winning-wallet-finder/live/README.md

live/ — find & track the genuinely-skilled ~3%

Finds the small fraction of Polymarket wallets with a real, repeatable edge — the ~3% the LBS/Yale study identifies — from the live data-api, caches everything locally, and tracks them forward. This is the going-forward system; the frozen-subgraph bulk approach lives in ../wide/.

Why this and not win rate

Win rate is survivorship-biased and decoupled from edge (see ../FINDINGS.md). A wallet is "skilled" only if it beats the prices it paid and that edge persists out-of-sample. We reproduced the research's own finding on live June data: favorite-rider wallets that looked +23.6% in-sample lost −7.4% once selected without look-ahead (see "The clean test" below).

The 5-gate funnel (skill.py)

A wallet counts as skilled only if it clears all five:

1. n ≥ 15 resolved bets (assessability; the paper's skilled avg ~79).
2. z = (wins − Σp)/√Σp(1−p) clearly > 0 — wins above what entry odds implied. This is the closed form of the paper's "randomize direction 10k×" benchmark.
3. Benjamini–Hochberg FDR @ 5% — at scale, thousands clear z>3 by chance.
4. Split-half out-of-sample — skill in the earlier half persists in the recent half (z_oos > 0). The gate that separates the real 3% from the lucky.
5. MM/bot cap (n ≤ 2500) — a thousands-of-bets grinder isn't info-edge.

Win rate is never a gate — only displayed. Wallets are tagged value (beats underdog/longshot prices — the copyable alpha), balanced, or favorite (high win% riding near-certain favorites — real but thin/uncopyable).

Pipeline

step	script	what
enumerate	enumerate.py [days]	recent liquid markets (Gamma end_date_min) → top traders → candidate pool (candidates.json), accumulates across runs
cache	cache.py / collect.py	pull each wallet's resolved bets once into cache.duckdb (24s → 0.003s on re-read). Stores res_t per bet, so any date cutoff reads the same cache
score	skill.py [N]	the 5-gate funnel over cached candidates → watch_skilled.json (webhook-compatible)
sharps	conviction_scan.py + validate_timing.py	conviction-profile scan → copy-positive-holder selection → watch_sharps.json (see "The repeatable find")
portfolio	portfolio.py	$1k paper book off the cache → portfolio.json (see "Paper portfolio")
dashboard	dashboard.py	self-contained dashboard.html — sortable, archetype-tagged, live recent-trade lookup
backtest	backtest_june.py [arch]	copy an archetype's June-1+ entries, $1000, no lag → P&L
clean test	clean_test.sh	the honest test: re-select on pre-June-1 data only, then backtest June-1+ forward

The cache is the point

cache.duckdb holds ~30k wallets / 14M+ bets, pulled once. Every score — any archetype, any cutoff, the clean OOS test — now runs in seconds instead of hours of API pulls. MAX_AGE_DAYS=14: the broad pool refreshes biweekly; the watchlist is force-refreshed daily (cache.invalidate) for forward tracking.

Schema v2 (2026-07-02) — token-keyed, provenance-tagged, archival. bets now carries asset (token id — the position identity), src/ts (endpoint provenance + close time), and resolved (False = early-sold position in a market that hadn't ended at pull time; its won is a curPrice mark, not an outcome — scorers filter these). p is stored raw (0 = avgPrice missing) and clamped to [0.001, 0.999] by get_bets on read, so "missing price" stays distinguishable from a real 0.1¢ longshot. Refresh is an upsert by token (plus superseded legacy rows), not a wallet wipe: each pull still covers the rolling WINDOW_DAYS, but rows that slide out of the window now survive, so per-wallet history accumulates into a permanent archive. The same-asset row from both endpoints (a partially-closed position) is deduped to the larger- stake row instead of double-counting. Failed pulls are returned empty but NOT cached and NOT marked pulled — they retry on the next call instead of masquerading as "no bets" for MAX_AGE_DAYS (pre-v2, an API error could cache a wallet as empty-and-fresh; that bug bit the watchlist in practice). Legacy v1 rows keep NULLs in the new columns until their wallet's next refresh. The migration runs automatically on first open (v1 → v2, exact-duplicate rows merged). Per-endpoint pagination is still capped at ~2k bets (max_pages); ../wide/pmkt.duckdb remains the deep-history subgraph dataset.

The clean test (why the favorites are a mirage)

clean_test.sh selects favorites using only bets resolved before June 1, then copies their June-1+ entries:

• In-sample (contaminated): 21 favorites, 99% win rate, +23.6%.
• Clean (pre-June-1 selection): 15 favorites, 68% win rate, −7.4% (−19% on the settled portion).

The +23.6% was selection bias. This matches the paper: ~60% of "lucky winners" turn into losers out-of-sample. Don't copy favorite-riders. The value archetype (beats underdog prices) is where real alpha may live — test it with backtest_june.py value.

Strategy backtests

• strategy.py — train (pre-May-30) / test (June1+) wallet selection on copy-ROI
  • z + monthly consistency + diversification. → selection.json.
• followability.py — pull entry timestamps (cached), drop wallets whose edge is in un-followable fast/live markets, re-rank on followable forward bets. → watch_final.json (the execution-realistic list).
• pnl_basket.py / pnl_focused.py — $1,000 capital-constrained copy sims with missed-trade accounting. Key result: the broad basket loses on $1k (can't follow 1,200 trades), but 1–2 wallets + a conviction (bet-size) filter clears out-of-sample. See ../FINDINGS.md.

