Skip to content
kanven

Track record

Tested on the past. Misses included.

We give Kanven only what was known the day before a real policy took effect, then compare its predictions with what actually happened. Every case we run is published here.

first results coming soon

We’re checking each historical outcome against its official source before publishing any score. Unverified numbers never appear here.

Cases being prepared

How we test

  • A hard knowledge cut-off. Kanven sees only evidence published before the day the policy took effect, and economic indicators from before that year. The live web search is switched off.
  • Disguised cases. AI models learn from text written after these events, so they may “remember” what happened. Each case is also run with the country and date removed; the score uses the disguised run, and we publish how often disguising changed the answer.
  • Ranges, not points. We check whether reality fell inside Kanven’s likely range. A good range holds the real value about 80% of the time — a range that is always right is too wide to be useful.
  • A simple benchmark. Kanven must beat the forecast “nothing changes” to earn a positive skill score.
  • Sources for every outcome. Real-world outcomes come from official statistics and are checked by a person before they count.

Limits

Policies rarely happen in isolation — the 2015 blockade overlapped with political unrest, Sri Lanka’s fertiliser ban with an economic crisis. We note confounders on each case. Past accuracy doesn’t guarantee future accuracy; it tells you how much weight to give a result. Read the full methodology.