Favorite–longshot bias: the corrected backtest — signature REPLICATES, adverse selection is favourable, generality UNPROVEN

Why this exists

The 2026-07-25 strategy-designs doc claimed Design"survives every robustness check" on decision-point tables whose code was never committed and whose source data is gone, while the committed script's own verdict was DOES NOT REPLICATE. This rebuild replaces both with a reproducible chain and applies every correction the review demanded.

Result— the FLB signature replicates (and the old sign flip was the bug)

Buyer net per contract, event-clustered, taker fees, single-event only:

Both pre-registered gates pass: longshot buyers lose, favourite buyers win — the CEPR DP20631 (Bürgi, Deng & Whelan) signature, reproduced out-of-sample on our own data. Note the low bucket is −1.13¢ here versus +5.56¢ under the old whole-life method: the correction flipped the sign, confirming selection-on-ever-touching was the defect.

Result— the missing decision-point analysis, rebuilt

Selling cheap YES as a maker at a fixed fraction of each market's trade life, held to settlement:

Positive at every decision point except 10–15¢ at 0.25 — i.e. it does not depend on a single observation point, which is exactly the fragility that killed Design(buy-favourites). The overpricing is direct: at 0.75, the 1–5¢ bucket implies 1.8% but settles 0.0%, Wilson CI [0.0, 0.7] — the CI's upper bound sits below the implied probability. **The 1–5¢ t-statistics are flagged deg** (degenerate): every market settled NO, so the CI, not the t, carries the evidence.

Result— adverse selection runs in our favour (the reconciliation)

Takers who buy longshots are more wrong than average — the side we would be filled on is the uninformed one. This reconciles the apparent contradiction with our own maker refutation: that result was about day-ahead weather fading, a different trade, in a band where the fee peaks and the informed flow is forecast-driven. Being the maker is not uniformly bad; it is bad there and good here.

What is NOT established

• Generality is unproven.ofseries are weather. KXCPI contributes 2 events,

KXFEDDECISION 1. Any cross-family claim is unsupported.

• Per-series, it is not uniform:ofpositive, but only KXHIGHLAX (t 2.55), KXHIGHCHI

(t 2.78), KXHIGHDEN (t 2.40) and KXCPI (t 4.74, n=2 events) clear t>2 — and KXHIGHTATL (−0.60¢) and KXHIGHTDC (−0.67¢) are negative.

• Effective n is ~203 events, not 1.2M trades. Weather brackets are ~12 markets per event and

~20 events per series; the trade count is not independent evidence.

• Fills are assumed, not simulated. We count trades that happened and assume we could have been

the resting maker. Queue position and partial fills are unmodelled — the live probe (kalshi_longshot_probe.js) exists to measure exactly that, and remains unarmed.

• Recent period only (~last 10–20 trading days per series). No regime variation.

Cross-family extension (same day, second pass) — and a bias it exposed

Extending beyond weather required discovering the real series universe (10,190 non-parlay series) and probing candidates for settled depth rather than collecting blindly. Adding KXMLBGAME (497 events) produced an apparently spectacular sports result — +6.78¢, t=21.5 — which did not survive checking:

C8 truncation guard (the artifact). The collector capped trades per market. Measured: 88% of MLB markets hit that cap, capturing a median of 1.6 hours of markets that run up to 75 hours. The captured slice was the end of the game, where the losing side genuinely is ~0% — so the "edge" was a sampling artifact, not a market fact. Capped markets are now excluded from the analysis (decision-point fractions are meaningless on a tail slice) and the collector paginates 40× deeper and records per-market completeness.

C9 minimum-n generality gate. The first cross-family verdict read "CROSS-FAMILY" off a family with 3 events (macro). A family now needs ≥30 independent events to count.

Corrected cross-family picture (post-guard):

Verdict: PARTIAL — holds inqualified families. And the sports figure carries its own selection caveat: after excluding 88% of MLB markets as truncated, the survivors are precisely the quiet games, so that t=32.7 should be treated as unproven until a full-depth re-collection runs.

Status

Designmoves from plausible, not internally reproduced → reproduced on a corrected, committed chain: longshot overpricing is robust in weather (200 events, t=3.09) and directionally present in sports (60 events, selection-caveated); macro remains underpowered (3 events). Adverse selection runs in the seller's favour. Generality is PARTIAL, not established. The evidence bar for risking money is unchanged: cross-family n, a fill model, and correlated-tail sizing must land first.

Fill simulation (the last caveat) — the strategy DOES NOT SURVIVE

The earlier passes counted trades that happened and assumed we could have been the resting maker on them. That assumption is the most common way a paper edge dies in production. It is now removed, using free public data: Kalshi's candlesticks endpoint serves 1-minute yes_bid/yes_ask OHLC, which is enough to require that a real buyer existed at our price before we book a fill.

Method (kalshi_flb_fillmodel.js, data via kalshi_collect_candles.js,markets / 417k minute-candles): post a resting offer to sellkalshi_collect_candles.jsYES at the quoted ask; it fills only if a later minute's yes_bid high reaches that price. Unfilled offers earn nothing and are counted.

