trade-review-loop

name

trade-review-loop

description

Closed-loop learning: snapshot every trade decision into a decision packet, run weekly diagnostics (false positives by regime, edge decay by setup, execution drag by condition), and emit advisory tuning proposals for scanner weights and guardrail thresholds. Use when working on decision packets, trade review, outcome attribution, weight/threshold feedback, or core/decision_packet.py / trade_review.py / weight_feedback.py.

Trade Review Loop

Canonical Contract

Every trade decision becomes a DecisionPacket (the unit of post-trade evaluation). Outcomes are backfilled when trades resolve; the weekly review aggregates packets into diagnostics; diagnostics produce advisory tuning proposals that humans promote OFF → SHADOW → LIVE. Nothing auto-applies.

Pairs with: [signal-scanner] (weights it proposes to tune), [execution-quality-lab] (execution drag), [decision-card-builder] (the card captured in each packet).

Decision packet lifecycle

decision (approve) ──build_packet──> record_packet ──> decision_packets.json
                                                           │
trade resolves ──self-study / digest──> backfill_outcome ─┘
                                                           │
weekly ──load_packets──> trade_review.weekly_report ──> weight_feedback.propose

• DecisionPacket (in core/contracts/) captures decision-time context: regime_state, volatility_state, setup_type, gate_disposition, policy_id, rank_score, edge_score, p_up_calibrated, plus a backfilled outcome (label, realized_return_pct, horizon_days, realized_slippage_bps).
• Recorded additively in webapp/main.py approve_trade (guarded; never affects the trade). Backfill via decision_packet.backfill_outcome(...) from the self-study / weekly-digest job once outcomes are known.

Weekly diagnostics (core/trade_review.weekly_report)

Diagnostic	Grouping	Metric
false positives by regime	regime_state	loss rate among resolved decisions
edge decay by setup	setup_type	predicted edge − realized return
execution drag by condition	volatility_state	avg realized slippage bps

Rates are computed only over resolved packets; coverage_pct reports how many are labeled. Wire weekly_report into the existing weekly digest job (schwab_skill/main.py scheduler) to ship it to Discord.

Feedback (core/weight_feedback.propose)

Turns the report into proposals (each: kind, target, direction, scope, evidence, confidence). Requires >= _MIN_SAMPLES resolved samples. Examples:

• high regime FP rate → increase QUALITY_MIN_SIGNAL_SCORE for that regime
• high edge decay for a setup → decrease that setup's ensemble weight
• high execution drag in a volatility bucket → decrease EXEC_POLICY_TIGHT_SPREAD_BPS

Advisory only. Apply through config + the OFF → SHADOW → LIVE rollout; never auto-tune.

Endpoints

• GET /api/cockpit/decision-packets?limit=N — recent packets
• GET /api/cockpit/review — weekly diagnostics + tuning_proposals

Adding an outcome backfill source (checklist)

1. When a trade resolves, compute label + realized return/slippage.
2. Call decision_packet.backfill_outcome(skill_dir, packet_id, label=...).
3. Ensure the packet_id is discoverable (store it on the order / self-study row).
4. Confirm weekly_report coverage rises; proposals appear once _MIN_SAMPLES met.

Key Files

• schwab_skill/core/contracts/decision_packet.py — DecisionPacket DTO
• schwab_skill/core/decision_packet.py — build / record / load / backfill
• schwab_skill/core/trade_review.py — weekly diagnostics
• schwab_skill/core/weight_feedback.py — advisory tuning proposals
• schwab_skill/webapp/main.py — approve hook + /api/cockpit/review, /decision-packets
• schwab_skill/self_study.py, schwab_skill/main.py (weekly digest) — outcome backfill points