Edge-1 dossier — Favorite–Longshot bias (take the favorite / short the longshot)

Verdict: FALSIFIED (out-of-sample, net of costs, vs the favorite null). Ticket: T-0015 (powered re-run of T-0009). Script: research/02_favorite_longshot.py. Date: 2026-07-24.

> History: provisional FALSIFIED on the Apr 18–29 window (~485k bets, > TEST ≈ 1 day; T-0009) → powered FALSIFIED on the FULL 53-day window > (29.6M bets, real 15-day June OOS; T-0015). The powered run confirms and > strengthens the falsification: the sign no longer flips across cuts, and the > provisional "adversarial longshot" mirage is gone.

> Honest > profitable. The gross calibration signal is real but ~2–4× too small > to survive Polymarket taker costs; net of costs it loses in every price band, > out of sample, and it does not beat simply buying every favorite.

---

Hypothesis

From the iter-002 calibration diagnostic: Polymarket over-prices mid-priced longshots in sports — tokens in a mid price band win *less* often than their price implies. If the longshot is overpriced, its complementary favorite is under-priced, so systematically taking the favorite side (mechanically: buying the favorite outcome, which is the same position as shorting the longshot) should carry a positive edge OOS, net of costs.

Method (powered — T-0015)

Universe: build_resolved_bets_full(use_cache=True) → 29,651,775 resolved sports bets across 83,019 markets, span 2026-04-18 → 06-11 (53 days). One row per real fill on a resolved sports market, gross realized PnL.
Split: calendar_split (cut on calendar date, boundary-safe, leak-free) — TRAIN = Apr18→May26 (17.3M selected fills), TEST = May27→Jun11, a genuine 15-day June OOS era (8.8M selected fills). Not the ~1-day tail a row-fraction cut produced provisionally.
Position transform: every fill re-expressed as our BUY of the favorite outcome (side priced ≥ 0.5) at q = max(p, 1−p). This is the mechanically-real "short the longshot," and it bounds every bet's return to ≈[−1, +1] — the tail-cap the ticket asks for, for free.
Band derived FROM TRAIN ONLY: bucket favorite price q in 0.05 steps on TRAIN; band = contiguous span of buckets whose TRAIN gross favorite-buy return is > 0. TEST never read during derivation. Derived band: q ∈ [0.50, 0.95).
Costs: polymarket_costs, sports defaults fee_rate=0.05, spread=0.02.
Null (standard control): the FAVORITE null — buy *every* favorite. On this favorite-only frame the favorite null coincides exactly with the uniform null, so "beats naive" == "band-selection beats buying every favorite." Verdict = TEST edge_vs_naive > 0 and TEST mean_net > 0 (honest profitability).
vbt.pro 2026.4.7 simulates the TEST book (hourly equal-weight) for Sharpe + total return + cost-adjusted equity (gate L1.2).

TRAIN calibration scan (favorite-price buckets, powered)

The signal is real but *small and cost-swamped*: gross edge is ~+1% in the mid band, but taker cost at those prices is ~3%, so mean_net is negative in every bucket:

favorite price	N	win_rate	mean_gross	mean_net
0.50–0.55	3,101,799	0.530	+0.0143	−0.0287
0.55–0.60	2,655,286	0.577	+0.0112	−0.0278
0.60–0.65	2,543,768	0.608	−0.0214	−0.0565
0.65–0.70	2,067,306	0.673	+0.0030	−0.0284
0.70–0.75	1,711,005	0.729	+0.0132	−0.0147
0.75–0.80	1,571,871	0.778	+0.0092	−0.0152
0.80–0.85	1,335,907	0.800	−0.0257	−0.0468
0.85–0.90	1,154,256	0.882	+0.0123	−0.0056
0.90–0.95	1,129,432	0.925	+0.0055	−0.0094
0.95–1.00	2,425,620	0.987	−0.0007	−0.0115

TRAIN vs TEST — the verdict table

Edge-1: take the favorite in band [0.50, 0.95) vs buy-every-favorite (favorite null)

split	N	gross	net	sharpe	naive_net	edge_vs_naive
TRAIN	17,270,630	+0.0028	−0.0286	−0.0394	−0.0265	−0.0021
TEST	8,844,141	−0.0114	−0.0431	−0.0588	−0.0399	−0.0032

TRAIN edge is already negative (−0.0021) and TEST stays negative (−0.0032); TEST net return is deeply negative (−0.043). Band-selection does not beat buying every favorite, and the strategy is not profitable after costs → FALSIFIED. Unlike the provisional run, there is no TRAIN→TEST sign reversal — the falsification is stable across the real OOS era.

Supporting / adversarial runs (context, not the verdict)

Adversarial — buy the longshot in the same band: TEST edge_vs_naive −0.0075 (vs uniform null). The provisional run's apparent "+0.28 longshot edge" was a single-day regime artifact; on the powered window it is gone — buying the longshot does not work either.
Robustness — one position per market (de-clustered fills): TRAIN edge_vs_naive −0.0146, TEST −0.0095 (net −0.0386, N=14,853). Same negative sign as the all-fills run on both splits — the provisional sign-instability between fill-weighting and market-weighting has resolved.

Tail-variance honesty

unit	TEST N	median_net	mean_net	min	max	top-1% \|PnL\| share
all-fills	8,844,141	+0.2275	−0.0431	−1.045	+0.955	0.016
per-market	14,853	+0.2275	−0.0386	−1.045	+0.955	0.016

Median (+0.23) vs mean (−0.04) is the whole story of favorites: they win *often* but pay *little*, and the occasional −100% loss plus costs dominate the mean. Net expectation is negative.
Returns are bounded to ≈[−1, +1] by the favorite-buy construction; the top-1% of bets carry only 1.6% of absolute PnL — no handful of bets fabricates (or could hide) the result. Sizing = equal-notional per bet.

vbt.pro simulation (gate L1.2)

vbt.pro 2026.4.7, hourly equal-weight book over the TEST window (376 bars across the ~15-day June era, cost-adjusted): total_return −100.0%, Sharpe(1h-ann) −32.2, max_drawdown −100.0%. Unambiguously negative and consistent with the harness verdict (the compounding book is driven to zero by a persistent ~−4%/bar mean net). Pandas equity curve emitted alongside as fallback.

Verdict

FALSIFIED (powered). "Take the favorite / short the mid-priced longshot" does not beat buying every favorite out of sample and is not profitable net of costs. The gross favorite-longshot mispricing exists (~+1% for mid favorites) but is smaller than Polymarket taker costs (~3%). On the powered 15-day OOS the result is now robust: negative on TRAIN and TEST, negative under both fill-weighting and market-weighting, and the provisional longshot mirage does not reproduce.

Caveats

Cost sensitivity is the whole story. The gross edge is real but ~2–4× smaller than the ~3% taker cost. A materially cheaper execution path (maker rebates, tighter spreads) is the only thing that could revive this — a cost sensitivity sweep is the natural follow-up before fully closing the book.
Fill-clustering: all-fills weights high-volume markets heavily; the de-clustered per-market run now agrees in sign (both negative), removing the provisional concern. by="market" split remains the strictest unit.
Return convention: the favorite-buy transform uses the standard BUY payoff (payoff/q − 1). Both strategy and null pay the identical cost model, so edge_vs_naive is apples-to-apples.
Single 53-day macro window. Apr–Jun 2026 is still one macro regime; a multi-quarter replication would harden the verdict further, but the 15-day OOS era already gives real (not ~1-day) power.

Reproduce

uv run --extra backtest python research/02_favorite_longshot.py

uv run pytest -q → green.