7,426 forecasts timestamped in Bitcoin (OpenTimestamps, open data, market won)

[neuportal]

Hi all,

Since September we have been running an open experiment: our statistical models make probability calls on Polymarket questions, and every call is stored next to the market price at that moment. Each daily file is then timestamped into Bitcoin with OpenTimestamps, about two minutes after the calls are made. Nobody, including us, can quietly edit a forecast after seeing the outcome.

The result is not flattering to us: the market price beat our models. Brier score 0.161 for the market vs 0.194 for us over 7,048 resolved calls. Every blend of model and price scored worse than the price alone.

Why post it here: it is a working use of Bitcoin as a public clock. 1,697 proof files, each anchored in a block, and a script that checks them all.

How to check without trusting us:

• Proofs and verify script: https://github.com/neuportalai-ui/experiment-proofs
• Pick any file, drop the .ots and then the .json on opentimestamps.org, and see which block it is anchored in.
• Or clone the repo and run ./verify.sh (needs pip install opentimestamps-client). It compares every file's sha256 with its proof; ./verify.sh --node also runs ots verify against your own node.
• Full data (CC BY 4.0): https://doi.org/10.5281/zenodo.24051884

A mistake we left in the record: calls from 14-26 September were stamped late, in one batch on 26 September, after most of them had already resolved. They are flagged in the data (sealed_before_resolution = false) and excluded from the scores. Since then each daily run is stamped minutes after it is made. That is the lesson for anyone keeping a public track record: the timing of the proof is everything.

Write-up: https://neuportal.ai/blog/7426-forecasts-where-the-market-beat-our-models

No token, no signals, nothing for sale in this thread. Questions and criticism of the method welcome.