Verdict: DOES NOT SURVIVE fill simulation. Three findings kill it:

1. **Only the latest decision point is positive-and-significant — and its statistic is

degenerate.** At 0.75 allfills settled NO, so P&L variance collapses and t=7.68 is an artifact of price jitter, exactly the C6 trap. Withfills a true 6% rate predicts ~1.7 winners; observingis unremarkable, not significant.

1. **"Works only at the latest observation point" is the same fragility that disqualified

Design(buy-favourites).** Applying our own standard consistently, it disqualifies this too.

1. Half the offers never fill, and filled ones wait 1–2.5 hours. Capital is committed while

idle, so per-fill edge overstates the business: the net_c_per_offer column is roughly half the per-fill number even before that.

ForecastEx economics were checked as the friendlier venue (flat 1¢/contract embedded in the spread, $0 IBKR commission, no maker/taker split — versus Kalshi charging even on resting fills). Its flat cent removes another cent per fill, which does not rescue any point.

Honesty limit: queue priority, partial fills, and our own market impact are all ignored, and each makes real fills worse. These simulated results are an upper bound — the true strategy performs no better than the table above, and the table is already a failure.

The favourite side — the one candidate still standing (underpowered, not proven)

Prompted to search the literature rather than keep grinding our own data, the academic synthesis (QuantPedia, ~20 studies) ranks the documented edges by retail accessibility: inter-exchange arbitrage needs sub-second execution, intra-exchange arbitrage is largely closed, and only longshot bias remains "still live" with "no special technology needed" — expressed as buying favourites, not selling longshots.

That distinction is decisive against our own week's work: selling longshots is a maker trade (needs someone to lift a resting offer; ~50% never filled), while buying favourites is a taker trade — fills are structurally certain. The property that killed the previous candidate is absent by construction.

Two sampling corrections, both of which changed the answer (kalshi_favorites_liquid.js):

• Real ask, not traded price — a taker pays the ask. Measured gap: +0.60¢.
• Liquid minutes only (volume > 0) — sampling every quoted minute is dominated by stale asks in

untraded markets, and produced spurious significant negatives (50–60¢ t=−2.60, 95–99¢ t=−5.48) that vanish once a real trade is required in the same minute. You cannot lift a quote nobody is honouring.

Result (553 markets, event-clustered, real asks, taker fees, liquid minutes):

Verdict: directionally positive in the favourite bands, NOT statistically significant. The 75–95¢ range nets +1.6 to +2.6¢/contract with t = 0.5–1.4 on ~115 events — the right sign, matching both the literature and our own trades-based pass, but short of proof. At the observed variance, roughly double the events (~240) would be needed to reach t≈2.

This is materially different from the eight refuted strategies: it is underpowered, not refuted — the first candidate to end that way. It stays unarmed until the event count doubles.

External grounding + the structural argument (why the extremes, and why Kalshi ≠ a bookmaker)

The serious counter-argument, found and confronted. datagolf argues the favourite–longshot bias is not a behavioural bias at all: in bookmaker markets it falls out mechanically because bookmakers allocate roughly equal absolute margin (~0.8–1% per side) across all odds (Pinnacle, 27,150 matches 2012–2020). Every participant is rational, returns decline at long odds anyway, and — crucially — all bets stay negative-EV, favourites merely less so. On that account the pattern is real but unexploitable, and the football evidence agrees (favourites −3.64%, outsiders −26.08%: both losers).

Why that critique does not transfer to Kalshi. Kalshi is an exchange, not a bookmaker — no house sets one-sided odds with an embedded margin — and its fee is 7·P·(1−P)/100 cents, a convex curve minimised at the price extremes. Measured in our own data: 0.43¢ in the 90–97¢ band versus 1.74¢ mid-book — a 4× difference. A small gross edge can therefore survive net at the extremes while being eaten alive in the middle. The CEPR study of 300k+ Kalshi contracts reports exactly the outcome the bookmaker model does not predict: high-price contracts yield "small positive returns" net of fees — positive, not merely less negative.

Decomposition of our own edge (kalshi_fee_curve_decomposition.js, liquid minutes, real ask, event-clustered):

Both tails behave as the theory predicts, with opposite signs, and the loser side is strongly significant: buying deep longshots loses −4.82¢ (t=−4.18); buying deep favourites gains +3.01¢ (t=2.29) where the fee is cheapest. The middle is noise.

The caveat applied to ourselves:bands were tested. A single cell at t=2.29 does not clear the Harvey–Liu |t|>3 bar for a new factor claim under multiple testing (Bonferroni-adjusted p≈0.19). What is defensible is the pattern — a monotone-tailed signature predicted a priori by an externally documented bias, with the fee curve explaining where it survives — not that one cell.

Status unchanged: promising, not proven. The strongest single number in the whole programme is still the one that says buying longshots reliably loses (t=−4.18) — and the profitable mirror of that requires being a maker, which the fill simulation already killed.