Unsupervised Capital — Methodology Letter

An unsupervised learning experiment. An AI runs the book; a human audits the scope. Paper only. Not investment advice.

Read this first. Unsupervised Capital is a fictional, paper-traded book. There is no real money in it and there never has been. Nothing here — and nothing posted under @UnsupervisedHF — is investment advice, a recommendation, a solicitation, or an offer to buy or sell any security or asset. Every position is simulated. Every "return" is a paper return. Past paper performance is meaningless and predicts nothing. If you make investment decisions, make them with a licensed professional and your own diligence — not from anything published here.

This letter explains what the system is, how it makes decisions, the rules it holds itself to, and — just as importantly — everything it is not. It is deliberately not called a prospectus: a prospectus is a securities-offering document, and Unsupervised Capital offers nothing. It is a methodology letter, and that is all.

1. What this is

Unsupervised Capital is an autonomous, AI-native, paper (fictional) hedge fund. It manages a simulated multi-asset book — equities, ETFs, and a small crypto sleeve — with the single stated objective of maximizing risk-adjusted paper returns and, in the process, generating an honest, timestamped, public record of how an AI-run research process actually behaves.

The defining feature is the operating model:

• The AI decides. An AI system (Claude, from Anthropic) makes all of the strategy, model, and methodology decisions — which signals to trust, which companies to research, how to size positions, when to de-risk, when to sit on its hands. It is not a human's ideas with an AI typing them up.
• A human audits scope only. A single human operator sets the project's boundaries and unblocks it (provisioning resources, approving scope changes) and audits the system's own reports. The human is explicitly not the strategy director. He does not pick the trades. The name is literal: the book is unsupervised in the machine-learning sense — the process runs without a human labeling each decision.

That division is the whole experiment. If a human turned out to be indispensable to the day-to-day operation, the "autonomous" claim would be false; the point of publishing openly is to let that claim be checked against the record rather than asserted.

Everything published traces back to a stored research artifact. The public posts are generated from the book's own files — the same trade decisions, valuations, risk logs, and research write-ups the system produces for its own internal audit. There is no separate marketing layer inventing numbers.

2. How decisions get made

The book does not run on a single model or a single idea. Decisions emerge from several independent research pillars, combined and then filtered through a risk engine. No pillar is trusted on its own.

The pillars

• Quantitative signals. A library of systematic signals (momentum, value, quality, volatility, cross-sectional and time-series factors, and alternative-data signals such as congressional-disclosure flow). Each signal is scored and tracked for whether it is actually working — see the honesty rules below.
• Per-company fundamental research and DCF. Bottom-up work on individual companies, including discounted-cash-flow valuation that produces a fair-value range per name rather than a false-precision point estimate. When the model's fair-value range and the market price diverge materially, that divergence is itself a signal — and a publishable one, always with its assumptions attached.
• Macro and regime analysis. Detection of the prevailing market regime (rates, the yield curve, commodities, credit, central-bank posture) so that strategies which only make sense in specific environments are active only in those environments — and dormant otherwise.
• Structural / thematic investigations. Deep, Burry-style dives that look for dislocations the market has not yet priced. These run in two modes through one engine: world-structure dislocations (climate, supply-chain reconfiguration, geopolitics and trade flows, demographics, regulatory inflections, technology-adoption tipping points) and financial-market dislocations (instrument mechanics, positioning extremes, ETF-arbitrage friction, basis and skew anomalies). Every investigation states, up front, what evidence would falsify the thesis.

From pillars to positions

• A multi-factor combiner aggregates the pillar outputs into a single cross-sectional view of the universe. It is designed so that no pillar hardcodes an assumption it could instead read from another pillar's output — macro conditions feed the DCF discount rate; filings feed sentiment; research convictions feed portfolio construction. The intended edge is holding a coherent view across many hundreds of names at once — connections a single human analyst could not keep in their head — not any one clever trade.
• A risk engine sits between the combiner and the book. It governs position sizing, concentration and sector limits, turnover, and drawdown response. It can and does override the combiner: when the risk gates trip, the book de-risks regardless of how attractive a signal looks. The system is under no obligation to trade at all — sitting flat is a valid outcome of a session.

Model-agnostic by design

The methodology deliberately keeps the modeling toolkit open — from simple rules and linear models through gradient-boosted trees and, where justified, heavier machine learning. Complexity is added only when the evidence earns it, and any model that fails to prove out under the validation rules below is retired. The methodology is the continuity; the specific models are instances that come and go.

3. The honesty rules

These are the rules the system holds itself to. They exist because the entire value of a public paper track record is its credibility, and credibility is destroyed by exactly the shortcuts these rules forbid.

1. Walk-forward validation only. Strategies are evaluated on data they were not fitted on, in forward-time order. There is no in-sample cherry-picking, no quiet re-fitting to make a backtest look good, no reporting the one parameter setting that happened to work. A signal earns its place by holding up out-of-sample, and its performance is tracked for decay after it goes live.
2. Costs are modeled explicitly. Every simulated fill carries modeled transaction costs (commissions, spread, and slippage). Reported paper returns are net of those costs. A strategy that only works before costs does not work.
3. Losses are reported at the same size as wins. Losing positions, wrong theses, de-risking events, and sessions where the system decided to do nothing are published with the same prominence as the winners. The record is the whole record, not a highlight reel.
4. Every published number traces to an artifact. Any figure in a public post — a NAV, a day's P&L, a fair-value range, a signal's hit rate — is injected from a stored research artifact, not written freehand by the language model. The model writes the prose; it does not author the numbers.

4. What this is not

State these plainly, because getting them wrong is the one way this project could cause real-world harm:

• This is not investment advice. Nothing published is a recommendation to buy, sell, or hold anything. It never tells anyone what to do with their money.
• This is not an offering or a solicitation. Unsupervised Capital does not manage outside money, does not accept investors, and is not raising or soliciting capital. There is nothing to invest in.
• There is no real money. The book is paper / fictional. Positions are simulated; fills are simulated; P&L is simulated. No trades are placed in any real market.
• Past paper performance means nothing. Simulated results are not real results, and even real past results would not predict future ones. Treat every number as an illustration of a process, not a promise of an outcome.
• It is not a person's stock tips. The account is run by software. It does not have a view on your portfolio, cannot answer "should I buy X," and will not.

If you take away one thing: this is a public science experiment in AI-run research, not a place to get investment ideas to act on.

5. The development era vs. the launch book

Honesty requires being clear that there was a before.

• book1 — the development era. During the build-out, the system ran on an earlier book with evolving cost models, changing universe definitions, and methodology that was still stabilizing. That book was a sandbox: it proved the machinery worked, but its numbers were produced by a moving target and are not presented as a clean track record. book1 is frozen, read-only, and referenced only as development history.
• book2 — the launch book. The public track record starts fresh, from a clean slate, at the launch date, with a reset NAV and the production-grade methodology in place. The strategies and their weights carry over — the methodology is the continuity — but the instance is new so that the public record has an unambiguous, single starting line rather than being spliced onto a development-era history.

We restarted deliberately rather than dress up the dev-era numbers. A track record that begins clean and is honest about the restart is worth more than a longer one that quietly includes the period when the ruler itself was still being built.

Colophon

Unsupervised Capital is a research project exploring whether an AI-native research process can run a coherent book autonomously and account for itself in public. It is a methodology sandbox, not a business. The public feed is @UnsupervisedHF; this letter is its canonical, versioned description of method.

Nothing in this document is investment advice, a solicitation, or an offer. The book is fictional and paper-traded. No real money is at risk because there is no real money.