From cognitive edge to structural edge
The project reorganised itself around access barriers rather than analytical effort.
Claim tested
That careful analysis of well-covered companies can produce a durable advantage.
Result
It cannot, at least not for me, and probably not for most people.
Consequence
The search narrows to small, rarely traded companies — where professional funds are structurally absent, not just outnumbered.
The distinction
There are two ways to be right about a security when the market is wrong.
The first is cognitive edge: seeing something in the available facts that others have looked at and misread. It is the more flattering of the two, and the one most people implicitly assume they have.
The second is structural edge: reaching a place where the relevant capital cannot go. Too small to size into. Too illiquid to exit. Excluded by mandate. Reported in a language nobody on the desk reads. Available only in a form nobody has aggregated.
Cognitive edge requires you to out-think people who do this professionally, full-time, with better data and a research budget. Structural edge requires only that you be small enough and patient enough to go where they are contractually or mechanically unable to follow. One of those is a repeatable condition. The other is a claim about yourself.
Cheapness alone is not a signal
A company can be cheap, uncovered and a trap, all at once. The three states are distinguishable, and the distinction is the whole game.
Nobody has done the work. The facts support a higher value and there is no offsetting reason for the discount other than obscurity.
People have done the work and drawn the wrong conclusion, usually by extrapolating a temporary condition or misreading an accounting artefact.
The market is right. The discount reflects a real and continuing impairment, and waiting will not help because the value is falling faster than the price.
The discriminating axis is the direction of value while you wait.
Not the size of the discount. If intrinsic value is compounding, time works for you and being early is survivable. If intrinsic value is eroding, time is the mechanism by which you lose, and no amount of statistical cheapness rescues it.
What follows from taking this seriously
Coverage becomes a screening criterion in its own right. A company followed by nineteen sell-side analysts has already had every publicly available fact reassessed by people who are paid to do it. The reassessment loop closes before I can get near it.
Survivability precedes payoff. At the bottom of a cycle the question is not whether the cycle turns, but whether this particular balance sheet lasts long enough to see it. A company with near-zero debt and one with net debt exceeding its market capitalisation are not variations on a theme, they are different instruments.
Kill criteria must be falsifiable. A list of things that worry you is not risk management, because every item on it can be reinterpreted favourably at the moment it matters. A kill criterion is a specific, pre-committed, primary-document-checkable condition, agreed while you are calm.
Inversion, applied systematically. Do not only ask how this makes money; ask how it loses money, and then avoid that. Building the case against your own idea is unpleasant, which is precisely why it is informative.
Founding premises
Every design decision in the specification descends from four conclusions reached before it was written. If a future change contradicts one of these, the premise should be revisited explicitly rather than eroded silently.
- Edge is structural, not analytical. Analytical capability is commoditising. What persists is access to places institutional capital cannot go: too small, too illiquid, too foreign, too unaggregated. The system's job is to point attention there.
- Numbers from a database, reasoning from the model, judgment and decision from the human. Each layer has a fixed role. Numeric data is never sourced from an LLM. Judgment is never outsourced entirely. The final call is never automated.
- The forward decision record replaces backtesting. A concentrated, multi-year, judgment-driven strategy produces too few independent decisions for historical validation, and its inputs cannot be reconstructed point-in-time. The journal is the validation mechanism.
- Approximately right beats precisely wrong. If a thesis requires a model to work out, it is too complicated. Coarse ranges, decomposed scores, no composite ranking.
The uncomfortable corollary
If the edge comes from inaccessibility, then the opportunities will be in places that are genuinely tedious to reach. There will be no API. There will be no analyst deck summarising the last five years. The filings will be PDFs on a company website with inconsistent naming, and somebody will have to read them. The reason the opportunity exists is precisely that this is annoying enough that institutions with better things to do have not bothered.
That turned out to be an extremely accurate prediction of how the rest of this project went.