Polymarket Insider-Trading Detection — Reproducibility

Public-data surveillance pipeline. Scopes event-driven Polymarket markets, engineers per-trader features, applies named explainable heuristics, and ranks accounts by insider-pattern likelihood.

Ethics & redaction. Outputs are investigative leads, not accusations. Wallet addresses in this repo are truncated (0xabc123…); only publicly documented / charged cases are referred to by name. The machine-readable lead lists (flagged_accounts.csv, trader_scores.csv, wallet_funding.csv, …) are git-ignored and not published — they regenerate from the code against live public data.

Requirements

• Python 3.10+, standard library only (urllib, json, csv, math, datetime). No pip install.
• No API keys. All endpoints are public: Gamma, Data-API, CLOB, and a keyless Polygon RPC.
• Network access. Be patient: the collector self-throttles and caches.

Run order

# 1. Scope the universe (event-driven markets in the window) -> markets.csv, info_timestamps.csv
python3 collect.py scope

# 2. Collect the scored example end-to-end (the Axiom / ZachXBT complex).
#    Rebuilds cache/axiom_complex.json + cache/axiom_participants.json + trade/user caches.
python3 collect.py complex axiom

# 3. Rank -> writes ALL FOUR trader CSVs: trader_scores.csv, flagged_accounts.csv,
#    trader_features.csv, trades.csv  (runs feature engineering internally)
python3 rank.py

# (optional) features.py standalone also writes trader_features.csv + trades.csv
# python3 features.py

# 4. On-chain funding & cohort-clustering forensics for the flagged set -> wallet_funding.csv
python3 funding.py

# (optional) raw-RPC funding for a single wallet (capped at 10k-block getLogs windows)
# python3 collect.py funding 0x1d9af6…

Steps 1–4 regenerate every result and CSV in the report. Everything is idempotent and cached in ./cache/ — re-running never re-downloads. To force a clean rebuild, delete ./cache/.

What each file does

File	Role
collect.py	All data pulls (Gamma/Data/CLOB/Polygon-RPC), caching, scope + complex collection, on-chain funding
features.py	Per-trader feature engineering (lifecycle, timing-vs-info, concentration, conviction, evasion)
heuristics.py	Named rules → 0–1 sub-scores, with documented thresholds and the case each is calibrated against
rank.py	Composite score, tiers, two-axis view; writes the scoring CSVs
funding.py	On-chain funding + cohort clustering (Blockscout) for flagged wallets
report.md	Methodology + findings

Outputs (CSV)

markets.csv, info_timestamps.csv, trades.csv, trader_features.csv, trader_scores.csv, flagged_accounts.csv, wallet_funding.csv. trader_scores.csv carries the per-heuristic breakdown, so any row can be read as why the account is flagged.

Data-source and reliability notes

• Gamma /public-search, /events?slug=, /markets — /markets?slug= is unreliable (returns empty for some markets), so the complex is built by fetching the event by slug and reading its child markets. Page size caps at 100; paginate with offset.
• Data-API /trades?market= and /trades?user= (cap 500). For mega-markets the collector uses filterType=CASH&filterAmount= to keep the pull to material players. Post-resolution /holders shows only losing/dust balances (winners redeem to ~0), so winners are found via trades, not holdings.
• CLOB /prices-history — pass bounded startTs/endTs; interval=max can hang.
• Polygon RPC — polygon-rpc.com is keyed off; keyless working endpoints are rotated (publicnode, 1rpc, drpc, onfinality). eth_getLogs ranges are capped, so funding pulls are chunked and run only for flagged candidates.

Known limitations (see report.md §7)

• The scored population is one event (the Axiom complex) — a worked example, not the platform. Broadening to the Maduro / d4vd / Iran families is the obvious next step.
• Trade pulls for the complex are cash-filtered (≥$1k); micro-bettors are excluded from scoring.
• On-chain funding is optional here; "fresh account" uses Polymarket trade-history tenure (with a censoring guard for ≥500-trade wallets) rather than the wallet's on-chain birth.
• The information timestamp is semi-manual and is the dominant error source for novel markets.
• Outputs are investigative leads, not accusations.