Automated Investing — STRATEGY
Problem
Trading is a data-intensive domain with the cleanest, most honest feedback signal in any RDCO bet (P&L). The founder lacks the ability to watch markets continuously; an always-on agent reacting at the right time is exactly what RDCO’s targeting-system + agent architecture should be good at. The bet is whether RDCO’s agentic infrastructure (loops, sub-agents, vault, instrumentation) can produce a strategy that beats a discipline-grade calibration baseline (Brier ≤ 0.12) on a real venue.
Approach
Single-threaded staged build now, multi-agent later, same interfaces. Per architecture-vision.md, the eventual shape is five agents (Strategy Research / Paper Testing / Execution / Monitor-Risk / Reporting), but today everything runs in one process with the autoinv Python package as the shared library, with component boundaries deliberately drawn so future agent splits are additive rather than rewrites. No real capital deployment until a strategy beats the discipline gate (0.12 Brier) on a live event. Curriculum-disciplined progression through 5 quant levels (probability → stats → linalg → calculus/opt → stochastic + advanced) with explicit deliverables per level. Why-this-not-that: don’t build the multi-agent system upfront (“I tend to overengineer up front”); don’t skip paper-testing to chase live capital; don’t build agents-for-agents’-sake — each split needs a load/failure-isolation/specialization justification.
Personas
- The founder (single user, internal-facing): wants a portfolio he doesn’t have to babysit. Reached via vault writeups + iMessage/Discord status updates from Ray.
- Ray (COO agent): operating the loop today across all five future-agent roles manually.
- TBD — needs founder distillation: this bet currently has NO external persona. It’s an internal-capability bet, not yet aimed at external customers. If/when it productizes (signal-as-a-service? educational content? case study?), persona work happens then. Worth surfacing as an explicit open question whether this stays internal-only.
Key metrics
- Discipline gate (the load-bearing gate): Brier ≤ 0.12 on live event before real capital
- PM1 baseline result: top-650 Polymarket markets ($7M+ vol) Brier 0.033 at 3-day window — venue is hyperefficient at high volume. Correction (PM1b): three of five volume bands between $10K and $100M FAIL the 0.12 gate, so $100K-$2M band is the plausible directional-strategy target
- Curriculum progression: L1 (probability) and L2 (statistics) deliverables shipped; L3 (linalg) next
- Strategy candidates filed: 0 as of 2026-05-01 (handoff schema awaiting founder approval per Apr 23 audit item #10)
- P&L: $0 — no live capital deployed; will not be deployed until gate passes
Work tracks
- Strategy-candidate handoff schema approval (founder-gate item #10 in Apr 23 audit): question / hypothesis / entry-exit rules / data sources / expected Brier-Sharpe. Owner: founder (15 min). Status: drafted in architecture-vision.md, awaiting sign-off. Unblocks Ray autonomously filing candidates.
- Curriculum L3 (linear algebra): S&P 500 PCA + Markowitz mean-variance optimizer from scratch with cvxpy. Owner: Ray. Status: queued.
- PM1c+ follow-up on $100K-$2M Polymarket volume band: directional-strategy candidates against the band that fails calibration gate. Owner: Ray. Status: gated by item #1.
- Paper-testing loop formalization: walk-forward windowing (currently in-sample only), automated gate evaluation (Brier < 0.12 + permutation test p < 0.05 + bias audit PASS), per-strategy paper-trading registry. Owner: Ray. Status: partial;
autoinv.engine.Backtest+BiasAudit+permutation_testshipped. - Execution agent for prediction markets (Stage 2+): concrete
PolymarketExecutionHandlerrequiring wallet auth. Out of scope until a strategy passes the gate. - Reporting agent: automated P&L dashboard, monthly summary, tax-relevant transaction log. Owner: Ray. Status: not started; today’s reporting is vault writeups + iMessage updates.
Targeting system (RDCO 4-layer thesis)
Sub-process: strategy quality. Defined operationally as: Brier score on live events (calibration), Sharpe / drawdown profile, bias-audit PASS, permutation-test p < 0.05. Concrete and measurable per the discipline gate.
P&L meta-layer: real-money risk-adjusted return after fees + slippage + capital cost. Currently $0 because no live capital is deployed. Per the recursive-structure thesis: even a strategy that passes sub-process gates can be vetoed at the P&L layer if expected returns don’t survive sizing constraints (Kelly-fractional) or capital opportunity cost.
Instrumentation (RDCO 4-layer thesis)
Built: Polymarket public Gamma + CLOB API client; yfinance pull for equities; Brier-score computation across time windows; bias audit; permutation test; drawdown report; strategy report; experiment writeups in vault with embedded results/plots. Vault as longitudinal sensor (every experiment is a dated md doc with reproducible script).
Gaps: live-monitoring (no real-time drawdown watchdog yet); paper-trading registry tracking history per strategy; Kelly-criterion position sizing (deferred from consolidation pass per Halls-Moore Ch 12); per-strategy capital allocation tracking; tax-relevant transaction log.
Tools (RDCO 4-layer thesis)
Built: autoinv package (engine, validation, stats, portfolio); event-driven Backtest orchestrator; SimulatedExecutionHandler (commission + slippage); abstract ExecutionHandler base ready for live subclassing. Demo template script per strategy. uv-managed Python 3.12 env at 01-projects/automated-investing/.venv/.
Gaps: walk-forward windowing; automated gate evaluation; concrete PolymarketExecutionHandler (wallet auth); BrokerageExecutionHandler for equities (Alpaca / IB / TD candidates per index.md); order state tracking; fills reconciliation; real-money idempotency.
Feedback loop (RDCO 4-layer thesis)
Built (curriculum loop): each level has a deliverable; gate-to-next is concrete; experiment writeups are the diagnostic artifact. PM1 → PM1b correction is a live worked example of the loop closing (initial conclusion was wrong; band-filtered re-analysis surfaced the actual signal).
Partial (strategy loop): research-agent → paper-tester boundary is the immediate friction. Without the candidate-handoff schema (work track #1), Ray re-asks the founder for ratification per idea instead of running the loop autonomously. That’s the bottleneck blocking the loop from closing on the strategy side.
Open questions for founder
- Approve strategy-candidate handoff schema as drafted (Apr 23 audit item #10) so Ray can file candidates autonomously?
- Is this bet permanently internal-only (“portfolio Ben doesn’t have to babysit”), or is there a future productization path (signal-as-a-service, educational content, MAC-style framework)? Materially affects whether L4-L5 work has external-facing implications.
- After PM1b correction surfaced the $100K-$2M Polymarket volume band as the directional-strategy target, do you want me to file specific candidate strategies against that band, or is the next priority L3 curriculum progression?
- Risk tolerance on first live deployment: what’s the maximum capital you’re willing to put behind the FIRST strategy that passes the discipline gate? (Sets the sizing-constraint side of Kelly work.)
Related
- ../../06-reference/2026-04-30-rdco-thesis-targeting-systems-feedback-loops
- ../../06-reference/2026-04-30-rdco-bet-architecture-playbook
- architecture-vision
- infrastructure-plan
- experiments/pm1-polymarket-baseline
- experiments/pm1b-polymarket-long-tail-correction