Method and evidence standard

The Equity Research Lab publishes company dossiers on the AI buildout and the software built on it. Each dossier is an investment memo whose numbers come from an offline pipeline and whose judgments are written by hand, kept apart, and labelled. This page states the rules the lab holds itself to and what the mathematics can and cannot say.

What a dossier is#

A dossier has two parts that never mix. The memo carries the situation, the thesis, where consensus comes from and where this view differs, the key drivers, the disconfirming evidence, what would change the view, and what is being monitored. The artifact carries every number: the reverse discounted cash flow, the scenario tree and its return distribution, the factor exposure, the drawdown and value-at-risk record, the expected move, the relative valuation, and the capital-cycle ratios. The site renders the artifact's figures between the memo's sections. A memo can exist before its artifact does; it is then marked in progress and shows no numbers.

The evidence standard#

1. Pinned inputs. Every figure traces to named inputs with their SHA-256, coverage, and vintage, and to the commit of the code that produced it. A reader who checks out that commit and fetches the same public sources reproduces the artifact.
2. Disconfirmation first. Each memo states what would prove the thesis wrong before it argues why the thesis might be right, and records the evidence against it in its own section.
3. Judgments labelled as judgments. Probabilities, discount rates, terminal assumptions, and peer sets are tagged as the author's inputs wherever they appear, apart from filed and market figures, which are tagged by their source.
4. No price targets. The lab publishes what a price implies and what stated scenarios would pay. It publishes no target, no rating, and no recommendation.
5. A dated record. Thesis changes are logged with their date, version, and reason. The research ledger is derived from those entries, never written after the fact.

How the numbers are produced#

The pipeline runs on the author's machine. It reads public filings from SEC EDGAR, the daily factor files from the Kenneth R. French Data Library, rates and exchange rates from FRED, and daily prices downloaded for personal use. It computes the seven analytics, downsamples any series to at most four hundred points, and writes one artifact per company with a manifest of its inputs. Raw prices, factor files, and option chains never leave that machine; the artifact carries their hashes and vintages instead. The pipeline refuses to publish from an uncommitted or unpushed tree, so the commit named in every artifact is the code that ran.

What the mathematics can and cannot say#

A reverse discounted cash flow says what growth and margins a price requires under a stated cost of capital; it says nothing about whether those requirements will be met. A scenario tree says what stated branches would pay and how the payoff is distributed when the branch parameters are uncertain; the probabilities are the author's, and the distribution is only as honest as they are. A factor regression says how much of a stock's movement the market and the standard style factors explain over a sample, with the confidence intervals a short history deserves. A value-at-risk backtest says whether a risk forecast earned its confidence level in the past. Relative valuation says where a company sits among a stated peer set on a stated date. Capital-cycle ratios say what the filings show about how much capital the business consumes and what it earns on it. None of these is a forecast. Together they bound what a reasonable person could believe at the price.

Coverage and the ledger#

Coverage follows the thesis: the companies building the buildout and the companies being built on it, the well known and the emerging. A company enters the coverage ledger as a watchlist row, moves to in progress when its memo is being written, and is published when its artifact exists. Rows that are not published carry no numbers. The research ledger lists every dated thesis change across all dossiers, so the record of what the lab believed, and when, is public.

Limits and licensing#

Short listing histories widen confidence intervals and can make a rolling backtest impossible; the artifact then says so with an explicit reason rather than a shrunk window. Filed figures follow the latest filed value, which is right for a valuation today and wrong for a point-in-time study, and the artifact states that policy. Factor files arrive one to two months behind prices, and the artifact publishes both end dates. Market data are used under personal-use terms and appear on the site only as derived statistics, single as-of values, and transformed, downsampled series.

Disclosure#

Everything in the lab is personal research and education. It is not personalized investment advice, and nothing here is a solicitation. Coverage follows the thesis, not a portfolio.