Method
Once a day, at 04:00 Helsinki time, the pipeline collects the last 24 hours of news from about 55 feeds, Reddit and X, indexes it, refreshes the prices of every question it follows on Polymarket, and hands the bundle to a forecasting agent built on FutureSim, a research simulator for testing whether AI agents can forecast world events from a time-sealed news corpus.
The agent searches that corpus, reads the resolution rules, and submits a probability for each open question. It keeps a memory of its own past calls and their outcomes; each day starts with a fresh context so it cannot simply repeat yesterday.
Agents in the book today: minimalHarness_gpt-55_001, minimalHarness_gpt-56-sol_001, minimalHarness_gpt-56-sol_002. The second one, when present, is an A/B control with the same model and no accumulated memory.
A question counts only if the bot forecast it while it was still open and then watched it settle. Its number and the market’s price on the market’s last open day are scored with the Brier score, the squared distance from what happened. Both are then averaged per event — a ladder of price thresholds is one event, however many rungs it has — because otherwise one lucky call on a ten-rung ladder would look like ten wins.
Both numbers are taken from the same day, so neither side gets to see the answer first. Until 7 September 2026 the record used the bot’s very last forecast instead, which on about a fifth of questions was made after the outcome was already visible in the price; that flattered the bot, and the record has been recomputed without it.
When the bot disagrees with the price by at least 5 percentage points on a question with a real order book, it opens a 100-share paper position on the side it favours, pays Polymarket’s taker fee, and holds to resolution. It refuses disagreements above 25 points (more likely a stale belief than a wrong market), never holds two positions in the same event, and never trades inside ten days of resolution.
Nothing here is traded with real money yet. Three gates have to pass first:
After that, a $25 mechanics test with 5-share orders, and only then anything larger — and only after a Finnish tax ruling on how prediction-market gains are treated.
The analyst view of each day, with every chart and table, is kept at /forecasts/2026-09-07/.
Not financial advice. Paper results are simulated fills at the quoted price and can overstate what a real order would get.