The map · The tests · Chapter 13 of 18

What the tests removed

The claimA rule that sounds right is not evidence: every mechanical rule was tested on history under a bar set in advance, and most were removed.

About 8 minutes · the investor's route.

Everything in the chapters before this one was built from books, papers and careful reasoning. That is how most investment methods are built, and it is not evidence. In November 2026 the method was put on trial: twenty-six attempts on history, each one registered before it ran, every result logged, the failures too.

The rules of the trial

A test was written down first: what is claimed, which data, exactly what is measured, the bar, and what happens to the rule on either outcome. Then it ran once.11The attempt log, reference/backtest/ATTEMPTS.md in the private repository, one commit per registration ahead of the commit with its result. The public summary is reference/backtest/WHAT-WE-TESTED.md. A rule counted as holding only if it helped in both markets, the Nordics and the US, and in both halves of each period. From the last round it also had to beat what a thousand random relabellings of the same names give. Every look was counted, because after fifty tries one pass is what chance produces.22Harvey, Liu and Zhu, '... and the Cross-Section of Expected Returns', Review of Financial Studies 29 (2016): with many tests, a t-statistic above 3.

The history ran from 2016 to 2020 in the US and 2010 to 2021 in the Nordics, with every figure entering only on the day it became public, and with the companies that later went bust or were taken over put back in. A test on today's list of companies forgets exactly those, and our first run, which could only price survivors, flattered the method badly.

Decided: A rule that does not hold is removed from the code, not kept as a habit.
Why: Ruling E164: a rule about what is bought, held, sold or valued enters or leaves the method only on such a test. A rule about how a figure is read may be added freely.
Cost: The method lost most of its mechanical parts in ten days, including some the owner liked.

What went

The price stop. No stop level helped, with or without buying back, and with trading costs every one cost money. The exit is still written down before buying; for this owner it reads "no price stop".
The 15-50 percent band of chapter 4. The names inside it did no better than the rest.
The screener's ranking key. It did not sort later returns by decile in both markets. The screener now lists the names that pass in a fixed random order: the filters select, and the order no longer pretends to know.
Three hard kills: falling revenue, a squeezed margin, swelling receivables. They did not separate the names that did well from those that did not.
The debt limits, the quality tests as a drop rule, and the buy-price cushion as a binding line. None separated returns. The quality tests and the cushion are still printed, as information.

Open a rule to see what it claimed and what the test found. A rule held only if it helped in both markets and in both halves of each period.

Removed 16

The price stopattempts 10 and 18 · ruling E162

It claimed: Selling at a fixed fall under the buy price limits the damage.
The test found: No stop level helped any selection, in either market or half. Without buying back, every stop cost heavily. With real trading costs even the mildest one lost money.

The 15 to 50 percent bandattempts 8, 12 and 21 · ruling E161

It claimed: The bargains are among companies that have fallen 15 to 50 percent from their high.
The test found: The companies inside the band did no better than the ones outside it, in either market.

The ranking of the screenattempts 8 and 24 · ruling E167

It claimed: Operating profit against assets, added to operating profit against the price of the whole business, puts the best companies first.
The test found: It did not sort later returns by tenths in both markets, and its best fifth trailed the index. The list is now in a fixed random order.

Warning sign: falling salesattempts 17 and 24 · ruling E163

It claimed: Two quarters of falling sales end the reading.
The test found: It did not separate the companies that did well from the ones that did not. It is now information for the growth view.

Warning sign: a squeezed marginattempts 17 and 24 · ruling E163

It claimed: A sharp fall in the operating margin on flat sales ends the reading.
The test found: It did not separate returns.

Warning sign: swelling invoices and stockattempts 17 and 24 · ruling E163

It claimed: Unpaid invoices or stock growing much faster than sales ends the reading.
The test found: It did not separate returns.

Warning sign: debt above 3.5 times earningsattempt 25 · ruling E168

It claimed: A company that deep in debt is uninvestable.
The test found: It did not lift the average outcome. Whether it trims the worst losses is the open question below.

The screen's debt limitattempt 25 · ruling E168

It claimed: Net debt of at most 2.5 times yearly operating earnings.
The test found: It did not separate returns.

The screen's falling-sales testafter round 2 · with ruling E163

It claimed: A company whose sales fall two quarters running is not worth reading.
The test found: It went with the warning sign it copied.

Lower than a year agoattempt 25 · ruling E168

It claimed: Only read a company whose price is below where it stood a year ago.
The test found: It did not separate returns.

The run of falling quartersattempt 25 · ruling E168

It claimed: Three quarters of falling sales in a row is a soft warning.
The test found: It did not separate returns.

The quality tests as a reason to dropattempt 25 · ruling E168

It claimed: A company that fails the first three quality tests is dropped, or sold at its base value.
The test found: The tests did not separate the companies that did well. They are still printed, as information. This is the rule that sold CENTA.

The fourth quality test as a step in confidenceattempt 25 · ruling E168

It claimed: A buyer of companies whose purchases earn too little gets a larger cushion.
The test found: It did not separate returns.

The cushion as a binding buy lineattempt 25 · ruling E168

It claimed: Buy only at or under 85 percent of the value.
The test found: Companies bought under the line did not do better in both markets and both halves. The line is still computed, as information.

