live/: operationalizes the LBS/Yale "skilled ~3%" result against the live data-api. Enumerate recent liquid markets -> top traders -> candidate pool; cache every wallet's resolved bets once in DuckDB (~26k wallets / 12.5M bets, keyed by per-bet resolution time so any cutoff re-scores in seconds); 5-gate skill funnel (n>=15, z>0, BH-FDR, split-half OOS, MM/bot cap); dashboard + daily refresh. Key finding: copying the high-win-rate "favorite-rider" cohort looks +23.6% in-sample but loses -7.4% once selected on pre-June-1 data only (99% -> 68% win rate) — selection bias, reproducing the paper's "lucky winners revert" result on live data. Win rate != edge, again. wide/: bulk subgraph->DuckDB scanner (survivorship-bias-free over all wallets), but the public subgraph is frozen at Jan 2026 -> historical tool only. Large local data (*.duckdb, candidates.json, *_scored.json, history/) gitignored. README + FINDINGS updated with the current logic and the clean result. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
3.3 KiB
wide/ — bulk subgraph edge scanner
Find wallets with a real edge across all of Polymarket (~1.76M traders, 268k conditions) by bulk-ingesting the on-chain subgraph into a local DuckDB and ranking with SQL — instead of per-wallet API calls (which cap out at a few hundred wallets and rate-limit).
Why a local DB, not the API
The Goldsky orderbook subgraph times out on any orderBy over a
non-indexed field (numTrades, scaledProfit, …). You cannot ask it for "top
wallets." The only scalable pattern is: cursor-paginate every row by id (the
indexed key), land it locally, and rank in DuckDB. That constraint is the
architecture.
Why the win rate here is honest
The data-api hides losers: Polymarket only redeems winning shares, so losing
positions sit unredeemed and never enter /closed-positions. Measuring win
rate there reads ~90% when the truth is ~50% (see ../FINDINGS.md). The subgraph
records a marketPosition for every buy regardless of redemption, so the
survivorship bias structurally does not exist in this data.
What "edge" means
A high win rate is not edge — you can hit 90% by only buying 95¢ favorites. Edge is beating the prices you paid:
p = valueBought / quantityBought (entry price, 0..1)
won = the outcome you held paid out
z = (wins − Σp) / sqrt(Σ p(1−p)) standard deviations above odds-implied
z high over enough bets, on a wallet that isn't a market-maker, is the real
signal (the same one ../insider.py computes, here over the entire market).
Pipeline
| step | script | source | notes |
|---|---|---|---|
| 1 | ingest.py conditions |
subgraph | resolution + payoutNumerators → winning outcome |
| 2 | gamma_tokens.py |
Gamma | token_id → (condition, outcome_index); subgraph's outcomeIndex is null |
| 3 | ingest.py accounts |
subgraph | numTrades (market-maker filter), creationTimestamp (freshness) |
| 4 | ingest.py market_positions |
subgraph | the heavy table: entry price + win/loss per bet |
| 5 | score.py |
DuckDB | z, true win rate, profit + FDR + out-of-sample |
All ingests are resumable (per-table id cursor in _cursor), so a long
run can be stopped and restarted. Tables join in edge.sql.
Guardrails (so a 1.76M-wallet scan doesn't just surface luck)
- min-n + market-maker cap. A 300k-trade grinder posts huge z with no
information. Require ≥30 resolved bets and cap
numTrades. - Benjamini–Hochberg FDR. Scan 100k wallets and thousands clear z>3 by
chance (look-elsewhere effect).
score.pyreports how many survive 5% FDR. - Out-of-sample.
score.py --cutoff YYYY-MM-DDselects wallets on bets resolved before the date, then measures the same wallets forward. Edge that is real persists; edge that is curve-fit reverts to z≈0 — which is what every strategy in ../FINDINGS.md did. Do not size up on in-sample z.
Usage
pip install duckdb
python3 ingest.py conditions accounts # small tables
python3 gamma_tokens.py # token→outcome map
python3 ingest.py -p market_positions # heavy (~53M rows); parallel + resumable
python3 score.py --min-n 15 --max-trades 5000 --top 40
python3 score.py --cutoff 2026-04-30 --min-n 15 # in-sample vs forward