BTC 5-Minute Up/Down Bot

Automated trading bot for Polymarket's 5-minute Bitcoin Up/Down binary markets.

Current Status (Feb 19, 2026)

Live trading with safety filters. Previous version lost ~$56 from $130 deposited due to ghost fills, adverse selection, and bad trade filtering. v2 adds strict entry filters based on analysis of 162 real trades.

• Balance: $74.28 USDC
• Proxy wallet: 0x45bfb3aB984aFDA6c801b4a3cD5126c16926E42E
• Real performance (v1): 159 markets, 65.4% WR, -$96.64 PnL

How It Works

Polymarket runs a new BTC Up/Down market every 5 minutes (288/day). Each resolves to "Up" if BTC ends the window >= start price, "Down" otherwise.

The bot uses a GBM volatility model to estimate the true probability of each outcome. When model confidence diverges enough from market pricing, it places maker limit orders.

Core Formula

P_up = Phi(ln(S_t / S_0) / (sigma * sqrt(tau)) * 0.95)

• S_0 = BTC price at window start
• S_t = current BTC price (Binance websocket, real-time)
• tau = seconds remaining in window
• sigma = EWMA per-second volatility (10-min lookback)
• 0.95 = mean reversion dampening factor

v2 Entry Filters (from real trade analysis)

Filter	Value	Reason
P_ENTRY	>= 0.84	Minimum model confidence
MIN_BUY_PRICE	>= $0.70	$0.65-0.70 bucket was 41% WR, -$43 total
MIN_EDGE	>= 0.00	Negative edge trades lost $20.75
MAX_TAU	<= 250s	Early entries (280s+) had high reversal rate
MIN_TAU	>= 30s	Too close to expiry = no time for fill
SKIP_HOURS	0, 3	Midnight=50% WR, 3AM=63% WR
VOL_KILL	0.05-0.30%	Skip extreme volatility

Expected impact: ~13 trades/day (vs ~160 before), 84% WR on filtered subset.

Order Execution & Safety

• Fill verification: After placing GTC limit order, waits 3s then checks get_order_fill(). Cancels unfilled orders.
• Post-expiry phantom check: Re-verifies order status after window expires. Skips trades with filled=0 or CANCELED.
• Redemption verification: Checks USDC balance before/after redeem_positions(). If $0 received, adjusts internal bankroll.
• Kelly sizing: Scaled Kelly fraction (0.05-0.50), sized off internal bankroll.

Lessons Learned (v1 post-mortem)

1. Ghost fills: GTC limit orders can show MATCHED but deliver 0 shares. Always verify size_matched.
2. Adverse selection: Winning limit orders don't fill (price moves away), losing orders always fill. Paper WR (77%) ≠ real WR (65.4%).
3. Silent redemption failures: redeem_positions() can return True but deliver $0.
4. BTC price ≠ market resolution: Polymarket's oracle can resolve differently than our Binance feed.
5. PnL must be verified on-chain, not assumed from internal logic.

Architecture

live_trader.py     Main async loop, filters, Kelly sizing, order management
config.py          Constants, env vars
price_feed.py      Binance BTC/USDT websocket + EWMA volatility
model.py           GBM probability model, fee calculations
market.py          Polymarket CLOB client (orders, books, redemptions)
backtest_realistic.py  Backtester with adverse selection modeling
replay_filtered.py     Replay historical trades through filter configs

Data Files

data/live_trades.csv                    Active trade log (clean, v2)
data/live_trades_OLD_GHOST_FILLS.csv    v1 trades (includes ghost fills, unreliable)
data/btc_prices_2s.csv                  BTC price history (2s intervals, 143k rows)
logs/book_snapshots.csv                 Order book snapshots (39k rows)

Setup

uv venv && source .venv/bin/activate
uv pip install -r requirements.txt

Requires ../.env with:

PK=<polygon_private_key>

Usage

# Start live trading
cd btc-bot
.venv/bin/python live_trader.py

# Monitor
tail -f data/live_trader.log
cat data/live_trades.csv

Bankroll

• Starting bankroll: $74.28 (reset Feb 19, 2026)
• Kelly fraction: 0.05-0.50 (scaled by bankroll growth)
• Min bet: 5 shares / $1
• Target: $9/day profit ($270/month to cover inference costs)