Agentic Fund Orchestration

Agentic Fund Orchestration

Coordinates the other skills in this repo into a single, governed decision loop. Reference design: TauricResearch TradingAgents (analyst → debate → PM → risk → execution).

Mandatory framing

• Notification-first. Run for 6+ months proposing trades to a human before any automated

execution. Automated trading bugs are expensive bugs.

• Hard caps and the kill switch live in deterministic code, outside any LLM. Agents propose;

the deterministic risk layer disposes.

• Educational; not investment advice. Paper-trade everything first.

The team

The daily loop (notification-first)

1. INGEST    pull EOD prices (yfinance) + macro (FRED).            [data-sources]
2. REGIME    regime-detection -> exposure_multiplier.
3. ANALYZE   fundamental-analysis -> valuation context + defensive-sleeve ETF choice;
             any NEW candidate signal must clear the backtest gate before it can trade.
4. SIGNALS   trend-following + analysts -> per-asset signals.
5. CONSTRUCT portfolio-construction -> target weights (× exposure_multiplier).
6. RISK      risk-management -> vol target, drawdown de-risk, CPPI, caps -> risk_scale / veto.
7. DIP       dip-tranches-strategy -> any reserve tranche firing today?
8. REBALANCE rebalancing -> minimal trade deltas (calendar check, threshold act, no-trade bands).
9. TAX       tax-loss-harvesting -> swap any underwater taxable lots.
10. NOTIFY   email/Telegram the proposed trades + plain-English rationale + risk report.
11. EXECUTE  human approves -> execution agent places orders (Alpaca/IBKR).
12. LOG      append every signal/decision/order to an immutable audit log; update metrics.

Shared state (cuts tokens & drift)

Agents read/write a single structured state object (JSON) rather than re-deriving facts through long dialogue. Use structured outputs for facts; reserve natural-language debate (bull vs bear) for the steps where reasoning genuinely adds value.

{
  "date": "2026-05-29",
  "nav": 1000000,
  "regime": {"exposure_multiplier": 0.7, "score": 0.32},
  "signals": {"SPY": "in", "VXUS": "out", "GLD": "in"},
  "targets": {"RSP": 0.18, "VXUS": 0.12, "...": "..."},
  "risk": {"verdict": "scale", "risk_scale": 0.7, "current_drawdown": -0.04},
  "dip": {"tier_fired": null},
  "trades": [],
  "audit_log_ref": "logs/2026-05-29.jsonl"
}

Non-negotiable guardrails

• Human approval for: any new live strategy, trades above a notional threshold, any leverage change.
• Paper-trade for weeks; require live-paper Sharpe/drawdown to match backtest within tolerance.
• Deterministic kill switch + hard exposure caps the agents cannot override.
• Full, immutable audit trail (Git-committed logs or append-only store).

Pitfalls to design against

Overfitting (walk-forward, deflate Sharpe for # of trials), look-ahead bias (point-in-time data, decide on prior close), survivorship bias (yfinance lacks delisted names), transaction costs (#1 killer of paper-profitable strategies), regime change (test across 2008/2020/2022), LLM overconfidence (bound by the deterministic risk layer).

Metrics contract (every agent reports, net of costs)

CAGR, Sharpe, Sortino, Calmar, max & current drawdown, realized vs target vol, gross/net exposure, turnover, win rate + payoff, effective # of bets, time-in-drawdown, per-sleeve attribution.

Starter stack

yfinance + FRED → vectorbt / bt research → Alpaca paper → human-approved go-live with hard caps in code. See the research/ notes 05 and 06 for the full architecture and data/API details.