AI Execution Engine

The DRL agent optimizes how orders are executed — not what to trade, but the optimal way to slice and time each order:

State space (6 dimensions): - Current bid-ask spread (bps) - Order book depth imbalance [-1, 1] - Trade flow imbalance [-1, 1] - Time remaining in execution window [0, 1] - Percentage already filled [0, 1] - Recent 5-min realized volatility (bps)

Action space: - Slice percentage (0-100% of remaining quantity) - Aggression level (passive / mid / aggressive) - Wait time before next slice (0-5000ms)

Reward: -slippage_bps - market_impact_bps + time_bonus

The model is trained in the research lab (Python/Ray RLlib) and exported as a lookup table for the TypeScript runtime. Fallback: if no trained model is available, uses the Predictive Spread Model (EWMA-based).

For orders above $5,000 notional, the SOR splits execution across multiple exchanges simultaneously:

How it works: - Probes top-of-book on every venue where the user has credentials - Drops venues whose effective price is worse than best by > 8 bps - Allocates proportionally to (1 / spread_bps) — tighter spread = more allocation - Drops legs below $500; redistributes to the best leg - Maximum 4 parallel legs

Result: Better average fill price, reduced market impact, and natural diversification of exchange risk. Each leg independently uses the full TWAP / maker-first / slice-loop machinery.

The Time-Weighted Average Price (TWAP) slicer learns from historical fills:

• Reads bot_orders.slippage_bps over the last 30 days
• Buckets by (symbol, hour-of-day UTC, notional size: small/mid/large)
• Returns median slippage per bucket
• Adjusts sliceCount and delayMs based on learned patterns

Fail-open design: sparse buckets (< 5 samples) return null and fall back to static defaults. The adaptive layer can only improve execution, never make it worse than baseline.

A guardrail (Phase 76b) continuously compares adaptive TWAP performance against the baseline and alerts admins if the adaptive path underperforms.

Extends the Smart Order Router with a slippage-aware score per venue:

effectiveCostBps = priceWorseBps(vs best) + expectedSlippageBps(venue)

Expected slippage is the notional-weighted median of historical fills per (exchange, symbol, hour_bucket) over 30 days. Cold venues fall back to global median, then to a static 8bps default.

This means the router learns which exchanges actually deliver better fills for specific symbols at specific times — not just which has the tightest quoted spread.