Strategy designs from the literature, tested on our own settled data — one survives, two die

Where the designs came from

Web + corpus search for documented prediction-market edges surfaced one directly relevant academic result: Bürgi & Whelan, "Makers and Takers: The Economics of the Kalshi Prediction Market" (CEPR DP20631 / GWU 2026-001), on 300,000+ contracts. Two claims, both testable on our independent sample:

1. Both sides lose. "Takers lose almost 32% on average, while makers have an average loss of

about 10%."

1. Favorite–longshot bias. "Low-price contracts win far less often than required to break

even, while high-price contracts win more often and yield small positive returns."

Design— passive liquidity provision (be the maker): REFUTED, and now corroborated

Already measured and killed earlier today:ofout-of-sample weather cities positive, pooled −1.61¢/contract, t = −3.92, Walk-Forward Efficiency −1.41 against a 0.50 floor.

Bürgi & Whelan independently confirm it: makers lose ~10% on 300k contracts. Our refutation was not a fluke of our sample — it is the published result. This design is dead, twice over.

Design— buy favorites: REJECTED on decision-point sensitivity

The paper's "high-price contracts yield small positive returns" looked like it replicated: at a decision point 75% through a market's trade life, buying at 85–95¢ returned +4.83¢/contract, t = 2.33. But moving the decision point destroys it:

The sign flips three times. An edge that exists only at the latest observation point is a settlement echo — by then the market is largely resolved and the "forecast" is nearly free information. Not tradeable. Rejected.

Design— fade the longshot: SURVIVES every robustness check so far

The other half of the FLB is the robust half. Buying cheap YES (1–15¢) loses at every decision point tested:

Contracts priced around 5¢ settle YES roughly 1% of the time — a ~5× overpricing. That is the classic longshot bias, it is stable across the market's whole life (not a settlement artifact), and it matches the paper's finding that sub-10¢ buyers lose over 60% of their money.

The trade: in markets where YES is cheap, take the other side (sell YES / buy NO). Fees favour it — Kalshi's fee is 0.07·C·P·(1−P), near-zero at the extremes (0.63¢ at 90¢), so the cost that destroys mid-range strategies is smallest exactly here.

Why this is still NOT a green light

The risk shape is brutal and the expectancy hides it. Selling a 5¢ longshot wins 5¢ about 99% of the time and loses 95¢ about 1% of the time. That is picking up pennies in front of a steamroller: positive expectancy, catastrophic negative skew, and a Kelly fraction so small that realistic sizing makes the dollar return trivial. A single 1-in-100 event erases ~19 wins.

Unresolved before this could ever be traded:

• Fill realism. Can you actually get filled selling at 5¢ in size, or is the book thin exactly

where the edge is? Unmeasured.

• Base-rate confound. Our sample is strike-ladder heavy (weather, crypto, spreads) where most

strikes settle NO by construction. The edge must be shown to exceed what naive "always sell YES" earns — not yet isolated.

• Tail sizing. Requires an explicit Kelly/drawdown study before any size is chosen.
• Correlated tails. Ladder strikes on one event resolve together, so "1% of the time" losses

arrive in clusters, not independently.

The methodology error worth recording

The first version of this test found the exact opposite result — longshots +76%, favorites −16.7%, Spearman −0.857 — and it was wrong. The bug was definitional: it bucketed every trade across a market's entire life, so a contract that touched 90¢ mid-life and then collapsed counted as a "favorite." Diagnostic: markets that traded ≥85¢ and settled NO had a median price range of 93¢ and a median final traded price of 1.0¢. Those were peaks of dying markets, not favorites at any decision point.

The tell was that the result contradicted a 300k-contract published paper. When your analysis contradicts established work, the prior is that you have a bug — the fix was one decision per market at a fixed point, and it reversed the entire conclusion.

Status

No design here is cleared for capital. The one survivor is worth a fill-realism study and a Kelly/tail analysis next — not an order.

Independent review addendum (2026-07-25, #2954) — Designstatus corrected

The committed artifact contradicts the Designverdict above: the reproducible script (kalshi_favorite_longshot_replication.py) emits VERDICT: DOES NOT REPLICATE — F1/F2/F4 all false, Spearman(net, price) = −0.857 (the wrong sign for FLB), and the 1–15¢ bucket is +5.56¢/contract on the whole-life conditioning. The decision-point tables that Design 2's rejection and Design 3's survival both rest on have no committed code and cannot be regenerated.

Both tables can be honestly described at once: the whole-life bucket conditions on ever traded cheap (selection-on-ever-touching — eventual winners pass through cheap prices, inflating the low-bucket settle rate to 10.3% vs ~1% for is-cheap-now), so it is biased for longshots and is not tradeable evidence; the decision-point conditioning is the decision-relevant one — and is the missing artifact.

Corrected status of Design 3: plausible (CEPR DP20631 externally + the 2026-07-23 census direction), NOT internally reproduced. Restoration requirements are tracked in #2954: commit the decision-point code, cluster by EVENT not market, account the trade as the SELLER (net = −buyer − 2·fee) reconciled against the maker refutation, and break low buckets out by category. No sizing, shadow, or Act-path discussion before that lands.