The repeatable find (conviction_scan.py + validate_timing.py)

The best result. Score wallets on their high-conviction bets only — the top 20% by stake size (per-wallet p80, not a flat $200): the edge is wallets that win 70–80% on genuinely-uncertain (~0.4–0.6) markets — real skill, not favorite-riding. The per-wallet cutoff reproduces flat-$200's win-rate lift while adapting to each wallet's scale (a whale's $200 bet isn't conviction; a minnow's is).

• conviction_scan.py — train pre-June / validate June on conviction bets → 218 matches, 62/83 profitable forward (p≈0). → conviction_wallets.json.
• validate_timing.py — the copyability selection (rewritten 2026-06-23). For every conviction wallet it runs a flat-$50 copy replay and keeps only the ones genuinely profitable to copy: copy_pnl > 0 and a real hold-to- resolution edge (held_pnl > 0, held win-rate ≥55% over ≥8 resolved held bets), active in 30d, median lead ≥1h (drops sub-hour snipers). It also precomputes every stat the dashboard renders — including copy_pnl, the authoritative flat-$50 copy P&L (replay their entries, mirror their exits, settle held bets at CLOB resolution by token_id). → ~14 copy-positive holders in watch_sharps.json (read live by jaxperro.com/trading; the table defaults to sorting by Copy P&L).

Copy P&L is the one number that matters for picking copy targets. It replaced an earlier lead-time gate. The lesson (see FINDINGS "the scalper trap"): position win% over-counts scalpers — cache.won (curPrice ≥ 0.5) scores a sell-at- profit as a "win", so a wallet can show ~100% conviction win% yet lose money when copied (ArbTrader: 99.5% win, −$790 copy P&L). Conviction itself = the top 20% of a wallet's own stake sizes, measured at the position level (a wallet's total stake in a market, not per-trade — a scalper splits one position over many small buys). Judge by Copy P&L, never win%.

Paper portfolio (portfolio.py → portfolio.json)

A $1,000 paper book that mirrors a chosen set of wallets' conviction bets at a flat $50/trade, computed off the cache (not client-side). The cache stores each bet's resolution time (res_t), so capital recycles at the true resolution moment — fixing the phantom capital-lock the old browser replay suffered (it missed resolution dates for high-volume wallets and skipped bets it could afford). Edit the WALLETS list at the top of portfolio.py to change who's followed. Output portfolio.json (equity, splits, current/resolved/missed tables, per- wallet) is read by the dashboard in one request. Hold-to-resolution model (the cache has no sell events — the right model for the holders we select); a small live /positions pull supplies the "current open bets" panel.

Caveat carried through the whole stack: every P&L here is idealized — no slippage, fees, or copy lag, and the wallets were selected in-sample. Treat the headline % as a ceiling, not a forecast (the project's own history: an in-sample copy backtest hit +168% then collapsed out-of-sample).

Dashboard feeds (jaxperro.com/trading)

The dashboard (in the jaxperro repo, trading/index.html) is static and reads two precomputed JSONs from this repo via raw.githubusercontent:

• watch_sharps.json — the sharps table (copy-positive holders + per-wallet Copy P&L, win%/record, avg bet, leads).
• portfolio.json — the $1k paper book up top.

Both are committed + pushed by daily.sh, so the page is just a renderer (no per-wallet API calls). It keeps a client-side replay as a fallback if the feed is unreachable.

Copy execution (separate, in progress)

Actually placing the copied trades lives outside this finder — copybot.py + sync_floors.py (a teammate's work). sync_floors.py recomputes each followed wallet's p80 conviction floor from the fresh cache into ../config.json so the bot's entry gate stays in parity with the dashboard's top-20%-by-stake definition. This finder's job is selection + tracking; execution is downstream.

Daily (daily.sh, launchd 10:00)

1. discover (enumerate.py 14) — recent liquid markets → candidate pool.
2. freshen — force-refresh the watchlists, then collect.py tops up new/stale.
3. re-score (skill.py) → watch_skilled.json.
4. sharps (conviction_scan.py + validate_timing.py) → conviction_wallets.json
   • watch_sharps.json (the copy-positive holders).
5. floors (sync_floors.py) — copy-bot conviction floors → ../config.json (local).
6. portfolio (portfolio.py) → portfolio.json (the $1k paper book).
7. dashboard (dashboard.py) + snapshot to history/.
8. publish — commit + push the JSON feeds, then ping Discord (daily_webhook in the gitignored ../config.json).

Schedule via launchd/cron (Mac must be awake).

Usage

pip install duckdb
python3 enumerate.py 180        # build candidate pool (last 6 months)
python3 collect.py              # cache all candidates (one-time, slow; resumable)
python3 skill.py                # -> watch_skilled.json (seconds, from cache)
python3 dashboard.py            # -> dashboard.html
./clean_test.sh                 # the out-of-sample verdict

Local data (*.duckdb, candidates.json, *_scored.json, history/) is gitignored — regenerate via the steps above.