Two green flagsattempt 25 · ruling E168

It claimed: Cash flow at least as large as profit, and a rising margin, each raise confidence in a company.
The test found: Neither separated returns. They are out of the score.

The planned cut to growthattempt 23 · ruling E160, struck

It claimed: Growth derived from a company's own history is too high and should be cut.
The test found: It was not too high: the growth that followed was above the derived figure. The cut was struck before it was ever used.

Tested and kept 4

Three earnings misses in eight reportsattempt 17

It claimed: A company that keeps missing its own numbers is one to leave.
The test found: The only warning sign with data that separated returns. It stays.

At most six percent in one companyattempt 11 · ruling E157

It claimed: Spreading the money matters more than picking the best three.
The test found: Fewer names did not raise the typical outcome. They only widened the spread: three names are a lottery ticket on the same average.

Growth derived from the company's own historyattempt 23 · ruling E144

It claimed: A company's own organic growth, capped, is a fair base for its future.
The test found: It did not overestimate what followed, on any date or in either half.

Positive spare cashattempt 25

It claimed: A business that does not produce cash cannot be valued on its cash.
The test found: It could not be tested: every company the method can value has it. It stays, as a condition of the arithmetic.

Could not be tested: done by hand 4

Cuts to the company's own forecastno data

It claimed: Two cuts in four quarters end the reading.
The test found: There is no history of forecasts to test it on. It stays, by hand.

Both top executives leavingno data

It claimed: The chief executive and the finance chief both gone within a year ends the reading.
The test found: No data. It stays, by hand.

The sentence on what the company sellsby hand

It claimed: If the company's own words cannot say what it sells, to whom, and why they come back, it is not valued.
The test found: No history can test a reading. Its test is the shadow book.

The belief written before the priceby hand

It claimed: A growth view written after seeing the market's number is void.
The test found: It cannot be tested on history. The same discipline was applied to these tests themselves.

Still open 2

Do the debt limits trim the worst losses?attempt 26 · April 2027, on new data

It claimed: The debt limits and the first three quality tests may cut the share of companies that lose most of their value, even if they do not lift the average.
The test found: Registered, not yet run. It is measured on the middle outcome and on the share of names losing more than three quarters, on dates no earlier test has read. If it holds, those filters come back.

The nine-point accounting scorea candidate, for new data

It claimed: A known score of nine accounting signals separates strong companies from weak.
The test found: It passed once, and later checks weakened it. It is not a rule. It waits for new data.

What stayed

The valuation itself; figures only from the companies' own filings; the growth view written before the price is seen; positive free cash flow; at most six percent of the portfolio in one name; and the hand parts no data can test: reading the reports, the issuer's own sentence on what it sells, the judgement that the business is understood. Their test is the shadow book of chapter 12, which prices every "no" against the index.

What the trial did not find

An edge. In the Nordics the cheaper companies by fair value beat the dearer ones in most of the years the method was developed on, but a check on two later years, registered in advance and run once, did not confirm it. Where the sort worked, it worked by avoiding the dearest names rather than by finding winners. Against four published value methods it won none of twelve comparisons. The honest summary is the one the README gives: a broad index fund is the base, and this method is the part that is studied, not trusted.

Faults in our own data

Half the work was finding where the history lied. Share counts restated for later splits beside unadjusted prices; report dates that were only period ends; figures in another currency than the share; one security under two tickers; a ticker reused after a takeover; a vendor's single bad day read as a company going bust. Each was found by reading full output or by an independent review, logged, fixed, and the affected runs redone as new attempts, never quietly.

The history is closed

After the last round the owner closed these years: the method is not trimmed further on data it has already been fitted to. One question is left open on purpose. The debt limits may cut the losses at the bottom even though they did not lift the average; that is registered now, measured on the median and the share of names losing three quarters of their value, and it runs only on new data, in April 2027.

Where to check this

The attempt-by-attempt account: reference/backtest/WHAT-WE-TESTED.md. The scripts: tools/backtest/, against the engine in vss/. The rulings: E160 to E169 in reference/archive/. The price and statement data came from a commercial vendor under a private licence, so its figures are not published; with your own licence the scripts reproduce them.

Try it

ls tools/backtest/

The test scripts. With a data licence of your own they reproduce the results, or show where they are wrong. From the repository's root, after the install in the manual.

In three lines

Each test was written down before it ran, and a rule held only if it helped in both markets and in both halves of each period.
The stop, the band, the ranking, three warning signs, the debt limits and the cushion as a buy line were removed.
No edge was found. The history is closed, and the next test runs on new data in April 2027.

Sources for this chapter

The attempts and their rule — read from this repository: reference/backtest/WHAT-WE-TESTED.md; tools/backtest/; reference/archive/FRAMEWORK-EDITS-2026-11-08.md E160-E169
Multiple testing
Harvey, Liu and Zhu (2016), Review of Financial Studies 29
Outside benchmarks
Ken French data library; Jensen, Kelly and Pedersen (2023), Journal of Finance, jkpfactors.com (CC BY-NC 4.0)

The slow way, made survivable. Built 2026-11-09 from the repository as it stood that day. NAME A and NAME B are fictional. One real company, CENTA, is followed through the book, with figures from its own filings and the owner's own trades. Not financial advice: one private investor's method, shared as it is.