Research · study
Do divergences converge?
The Scopes Desk
The first Scopes study, computed from the intraday record: when the exchange price and the options-implied “Scopes Fair Value” disagree by more than 5 points about 3 hours before a contract closes, does the gap converge by close, and when it does, is it the market moving to the fair value, or the fair value drifting to the market? The record is young; the numbers below fill in as the collector banks more sessions, and nothing is asserted before the sample supports it.
The current read
Across 2029 contracts over 32 sessions, 69 showed a divergence over 5 points about 3 hours before close. 31.9% converged to within 2 points by close [95% CI 21.6%, 43.2%, clustered by session].
Among the converged cases, the market (Kalshi) closed on average 2.8% of the gap, the rest was the fair value moving toward the market.
How it's measured
One trigger per contract: the snapshot nearest the chosen time-to-close, taken only when the divergence there exceeds the threshold; the outcome is read at the contract's last recorded snapshot. Convergence alone isn't the whole story, so each converged case is split by which side did the moving: the market correcting toward the fair value, or the fair value drifting to the market.
The interval is a session-level cluster bootstrap: whole trading sessions are resampled together, because snapshots within a day and overlapping bands on the same underlying move together, so the honest sample size is the number of sessions, not the thousands of snapshots.
Two things this page does not yet claim, by design: that the convergence is more than the mechanical regression toward the mean of a noisy measurement (that needs a permutation null), and that any single threshold is special (the finding only becomes citable once it's pre-registered or confirmed on held-out sessions). The record has to earn those claims.
Run it yourself
Choose the divergence threshold, the time-to-close, and the convergence target; the session-clustered estimate recomputes. Slices below 20 sessions report counts only, no verdict. It runs on the trailing 30 days for free, a paid plan runs it over the full archive.
Running the study…
More from the record
- Frontier models vs the market: 2026 Senate controlA pre-registered contest: each LLM forecasts every contested 2026 Senate race and who controls the Senate, with our poll composite as context, scored by Brier against the results and the market. Scopes keeps the record.
- Does context make a model commit?A within-model A/B on the weather desk: given only two probabilities, GPT and Grok abstain on nearly every read; given a public-domain NWS forecast, they commit. The effect is on whether a model answers, not on accuracy.
- Frontier models vs the market: the 2026 MLB postseasonA pre-registered contest: each LLM forecasts every postseason series with public context (records, rotation, rest, injuries), scored by Brier against the outcomes and the market. Scopes keeps the record.
- The MLB binding artifact: what we caught and removedA fair-value game-binding bug inflated MLB divergences and manufactured ~226 flags before Aug 24. They predate the fix and are withheld from scoring; nothing deleted.
- When our own benchmark lost to the marketThe two statistical desks were losing to the market. The cause was tail overconfidence, not timing and not stale data, and what we changed.
- In-play contamination: what changed and whyWhy flags minted while a game is in play or a contract is near settlement are excluded from the scorecard.
- Prediction markets and polls: five pre-registered claimsFive claims registered before the 2026 midterms resolve, to be scored mechanically once the results are in.