Sanctuary Research
ResearchBriefingsMethodologyOpen the desk
Sanctuary Research

How Sanctuary scores a name
(and what we refuse to claim)

Our system is built to surface probabilities, not promises. Here's how each piece works.

01 — Score

Nine academic factor models, one composite

Every name in the large-cap universe is scored by nine academically-documented factor models — value, momentum, quality, profitability, investment, size, volatility, and more. Each model is evaluated on the stock's own earnings clock, the period between its earnings announcements, so that timing matters.

The composite score weights each model by its walk-forward-validated predictive power. A model that only worked in backtests gets a smaller vote than one that has continued to work on data it never saw during development.

02 — Context

Grounded macro, challenged claims

Macro context is anchored to FRED data releases and prediction-market odds — not sentiment or headlines. When the model references GDP growth, inflation expectations, or rate probabilities, those numbers come from the Federal Reserve's own datasets or from markets where real money is on the line.

Every sector-level stance passes through a challenge layer that looks for stale evidence and forces disputed claims to earn their place or get struck from the thesis.

03 — Debate

Four models, three rounds, a quorum gate

Four frontier AI models argue each candidate position over three rounds of structured debate. Their job is to break the case, not make it — they surface risks, contradictory evidence, and edge cases.

Consensus is required: a quorum gate rejects any name that cannot survive the critique. If the models cannot agree that a position holds up under scrutiny, it does not make the book.

04 — Monitor

Alerts, not auto-trades

Once you build a position, the system monitors for drift, drawdown, concentration, and thesis-break events. When something triggers, you receive an alert with the reasoning.

You approve or reject every rebalance. Nothing trades on its own. The system proposes; you decide.

What we refuse to claim

Probabilities, not predictions

  • We do not claim to predict the market. We publish odds on which names have historically been associated with better outcomes under similar factor conditions. Those odds can — and do — fail.
  • We do not claim historical performance guarantees future results. The composite-decile statistics we cite (e.g., top decile outperformance) are historical, out-of-sample research results. They describe what did happen, not what will happen.
  • We do not place trades. Sanctuary is a research desk, not a broker, advisor, or fund. You act on your own judgment with your own capital.
  • We do not pretend AI is infallible. The debate system is a filter, not an oracle. Models can share blind spots, hallucinate citations, or converge on a flawed consensus. The gate catches many errors; it cannot catch all of them.
  • We do not optimize for engagement. The goal is to surface useful research, not to keep you clicking. We'd rather lose a subscriber to boredom than cost one real money through hype.
Historical statistics

What the decile numbers mean

When we cite composite-decile returns (e.g., "top decile averaged 15.6% annually"), those figures come from an out-of-sample validation period — data the models did not see during development. Still:

  • Past returns do not predict future returns.
  • Decile membership changes monthly; any name can move between deciles.
  • The validation period covers a specific market regime. Different regimes may behave differently.
  • Transaction costs, taxes, slippage, and timing are not included in the decile statistics.

Treat these numbers as evidence that the scoring system captured something real in the validation window — not as a forecast of your personal returns.

Ready to see the research in action?

Explore the research Read the daily briefing Open the desk

Sanctuary Research provides quantitative investment research and educational content — not personalized investment advice. Markets involve risk; do your own due diligence.