One point per company, at its own most recently filed quarter. Where the US market sits right now.
Results commentary lifted from 2,954 companies’ own 10-Q and 10-K filings — searchable across 11.3M characters of what they wrote about their own quarters.
Quality gates applied to the chart and the table below.
Selecting a sector highlights it against the rest of the market rather than colouring all twelve at once — at this density twelve hues are neither readable nor colour-blind separable. Use the panels below to compare sectors.
A composite is only worth having if a portfolio built from it beats what you would have got by picking at random. That is a testable claim, and this is the test.
Every company gets a percentile on value, quality, cash generation, balance sheet and growth — compared only with its own sector, so a software company is never called expensive for not looking like a utility. The five combine into one score out of a hundred.
The ten highest scorers in each of the twelve sectors, about 120 companies. Equal amounts in each, so one large holding cannot decide the outcome.
Bought on the first trading day of January, sold twelve months later. No trading in between, no reacting to news.
Beating the S&P would only show these companies are smaller than the S&P. So each basket is compared with 2,000 random baskets of the same size, drawn from the same companies on the same day. That asks the real question: was the choosing worth anything? The S&P is drawn too, because the answer to that turned out to matter.
Over the seven years the basket returned +146%. The companies it drew from returned +94%, so the score really did pick better than the pond it was fishing in — about 3.9 points a year better, and in four of the seven years by more than luck can account for.
The S&P 500 returned +206% over the same seven years. $10,000 became $24,600 in the basket and $30,600 in an index fund — both in dollars, since a sterling investor would also have worn the exchange rate, which is not modelled here. The basket beat the S&P in two years of seven, and trailed it by 3.6 points a year.
Both are true at once, and the second is the one that decides anything. The score adds value relative to its universe. That universe — the smaller end of the US market — was a worse place to have been than the mega-caps carrying the index, and picking well inside it did not close the gap.
Read it this way. Each marker sits at what the basket actually returned; the grey band behind it is where nine out of ten random baskets of the same size landed. A filled marker is a year that finished clear of that band; a hollow one is a year the score cannot be told apart from luck, whether it landed inside the band or below it. The score cleared the band in four of the seven years, and the median year sat at the 98th percentile of random selection — something that happens by luck about twice in ten thousand tries.
2020 was bad, not just unlucky. The basket returned 5.8% while the average random basket returned 15.8% — the 3rd percentile. A method that wins by ten points in four years and loses by ten in a fifth is harder to live with than one that wins by four every year, even where the averages match. Across all seven years the average advantage is +3.8 percentage points, and that average is not statistically significant (t = 1.21).
The method was designed while looking at this data. Using return on capital employed rather than ROIC, capping quality, capping growth at 50% — each of those was decided by examining this universe and fixing what looked wrong. The data in this test is out-of-sample; the method is not. The only cure is to fix the rules now and test them on years that have not happened yet.
What is genuinely solid. The universe at each January is rebuilt from companies that had a price and had filed accounts on that day — including the ones that have since delisted, which most backtests quietly drop. Two thirds of the companies in the price history no longer trade. Leaving them out is the single most common way a backtest flatters itself.
Operating companies only. Financials and Real Estate are scored on a separate, weaker model and are listed below them — the two are not comparable.
Scored on P/E, ROE, growth and payout only: free cash flow and enterprise value are meaningless for them. A leveraged mortgage REIT can look cheap on P/E while being the most rate-sensitive thing in the index.
Same axes throughout. Each panel shows one sector against the full market in grey, so the panels are directly comparable.
Sorted by market cap. Blank means not computable from the filings — never zero.