Alezanello/polymarket-arena-eliminated-agents

Polymarket Arena — Eliminated Agents (trades)

Per-trade log from a live learning arena in which hundreds of formula-based trading agents compete (paper money, but priced and resolved against real Polymarket crypto up/down markets — BTC, ETH, SOL, XRP, DOGE, BNB — across the 5m, 15m and 1h timeframes). This dataset contains the trades of the agents that were eliminated (kept out of the live roster). Elimination criterion for this export: Not in the per-timeframe keep set as of the 2026-06-19 deep-clean: KEPT = top-50 currently-active agents per timeframe (ranked superavits desc, respawns asc, realized_usd desc) plus any passing superavits>=1 AND respawns<3, plus Control/Hold baselines. Everything else is archived here (superset of the earlier >10-bankruptcy cull).. The random Control and Hold baselines are exempt. Sibling of polymarket-arena-capture.

How an agent lives or dies: respawns and superavits

Every agent starts each life with a paper bankroll of $100 and sizes each trade at 5% of its current bankroll.

• respawns (bankruptcies): if the bankroll falls below 5% of the initial stake ($5), the agent goes bankrupt → reset to $100, respawns += 1. So respawns = how many times it went broke (a chronic-loser signal).
• superavits (10× milestones): if the bankroll reaches 10× the stake ($1,000) → reset to $100, superavits += 1. So superavits = how many times it grew $100 into $1,000 (size-invariant skill metric).

Bankroll/respawns/superavits are shared across assets (one pool per agent per timeframe).

Files

• trades_5m.parquet, trades_15m.parquet, trades_1hr.parquet — one row per trade (entry → exit / resolution). 4,779,736 trades, 134 MB. Schema = the trades table.
• eliminated_agents.json — per timeframe, eliminated agents with respawns, superavits, n (trades), wins, realized_usd.

Key columns of trades

tf, agent, family, asset, side (Up/Down), entry/exit/ask (price), pnl (net cents/share = payoff+rebate−fee), pnl_usd (paper dollars), shares, fee, rebate, confirmed (1 = resolved by gamma, final), outcome_win, win_start/entry_ts/t (epoch), entry-time features (tnorm,dist,spread,sigma,spot_mom,basis), exec (maker/taker), slug/cond (public market ids).

⚠️ Use confirmed=1 for honest P&L (confirmed=0 rows are provisional, pre-gamma). All P&L is paper; no wallet/key/account data is included.