on ~100k Polymarket questions — 80% accuracy (text-only)
A simple text-only model beats complex LLMs at forecasting event outcomes using just question wording.
Deep Dive
An anonymous researcher built 'mystery:rf-v1', a Random Forest model trained on ~90,000 resolved Polymarket questions. Using only text features like TF-IDF and keyword flags, it achieved ~80% accuracy on a 15,000-question test set. The model, now in paper trading, competes with state-of-the-art LLMs as a benchmark, showing that question formulation alone can strongly predict binary market outcomes.
Why It Matters
Demonstrates that simple, interpretable models can rival expensive LLMs for specific forecasting tasks, potentially lowering costs.