web / scripts /02_analyze_wallets.py
LinJennyLin's picture
Deploy SharpFlow (nav + example wallets + Polymarket copy)
d4c51f2 verified
Raw History Blame Contribute Delete
3.67 kB
#!/usr/bin/env python3
"""Step 2 — fetch history for each candidate and compute metrics + patterns.
python scripts/02_analyze_wallets.py
Reads data/candidates.json
Writes data/wallets.csv (one row per wallet: metrics + pattern features)
data/ledgers/<wallet>.parquet (cached raw ledgers for the backtester)
"""
import sys
from pathlib import Path
import pandas as pd
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from pmct.client import PolymarketClient
from pmct import ledger as ledger_mod
from pmct import metrics as metrics_mod
from pmct import patterns as patterns_mod
from pmct import discover as discover_mod
from pmct.util import load_config, cache_path, read_json
def main():
cfg = load_config()
acfg = cfg["analyze"]
candidates = read_json(cache_path(cfg, "candidates.json")) or []
if not candidates:
raise SystemExit("No candidates. Run 01_discover_wallets.py first.")
client = PolymarketClient(sleep_sec=acfg["request_sleep_sec"])
# authoritative profit/volume from the leaderboard (one pass, reused for all)
dcfg = cfg["discover"]
lb_map = discover_mod.leaderboard_profit_map(
client, dcfg["leaderboard_windows"], dcfg["leaderboard_top_n"])
print(f"authoritative leaderboard entries: {len(lb_map)}")
# analyze leaderboard wallets first (authoritative profit = best copy targets),
# then fill the budget with the rest, so max_wallets isn't wasted on low-hex
# addresses that happen to sort first.
candidates.sort(key=lambda w: (0 if w in lb_map else 1, w))
candidates = candidates[: acfg["max_wallets"]]
rows = []
for i, w in enumerate(candidates, 1):
try:
raw = list(client.iter_activity(
w, acfg["activity_page_limit"], acfg["max_pages_per_wallet"]))
led = ledger_mod.normalize_activity(raw)
if led.empty:
continue
open_df = ledger_mod.attach_open_positions(client.positions(w))
stats = metrics_mod.compute(w, led, open_df)
feats = patterns_mod.extract(w, led)
row = {**stats.to_dict(), **feats.to_dict()}
row.pop("wallet", None)
lb = lb_map.get(w, {})
lb_profit = lb.get("lb_profit")
row["on_leaderboard"] = bool(lb)
row["lb_profit"] = lb_profit
row["lb_volume"] = lb.get("lb_volume")
# authoritative profit: trust the leaderboard when ranked, else our recon
row["profit_final"] = lb_profit if lb_profit is not None else row["total_pnl"]
# recompute rebate share against the AUTHORITATIVE profit so truncated
# (negative) reconstructions don't falsely flag profitable wallets
pf = row["profit_final"]
if pf and pf > 0:
row["rebate_share"] = round(min(max(row["rebate_pnl"] / pf, 0.0), 1.0), 4)
elif row["rebate_pnl"] > 0:
row["rebate_share"] = 1.0 # no net directional profit, only rebates
row = {"wallet": w, **row}
rows.append(row)
led.to_pickle(cache_path(cfg, "ledgers", f"{w}.pkl"))
print(f"[{i}/{len(candidates)}] {w} pnl=${stats.total_pnl:>10} "
f"wr={stats.win_rate:.2f} rep={feats.repeatability:.2f}")
except Exception as e: # noqa: BLE001 - keep going through the batch
print(f"[{i}/{len(candidates)}] {w} ERROR {e}")
df = pd.DataFrame(rows)
out = cache_path(cfg, "wallets.csv")
df.to_csv(out, index=False)
print(f"\nAnalyzed {len(df)} wallets -> {out}")
if __name__ == "__main__":
main()