Agentic Trading Pipeline
Edge detection → Shadow validation → Real deployment (with human approval)
1
Edge Detection · Mechanism-Grounded
Agent declares mechanism + named participant + durability bucket BEFORE backtest. Reads multi-exchange OHLCV (binance/coinbase/bybit spot + perp, 35d) and Polymarket book ticks. Tests T+1 through T+240s across BTC/ETH/SOL/XRP. Realistic fills: no lookahead, 2s latency, walk-the-book slippage cap.
AUTOMATED · RIGOR v3
2
Backtest · 9 Gates
Runs over 35-day window. Rigor v2 (statistical): Wilson WR CI, block-bootstrap PnL CI, ACF/effective_n, holdout 70/30, longest-streak vs H0, max-day fraction. Rigor v3 (informational): M (mechanism declared), R (regime cells positive 4/4), F (FDR-corrected cushion). v3 surfaces "is this real edge or variance" — never auto-blocks.
AUTOMATED · 9 GATES
3
Shadow Bot Built
Agent clones strategy, applies filters. Creates DB table, resolver, poller, dashboard page. Shadow uses 1s fill delay at WS ask — no real orders, no money at risk.
APPROVAL TO DEPLOY
4
Shadow Validation (7+ days)
Shadow runs live. Agent monitors daily: WR, PnL, drawdown, trade volume. Compares vs backtest predictions. Flags divergence > 2σ from expected.
AUTOMATED MONITORING
5
Real Bot Proposal
Agent presents: shadow PnL vs backtest, daily breakdown, per-asset performance, estimated real PnL after 2c execution drag. Human sets share size + wallet.
REQUIRES YOUR APPROVAL
6
Real Deployment · Strict Slippage
Agent builds real bot from shadow with the mandatory drift-abort gate: pre-send fresh-quote re-check, hard abort if ask drifted >2¢ above signal (mark
ABORTED_DRIFT, never chase). Plain limit orders only. Mirrors shadow's slippage filter exactly. First 24h at 5 shares.AUTOMATED (POST-APPROVAL)
7
Live Monitoring
Real vs shadow trade-by-trade comparison. Tracks: side match, fill slippage, missed trades, WR divergence. Alerts on underperformance > $X/day.
CONTINUOUS
8
Retirement
Agent detects edge decay: rolling 7d WR below break-even. Proposes stop with evidence. Shadow continues post-retirement for regime-change detection.
APPROVAL TO STOP
Automated
Requires Approval
Continuous
Strategy Evolution Loop
How the agent mutates, scores, and promotes strategies — and writes every outcome back to memory.
Market Data
↓
Baseline Strategy
↓
Agent proposes one mutation
↓
Backtest
↓
Paper trade / dry run
↓
Score EV, PnL, drawdown, fills
↓
Better than baseline?
YES ↓
Promote to candidate strategy
NO ↓
Discard
↓
Small capital live test
↓
Still better?
YES ↓
Promote to production
NO ↓
Discard
↓
Update strategy memory
Log failure
↻ feeds back into —
"Agent proposes one mutation"
↻