How it works

What the screen measures, how the portfolios are built from it, and what the numbers on this site mean. No company is judged on its story or its prospects. The rules run twice a year and the output is published as it comes.

The universe

The 500 largest US companies by company size, minus banks, insurers and utilities — not because they are bad businesses, but because their accounts work differently and the cash measures below produce nonsense for them. A company that stops filing accounts stops being scored: the last published figures would otherwise stand in for a business nobody is reporting on.

What gets measured

Three measures of whether a business actually produces cash, and one of how it behaves. Reported profit can be shaped by accounting choices; cash is harder to flatter.

The fourth is not an accounting ratio at all. Filing lag counts the days a company takes to publish, and it enters the ranking backwards — slower is worse.

A company slower than its peers is a company having trouble closing its books, and it is the only signal we found that is not already contained in the other three: its correlation with each of them runs between −0.07 and +0.10. It is read off the dates on the filing, not the figures inside it, so a screen built on the accounts alone cannot see it at all. It is also what makes the tighter books work: without it, holding fewer companies made the results worse, not better.

From four measures to one list

Each company is ranked on all four measures and the ranks combined, with filing speed entering in the opposite direction because slower is worse. Companies are placed in order on each measure and the places added up, so one freak reading cannot swamp the other three. The precise weighting, the treatment of ties and the handling of incomplete filings are the parts that took the work, and they are not published.

The bottom 250 are then removed, leaving 250 companies. The three published portfolios are cut from the top of what remains, at three depths. Each tighter book is a strict subset of the wider one — it contains no company the wider book lacks — and we would rather say so than let anyone discover it.

How many companies

Three books, at 25, 35 and 45 companies. Each number came out of the same test. The method was built using 1998 to 2014. The years 2015 to 2026 were set aside before any of it began and looked at once, at the end. Here is what every book size did on each half.

Read the two columns against each other. On the years the method was built from, the fewer companies the better — an eight-company book beat the index by 12.4 points a year. On the years it had never seen, that ordering reverses: the same eight-company book finished behind the index, and the fifty-company book did best of all.

Small books look brilliant on the data they were designed against and fall apart on data they were not. Everything under twenty companies failed that second test, which is why nothing smaller than 25 is published here.

Position sizing, and the second lever

By company size, uncapped. That is not a stylistic choice: equal weighting and square-root weighting both lost to the index in the years held back for testing, at almost every book size. The largest holding can run to a fifth of a book. Capping it would cost return, so it is left alone and said out loud instead.

Broad works differently, and here is why. Spreading money over more companies is the obvious way to calm a portfolio, and it runs out of road quickly: weighted purely by company size, a 35-company book swings 17.3% a year and a 45-company book 16.7%. Six tenths of a point for ten more holdings.

So Broad adds a second step. Each position is still sized by how big the company is, then divided by how much its share price has moved about over the past year. The same companies are held; the jumpy ones are just held in smaller amounts.

That takes 2.4 points off the year-to-year swing and costs about 1.7 points a year of return. A fair trade, and the reason it is offered as a separate setting instead of applied to everything.

The blunter version does not work. Dropping the most volatile companies from the list altogether — or the ones that move most with the market — calms the portfolio further and destroys what it is calming: on the years set aside, Concentrated goes from +5.23% ahead of the index to 1.27% behind it. Choosing steadier companies throws away the thing the screen is finding. Choosing steadier position sizes does not.

Rebalancing

Twice a year, in January + July. The particular months matter far less than they appear to. Running the identical rules on all six possible pairs of rebalance months moves the annual result by 2.8 points on Concentrated, 2.7 on Balanced and 2.1 on Broad, and all eighteen combinations beat the benchmark — the weakest of them by 4.2 points a year. Holding fewer names is what breaks this: at twelve companies the spread is 4.4 points and at six it is 5.7, which was the failure of an earlier version of this method. It held four to six names, and what looked like a strategy was substantially a choice of months.

One period, worked through

Every headline figure on this site is built from the same small calculation, repeated 57 times. Here is one of them in full, using Balanced over the six months from January 2025.

