Private-investor decision support — survivorship-free firm screening + macro fragility
Decision support for a private investor — a survivorship-free 476-firm panel and a self-improving entity model built on real SEC data, paired with macro-fragility and crowd-cascade simulations. Research-grade, auditable, and honest about what it can and cannot claim.
Outcome metrics
- Panel
- 476 firms, survivorship-free, from real SEC EDGAR filings
- Self-improving model (closed propose → verify → keep loop)
- +0.098 mean out-of-sample improvement, 5 / 5 holdout splits
- Honest verdict
- generalization-honest / defensive, not alpha
- Compliance posture
- research / paper-only — not investment advice
Engagement shape
A private investor does not need stock tips; they need decision support — a way to pressure-test theses against data and scenarios, with the uncertainty made explicit. Comuvia built exactly that, end to end, on owned infrastructure.
What it produces
- A survivorship-free firm panel. 476 companies assembled directly from real SEC EDGAR filings — no survivorship bias baked in, so the screen is honest about the firms that didn't make it.
- A self-improving entity model. Comuvia's first end-to-end closed self-improvement loop ran on this real data: it proposes candidate models, an owned verifier scores them out-of-sample, and only improvements that survive a touch-once holdout are kept. It cut the generalization gap on 5 of 5 splits (+0.098 mean out-of-sample).
- Scenario context. The same engine supplies macro-fragility (US recession stress) and free-data crowd-cascade simulations, so a firm-level screen sits inside a regime view.
The honest result
The discovered model is honest and defensive, not alpha — it generalises better than the baseline without overfitting, and the write-up says so plainly. That candour is the product: a decision-maker can trust a screen that reports its own limits.
What it proves
- Owned verifiers beat self-report. A model change only counts as "better" if it moves an out-of-sample number on a held-out set — never by assertion.
- Real data, reproducible. Built on public filings with a survivorship-free panel and a versioned holdout, so results are auditable.
Constraints
Research / paper-only. Not investment advice. Market-data features and private-firm data are gated and used only with explicit sign-off. The value is the decision discipline — honest screens, explicit uncertainty, reproducible runs — not a trade call.