Investing by The Assembly Method
I started with a quantitative ETF allocation engine and turned it into a broader personal investment intelligence product designed to support disciplined long-term investing with less monitoring and clearer evidence.
The model was not the product
The project began as a reproducible ETF decision-support system. It ingested market and macro data, computed leakage-safe features and regime signals, produced allocations, and evaluated them through backtests and robustness checks.
That technical foundation raised a broader product question. A useful investing product should not require constant market watching or encourage activity simply to appear useful. I reframed the system around helping an owner understand what matters, make better long-term decisions, and preserve control over the portfolio.
I separated decision quality from engagement
- Owner control is explicit through account policy, position boundaries, pause controls, and account-specific execution authority.
- Recommendations and intentional non-actions are expected to carry rationale, uncertainty, and evidence rather than appear as unexplained model outputs.
- The website handles setup, trust, controls, history, and deeper inspection while the recurring relationship is designed around useful financial correspondence.
- Doing nothing is a valid product outcome when the evidence does not support a portfolio change.
The system became a governed product stack
I kept the quantitative engine, but placed it inside a broader operating system. The product maintains owner and account policy, ingests portfolio and market context, computes deterministic and ML-assisted evidence, produces governed decisions, compares added complexity with simpler alternatives, and records the resulting rationale and history.
The repository separates ingestion, deterministic features, ML research, decisioning, backtesting, hosted operations, the customer-facing web product, and test workflows. New model complexity is not automatically considered product progress.
System map
From owner intent to a governed investment decision
The architecture is designed so model output cannot skip policy, evidence, permission, or reconciliation on its way to an owner-facing result.
1 · Owner boundary
The system starts with explicit authority
- Goals and account policy
- Position-level Manage / Keep / Exclude rules
- Permissions, pause controls, and tax constraints
2 · Intelligence + evidence
Research has to survive evidence gates
- Portfolio, market, macro, and risk inputs
- Deterministic features and ML research
- Decision rationale, uncertainty, and simpler baselines
3 · Governed outcome
Action is only one possible result
- Recommend action or intentionally do nothing
- Execute only inside account-specific authority
- Reconcile, remember, explain, and correspond
Multi-user hosting forced stronger boundaries
The hosted product uses Vercel for the web application, Cloud Run Jobs for scheduled runtime, and Supabase for authentication and owner-scoped durable state, alongside a separate API surface. Tenant isolation is enforced through owner identifiers, row-level security, API authorization, worker scoping, and protected secret boundaries.
I treated permissions as account-specific rather than global. Policy, position overrides, execution authority, tax constraints, and provider readiness remain separate so a browser interaction cannot quietly become permission to trade.
The strategy had to earn its complexity
One of the most important design decisions was making negative evidence a valid result. The investment engine compares its complexity against simpler reproducible alternatives rather than assuming that a more sophisticated model deserves promotion.
In the current long-history value scorecard, the replay achieved complete price coverage across 43 periods and passed the predeclared 36-period foundation gate. Sharpe and drawdown gates passed, but the CAGR advantage remained below policy. The recorded assessment is MIXED with a REVISE disposition.
That result is useful evidence. It separates the ability to build a sophisticated system from evidence that the additional complexity is actually improving the investment decision.
What exists today, and what remains unproven
- A live customer-facing Investing product and authenticated multi-user foundation.
- Owner-scoped policy, account, position-boundary, and hosted-isolation controls.
- Scheduled quantitative and operational workflows with reproducible evaluation machinery.
- A product direction centered on explanation, education, correspondence, and bounded automation.
- No claim that the strategy has proven superior returns or a durable investment edge.
The project has evolved from an allocation-model experiment into a broader product question: how to build an investment relationship that knows when to recommend action, when to explain, and when to leave the portfolio alone.