Does the ranking predict returns?

How it was tested

Test enough ideas on one history and some look excellent by luck. The test was set up to prevent that:

1. 1Every idea, its source and its predicted result were written down and published in the project's history before it was run. For seven of them (each category on its own, marked †), the 2020–2026 result had already been computed in an earlier check of the data before they were registered, so for those only the 2012–2019 test is blind.
2. 2Ideas were compared with the current weighting month by month and judged on months they were never tuned on. Each comparison is paired by month and measured on the stocks the two rankings share, so a difference is a difference in ranking rather than in which companies were ranked. A weighting can only rank what it has enough data to score: over 2020–2026 the current weighting ranked 440 stocks a month on average and the narrowest alternative, Free-cash-flow yield alone (within sector), 348. The narrower screen ranks almost only companies the current weighting also ranks (fewer than one a month is unique to it), so pairing on the overlap holds the current weighting to that smaller set instead of crediting it with the 92 only it can score.
3. 3The years 2012–2019 were set aside as an untouched test, with the rule for passing it, before any of that data was built.
4. 4A result counts only if it clears a bar that allows for how many ideas were tried: p < 0.05 / 28 and a t-statistic of at least 3, the threshold Harvey, Liu & Zhu (2016) propose for a new factor. With p from Student's t, that means a t-statistic of at least 3.21 over 2012–2019's 96 months and 3.37 over the 38 held-out months. Every run is kept, including runs a data correction replaced.

Each alternative against the current weighting, 2012–2019

Difference in average monthly rank correlation. The shaded band is the smallest difference the test could reliably detect (80% of the time). The two short marks either side of zero are the pass bar: a difference counts only if it reaches beyond them, a t-statistic of 3.21 over these 96 months, in the same direction as 2020–2026. A dot inside the marks did not pass; it may still be a real difference smaller than the band. Rows marked † had their 2020–2026 result on file before they were registered; this chart is their blind test.

Size alone (within sector)† shows the largest 2012–2019 difference, +0.027 (t +1.83), short of the pass bar. Survivorship could explain it. The test can only include companies that are still in the index today, and among 2012's members the small ones that survived are disproportionately the ones that did well. This data cannot separate that explanation from the passage of time. The share of each month's members the test could include rose steadily, from 57% in 2012 to 97% in 2026 (its correlation with the calendar is 0.998), and within 2012–2019 the Size screen's monthly result moved with that share about as much as the current weighting's did (correlations of −0.124 and −0.135). Rebuilding the departed members from free filings and prices would test it; that rebuild is queued with its prediction written down in advance.

If you had kept the best backtest

A common way to choose a strategy is to backtest many and keep the winner. That rewards luck as well as skill. Bailey, Borwein, López de Prado & Zhu (2017) measure how much: split the history into equal blocks of months, and for every way of picking half of them (12,870 ways here, from 16 blocks), take the weighting that did best on that half and see where it ranks on the other half. The probability of backtest overfitting (PBO) is the share of those trials in which the winner lands in the bottom half. Luck alone gives about 0.5; a weighting that is genuinely better pushes it toward 0. This uses the current weighting and the 27 alternatives above, each scored month by month against the current one on the stocks both ranked, and reruns nothing.

Over 2012–2026 the PBO was 0.20, better than luck alone would give, and the winner beat the current weighting on the other half in 80% of trials. But one weighting did most of the winning: Size alone, in 68% of trials. Its lead over the current weighting was +0.027 over 2012–2019 and −0.015 over the held-out months, where it fell behind. That is the Size result above, which survivorship could explain.

Over 2020–2026 alone the PBO was 0.73, worse than a coin toss, and the winner beat the current weighting out of sample in 34% of trials. Drop Size alone from the set and the PBO over every year is 0.60, worse than a coin toss.

So a low reading over the whole history did not survive either check. Keeping the best of many backtests would mostly have picked whatever the history happened to reward, which is why the screener's weights come from published research and every test here is registered before it runs.

What a simple trading rule on the ranking would have done

A rule that holds the top 25, keeps each until it falls past rank 40, skip names flagged as value or growth traps, at most 6 in a sector, 5 basis points of cost a side, rebalanced monthly. It is compared with the equal-weighted average return of the stocks the test could score each month. Those are the members still listed today: 57% of each month's members in 2012, rising to 97% in 2026, so the average leans toward the companies that survived.

In 2012–2019 it returned 0.09% a month below that average (t −0.67), a gap within the noise. In 2020 – mid-2023 it returned 0.71% a month below that average (t −2.56), a gap larger than chance readily produces. In mid-2023 – 2026 it returned 0.22% a month below that average (t −0.55), a gap within the noise. Read any live record of the ranking next to an equal-weight fund.

The same rule with a 10% stop-loss

One pre-registered variant of the rule above: a holding whose adjusted close ends a day 10% or more below its price at the start of the month is closed out at that close, and its slot stays in cash until the next monthly rebalance. The construction is Han, Zhou & Zhu (2016), Taming Momentum Crashes: A Simple Stop-Loss Strategy (SSRN 2407199) — the one stop-loss rule with published support behind it. It was registered with its prediction before it ran: a less negative worst month and a shallower deepest fall in all three periods, the gap against the average scored stock about unchanged or slightly lower, and turnover higher.

The prediction half failed. The stop cut both the worst month and the deepest fall in 2020 – mid-2023, and made at least one of them worse in 2012–2019 and mid-2023 – 2026. Turnover rose as predicted, to 1.4x / 1.8x / 1.8x the unstopped rule's. The stop closed 470 positions across the periods shown. A stop exits into a fall and holds cash until the next monthly rebalance, so it trades the falls it escapes for the rebounds it misses, and the periods above disagree about whether that is a good trade. One stop level was tested, 10%, fixed before the test ran and never searched over — a level chosen by what this history happened to reward would tell you nothing.

How long would it take to prove an edge?

The ranking's monthly correlation swings by about 0.12 from one month to the next, so any monthly ranking of large companies needs years of record before a realistic edge can be told apart from luck. That applies to this screener's live track record too. Here, the smallest difference the test could reliably see was about 0.037 over 2012–2019 and 0.062 over the held-out months.

What this means

So the screener is a structured, checkable way to look at a company across eight dimensions. It is not a forecast.

Limits of the test

• It cannot include the analyst-based inputs (estimate revisions and forward earnings growth, about a sixth of the live weighting): no free history of them exists.
• Companies that have since left the index are missing, which matters most in the earliest years.
• Large companies only, rebalanced monthly, with costs approximated and no taxes.