oddsx status

Pre-registered verdicts (TEST split)

Protocol verdicts as computed; read the SCORECARD and dossiers for caveats.

Dossiers

• drift-to-resolution
• favorite-longshot
• fundamental-arb-exec
• fundamental-arb-short
• fundamental-arb
• maker-economics
• toxic-flow

SCORECARD

Edge SCORECARD — powered re-run (T-0015)

All three edges re-run on the FULL resolved-sports window (build_resolved_bets_full — 29,651,775 bets, 83,019 markets, 2026-04-18 → 06-11) with a leak-free calendar split (calendar_split): TRAIN = Apr18→May26, TEST = a genuine 15-day June OOS era (May27→Jun11). Standard control = the FAVORITE null (the honest, cost-artifact-free control per T-0012), same Polymarket cost model (fee_rate=0.05, spread=0.02), equal-notional sizing, returns capped ≈[−1, +1]. Band/threshold derived on TRAIN only. edge_vs_naive = strategy mean_net − favorite-null mean_net (costs cancel → pure selection skill). Verdict = TEST edge_vs_naive > 0 and TEST mean_net > 0.

All three stay FALSIFIED on the powered window. The favorite null (buy the favorite) is undefeated: no edge beats simply backing the favored side, net of costs.

Notes per edge

• Edge-1 (favorite-longshot): the gross calibration bias is real (~+1% for mid favorites) but ~2–3× smaller than the ~3% taker cost; net-negative in every price bucket. More robust than provisional: the sign no longer flips between all-fills and per-market (both negative), and the provisional "longshot mirage" (+0.28) is gone (−0.0075).
• Edge-2 (drift / CLV): falsified as a tradeable edge — momentum's TEST net is marginally positive (+0.0153) but is entirely a fat-tail artifact (median −0.0070; a single +65× outcome; top-1% carries 19.6% of |PnL|), and it loses to buying the late favorite (+0.0464). Separately, closing-line value is strongly CONFIRMED: late-price Brier 0.022 vs early-price 0.201 (TEST) — the information is real but already priced by entry time.
• Edge-3 (toxic-flow): falsified decisively — the follow-flow return *worsens* monotonically as flow gets *more* one-sided (opposite of the thesis). The provisional "+1.45 edge vs uniform" was a null-is-worse trap; vs the sharp favorite null the edge is a clean −0.55. OP's "wide-spread = retail/safe" is FALSE (capped taker outcome flat across spread quintiles; near-close flow is if anything *more* informed).

Tail-honesty summary (median vs mean; top-1% |PnL| share, TEST)

Verdict

No edge flipped to validated on the powered window — all three remain FALSIFIED, and the falsifications are now backed by a real 15-day OOS era rather than the provisional ~1-day tail. Honest > profitable: the sharp favorite null survives every challenge, closing-line value is confirmed as *information* (not a tradeable edge), and the toxic-flow / wide-spread retail folklore is refuted in the data. Every signal being real-but-sub-taker-cost points away from a taker edge — the maker side (post limit orders, capture the bias as spread, pay no taker fee) is the natural next question.

Reproduce (each pastes its full stdout):

uv run --extra backtest python research/02_favorite_longshot.py   # Edge-1
uv run --extra backtest python research/04_drift.py               # Edge-2
uv run --extra backtest python research/03_toxic_flow.py          # Edge-3

uv run pytest -q → green.

Maker economics (T-0016, added iter-006)

Fee asymmetry CONFIRMED (maker beats taker by +0.26/fill — the fee is the only wedge), and the adverse-selection gradient is real (maker net worsens monotonically with flow one-sidedness). But even at 0 fee, and as an OPTIMISTIC upper bound (assumes every passive fill wins), making is net-negative. Conclusion: no harvestable edge from historical fills — taker OR maker. The one untested candidate is polymm's actual edge = maker WITH FRESH SPORTSBOOK ODDS (de-vig), which needs Track C forward data.

fundamental_arb (G4, pre-registered, Polymarket price-threshold markets)

Pre-registration and full results: dossiers/fundamental-arb.md. Calendar split on decision time: TRAIN 2026-04-10 → 06-30, TEST 2026-07-01 → 08-21. Benchmark = favorite null, same costs. Verdict rule: fewer than 200 TEST bets → INSUFFICIENT_DATA. Realized-vol-only limited test.

fundamental_arb_short (G4b, pre-registered separately, Polymarket price-threshold markets)

Pre-registration and full results: dossiers/fundamental-arb-short.md. Decision 1 hour before end (at least 30 minutes to resolution). Calendar split on decision time: TRAIN 2026-04-10 → 06-30, TEST 2026-07-01 → 08-21. Benchmark = favorite null, same costs. Same verdict rule as G4 (200-bet floor; mean net > 0, edge_vs_naive > 0 and cluster-bootstrap 95% lower bound > 0). Realized-vol-only limited test.

The protocol verdict stands as computed, but post-hoc diagnostics (not part of the pre-registration) point to stale YES prices in thin hourly markets: the TEST 95% lower bound is only +0.0097, the edge sits where the last trade is old, and the freshest-priced bets lose. Not promoted; a follow-up with executable entry prices is required before any paper trading.

fundamental_arb_exec (G4c, pre-registered separately, Polymarket price-threshold markets)

Pre-registration and full results: dossiers/fundamental-arb-exec.md. Same signal, universe and decision time as fundamental_arb_short, but each bet is filled at the first executable trade at or after the decision (within 10 minutes); returns, costs and the favorite null use that fill price. Calendar split on decision time: TRAIN 2026-04-10 → 06-30, TEST 2026-07-01 → 08-21. Same verdict rule (200-bet floor; mean net > 0, edge_vs_naive > 0 and cluster-bootstrap 95% lower bound > 0). Realized-vol-only limited test.

At the executable fill, TEST mean net is −0.2449 and the 95% lower bound −0.3901. Post-hoc diagnostics (not part of the pre-registration): the same 516 filled bets earn +0.179 net at the stale signal price and −0.245 at the fill, and 67% of fills are worse than the signal price. fundamental_arb_short's result was a stale-price artifact. No fundamental_arb variant is promoted or paper-traded.

poly_data lake

Total 27.26 GiB · /home/trbck/workspace/_repos/poly_data/processed/parquet · sizes, rows and timestamps from Parquet footers (no data scans).

Verification

oddsx boil verification — 2026-09-14 08:49

39/39 frozen checks pass.

Goals ladder

Retention log (last 10)