web / README.md
LinJennyLin's picture
Deploy SharpFlow (nav + example wallets + Polymarket copy)
d4c51f2 verified
|
Raw History Blame Contribute Delete
5.1 kB
metadata
title: SharpFlow
emoji: πŸ”·
colorFrom: green
colorTo: gray
sdk: docker
app_port: 7860
pinned: false

pmct β€” Polymarket copy-trade research toolkit

Research + backtest only. No live trading, no wallet, no funds. This toolkit scans Polymarket traders, finds the ones that are (a) long-term profitable and (b) trading a repeatable pattern, then backtests copy-trading them with conviction-based position sizing (ε€‰δ½θ·Ÿθ‘—ζŠŠζ‘η¨‹εΊ¦θ΅°).

You run it on your own machine β€” it talks to Polymarket's public read APIs.

Pipeline

01_discover_wallets.py   leaderboard + market whales      -> data/candidates.json
02_analyze_wallets.py    fetch history, compute metrics   -> data/wallets.csv
                         + pattern/repeatability features     data/ledgers/*.parquet
03_screen.py             apply the two gates, rank         -> data/shortlist.csv
04_backtest.py <wallet>  conviction vs flat, equity curve  -> data/backtest_*.png

Quick start

cd polymarket-copytrade
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp config.example.yaml config.yaml      # edit thresholds if you want

python scripts/01_discover_wallets.py
python scripts/02_analyze_wallets.py    # slowest step (one API pass per wallet)
python scripts/03_screen.py
python scripts/04_backtest.py           # backtests top of shortlist

The two screens (your rules)

Gate A β€” long-term consistent profit (screen: in config): total_pnl, roi, win_rate, max_drawdown_pct, monthly_profit_ratio (share of active months that were green), plus min_active_days / min_resolved_markets so a lucky week can't pass.

Gate B β€” repeatable, replicable pattern (patterns.py): each wallet is reduced to entry-price-band tightness, sizing consistency, topic concentration, and cadence regularity, scaled by trade count, into one repeatability ∈ [0,1]. Copy a strategy, not a coin flip.

Gate C β€” copyable, not market-making (max_rebate_share): many top leaderboard wallets profit from maker/taker rebates and liquidity rewards β€” an edge you cannot copy. We measure rebate_share (incentive income Γ· authoritative profit) and reject wallets above the threshold.

Authoritative profit. For wallets on the leaderboard we rank on Polymarket's own reported profit (lb-api/profit), not our reconstruction β€” the reconstructed PnL is only a cross-check / the fallback for unranked wallets.

Conviction sizing (sizing.py)

We never copy the target's dollar amount. Per entry we infer conviction ∈ [0,1]:

conviction = w_stake Β· stake_signal + w_edge Β· edge_signal
stake_signal = how big this bet was vs the target's own typical bet
edge_signal  = |price βˆ’ 0.5| Β· 2   (distance from a coin flip)

Our stake = bankroll Β· frac, where frac scales from min_frac_per_trade (no conviction) up to max_frac_per_trade (full conviction), then is clamped by a kelly_cap Β· kelly_fraction ceiling and a max_frac_per_market aggregate cap. Defaults are tuned for a <$1k, conservative bankroll.

Important caveats β€” read before trusting a number

  • Endpoints verified live (2026-07) against the real API: Gamma /markets, CLOB /price /prices-history, Data API /activity /positions /holders /value, and the leaderboard at lb-api.polymarket.com/profit /volume (windows: all 1d 7d 30d). If Polymarket later moves a path, fix it in pmct/client.py (one place).
  • History is capped at ~3500 most-recent activities per wallet. The Data API refuses offset paging past 3000 ("max historical activity offset of 3000 exceeded"). Enough for pattern/behavior analysis; for lifetime profit we use the authoritative leaderboard number instead of reconstructing.
  • Reconstructed PnL validated to ~11% of the authoritative leaderboard figure on a full-history wallet ($3.61M vs $4.05M) β€” the gap is open-position marking + rebates + methodology. Trust profit_final (leaderboard) over total_pnl.
  • Resolution is approximated from the activity ledger (net share balance + redemptions), not from on-chain payout events. Good enough for ranking; not an accounting statement. Win-rate is the weakest metric β€” prefer profit + consistency.
  • Backtest fills are approximate: entry/exit = target's price Β± slippage; copy_delay_sec is folded into slippage rather than modeled tick-by-tick. Real copy-trading is worse than backtest β€” you react after the target moves the price. Treat backtest returns as an optimistic ceiling.
  • Past performance β‰  future. Prediction-market edges decay and crowd out. This is a research scaffold, not investment advice.

Next step toward live (only after backtests convince you)

The live agent would poll data-api /activity for your shortlisted wallets, detect new fills, and place mirrored orders via py-clob-client with the same sizing.py logic. That needs a funded Polygon wallet + CLOB API creds β€” not included here on purpose, since you're at the research stage.