1. Set the book. On the last trading day of January 2025 the screen ranks the 500 companies on the four measures and keeps the top 35. Each is weighted by company size: a company worth 4% of the book's combined market value gets 4% of the money.
2. Hold for six months. Nothing is bought or sold until the next reset. Each holding's return over those months is its price on 31 July against its price on 31 January, with dividends included, weighted by its share of the book. Those weighted returns add up to +26.05%.
3. Take the trading cost off. At the reset, 45% of the book changed hands, and 0.20% is charged on the part that moved: 45% × 0.20% = 0.09%. Subtract it: +25.96% for the period.
4. Compare it to the index. The S&P 500 Total Return rose +5.64% over the same days. This period was a good one: most are closer, and in 22 of 57 the index won.
5. Chain the periods. Each period multiplies the last. $100 at the start of the record was worth $2,292 by January 2025; a further +25.96% takes it to $2,885. After all 57 periods it is $4,112.
6. Turn it into a yearly rate. The compound annual growth rate is the single steady rate that would take $100 to $4,112 over 28.5 years — that is 13.93% a year. It is an average, not a description: no individual period looked like it.

Every other figure comes off the same series. Volatility is how much those period returns vary around their own average. The worst fall is the largest drop from any peak in that chain to the lowest point after it. The difference against the index is this yearly rate minus the index's own, computed the same way over the same days.

Reading the figures

“11.63% a year”

The compound annual growth rate: the single steady yearly rate that would take you from the start value to the end value over 28.5 years and 57 six-month periods. Real years are nothing like steady.

“Worst 6 months: -27.55%”

The worst six months the portfolio had in the whole record. The usual measure is a maximum drawdown, which tracks the fall from a peak to the bottom day by day. This record only has two marks a year, so a drawdown computed from it would miss anything that fell and recovered in between and make the ride look smoother than it was. The worst half-year is the honest number this data can give.

The benchmark

“the S&P 500” is the S&P 500 Total Return index, with dividends reinvested and nothing deducted for fees or tax. It is measured with dividends reinvested and nothing deducted — no fund fee, no trading cost, no tax — which makes it a harder line to beat than any tracker you could actually buy. Our own figures are net of trading costs, so the comparison runs against us.

Costs

Every figure is after 0.20% on the part of the book that actually changes hands, and after a 30% loss booked on any holding that stops trading. In practice a reset means 40 to 80 orders, so at a typical European broker charging a couple of euros a trade the bill is roughly €100 to €160 a year — on a €30,000 portfolio, about half a point. Your own broker, and your own tax position, will differ.

Where this is weakest

Every item below inflates the historical record. Real results would have been worse.

• It is one path through history. 57 six-month periods from Jan 1998 sounds like a lot and is not: they cover the dot-com crash, the 2008 crisis, the long bull run after it and the 2022 fall. Four or five distinct market conditions, not 57 independent tests. A method that happened to suit those conditions would look good here.
• Backtests flatter. Rules tested on the past are chosen, consciously or not, because they worked on it. Published edges lose roughly a quarter of their return when tested on fresh data and more than half once the research circulates. We held eleven years back and looked once.
• The delisting penalty is an estimate. A holding that stops trading is booked at a 30% loss. Some are taken over at a premium and some go to zero, so 30% is a reasonable average and not a measurement.
• The edge is modest and the ride is not. The worst six months in this test was -27.55%. A few points a year is invisible next to that, and most people who abandon a method do it in exactly that moment.
• Small books look best and test worst. Over the full record a ten-company book returns more than any size published here. Over the eleven years set aside it returns almost nothing extra, and an eight-company book finishes behind the index. That gap is the clearest result in this whole project, and it is why nothing smaller than 25 companies is sold.

Two things that commonly flatter backtests are handled: the figures include dividends, and the universe includes companies that later went bust or were bought out, not just the ones still trading today.

Sources and timing

Company accounts and daily prices for US-listed companies, including ones that have since been delisted. A company's results are only used once they were actually filed with the regulator, plus a further 60 days of caution — using figures before they were public would let the backtest know things nobody knew at the time. Prices run to 31 August 2026.