We tested 49 causal chains. One worked.
Most causal claims in finance don't survive contact with data.
Yes No Capital was built to test relationships between prediction market events. Academic literature offers dozens of plausible mechanisms — inflation depresses incumbent vote share, oil spikes precede recessions, tariffs raise consumer prices, war depresses markets, rate cuts move equities, credit spreads forecast corporate distress. Forty-nine such chains, each sourced to a peer-reviewed paper.
Each chain becomes a keyword rule: match inflation events with election events; when the cause resolves YES, measure whether the effect resolves YES more often than the category baseline. The measure is straightforward — lift is the ratio, anything above 1.20 at n≥30 pairs earns a PREDICTIVE verdict, anything under 0.85 earns INVERTED, anything near 1.00 means no edge.
After 10,494 resolved pairs, one chain survives: nuclear policy events correlate with nuclear-related markets at lift 1.06, stability 0.98 across time-split walk-forward validation. That is weak. It is also the only one that passes all three gates.
Twenty-four chains show no edge despite 500+ pairs each — inflation→elections (lift 1.02), gdp→elections (1.00), oil→inflation (1.01), tariffs→gdp (0.98). At scale, the "obvious" macro relationships don't discriminate prediction market outcomes at all. Keyword pairing is not causation.
One chain inverted: antitrust events predict markets in the opposite direction academic theory suggested.
Most informative, an earlier version of this test looked much better. With a top-5-per-chain curation layer, two chains came in at lift 1.22 and 1.26 — above the PREDICTIVE threshold. Uncapping the sample collapsed both to baseline. The curation was picking causes with extreme resolution probabilities and inflating the measurement. Selection bias, built into the test.
This is why backtests are easy to publish and hard to trust. Sample-cap an hour earlier, filter by volume, restrict to a favorable window, and any chain can look predictive. The honest version of the scoreboard is rarely what publication incentivizes.
What works, specifically
- Walk-forward validation catches chains that had past edge but lost it. Stability 0.71–0.76 on several near-miss chains flagged them as unstable even when overall lift looked acceptable.
- Uncapped sampling catches chains that only look predictive under curation.
- Directional spread — the difference between P(effect|cause=YES) and P(effect|cause=NO) — catches chains with overall lift but no discriminant power.
What still does not work
- Temporal coupling — pairing cause events with effects that resolve in a specific following window. Prediction market data does not cleanly support this yet.
- Foresight testing — whether cause probability shifts predict effect outcomes before resolution. Requires per-event probability trajectories, infrastructure not yet in place.
These are honest gaps. The scoreboard at yesnocapital.com/track-record is updated hourly as new events resolve. Chains auto-deprecate when they accumulate 100+ pairs at baseline lift — the library is self-cleaning.
One surviving chain is not a product. It is a start. The point isn't that nuclear_markets works; it is that most things don't, and making that visible is the only honest version of this kind of tool.