CME proof ledger

Do the signals work yet?

This page is where Convexly proves, over time, whether its Coherence Signals turned into useful settled outcomes. Of 844 resolved signals, 425 have a PnL the model actually computed, leg by leg. Across those, realized net at prescribed size is -$403.0k, with 41.2% of them positive (175 of 425, 95% CI 36.6% to 45.9%). That is a loss, published as-is.

The other 419 are excluded because nothing about them was measured. Their emitted signals carried no per-leg trades, so the resolution model assumed both sides: a total loss of stake on the unfavourable ones, and the emitter's own modeled expectation on the favourable ones. Neither direction is a result, so neither is counted. The code no longer does this; the rows it already wrote are excluded here rather than rewritten.

Four limits on the figure above, all of which cut against reading it as a verdict.

• It is selected on data availability, on whether the emitter recorded a per-leg position, which is not known to be independent of outcome.
• The two cohorts are date-disjoint: every excluded signal was emitted between 2026-04-29 and 2026-05-12, every included one from 2026-05-13 onward. So this is one era of the pipeline, not a sample across its life.
• The included signals are not independent observations. They span 33 distinct markets across 425 signals, so the same event groups are re-emitted many times and any total built on them is pseudo-replicated.
• The pre-registered matched-control test (#294035) has not accumulated enough resolved pairs to say whether these signals beat a control, so nothing here separates the result from the market conditions it was earned in.

One finding the exclusion would otherwise have hidden, so it is stated here rather than deleted with the rows. Among the excluded baskets built on event groups typed mutually exclusive, 143 of 143 legs resolved the same way. Under mutual exclusivity at most one of them can. Those groups were therefore not mutually exclusive, which falsifies the classifier that typed them rather than the arithmetic that priced them. That is a methodology defect in the signal construction, and it is the more useful thing on this page than the dollar figure.