Every input is public and keyless: Robinhood’s unauthenticated endpoints for tradability and prices, SEC EDGAR’s XBRL frames for financial statements, and FRED’s open CSV series for the macro overlay. Nothing is bought, licensed, or estimated by a model.

The composite is five pillars. Inside each pillar, every metric is converted to a rank across the whole screened universe, turned into a z-score, and then had its sector median subtracted — so a company is measured against its peers, not against whichever sector happens to be cheap this quarter. Momentum is the deliberate exception: it is never sector-neutralised, because industry momentum is part of the signal.

Missing data is never treated as zero. If a metric is absent, it is dropped and the remaining weights inside that pillar are renormalised. The three penalty metrics — accruals, asset growth and share issuance — are the exception: they are imputed at neutral and keep their weight, so a company cannot dodge a dilution penalty simply by not reporting. Companies below the coverage floor are excluded outright rather than scored on thin evidence.

The market temperature gauge on the board does exactly one thing: it moves how much weight valuation carries, between 10% and 30% of the composite. When the broad market is historically expensive the screen becomes more price-sensitive; when it is cheap, growth and momentum get more say. It never scores an individual company.

The construction is a building-blocks decomposition rather than a fitted model, for one reason: every term is separately inspectable and separately arguable. If you disagree with the number, you can find the term you disagree with, and the board shows you all five in the expanded row of every company. Nothing is a black box, and nothing is regressed.

E[r] = cash yield + faded growth + valuation reversion − dilution − market drag

• Cash yield — what the business currently throws off per dollar of its price: the tax-normalised earnings yield and free cash flow to equity, averaged when both exist, because they disagree for real reasons and the midpoint is steadier than either. Banks and insurers are earnings-only: their operating cash flow includes premium float and deposit flows, money that belongs to policyholders and depositors rather than shareholders. Capped at ±15%, since a higher figure is usually a non-repeating year rather than a P/E under seven.
• Faded growth — trailing three-year revenue growth, halved, capped at 25%, then decayed toward a 4% terminal rate that stands in for nominal GDP. The halving is not conservatism for its own sake: realised growth persistence beyond about two years is close to zero in the data, so carrying a 60% trailing rate into a five-year number would be indefensible. The longer the horizon, the more weight the terminal rate takes.
• Valuation reversion — half the gap between the company’s own value yield and its sector’s median, assumed to close over the horizon and amortised across it. Multiples do revert, slowly and incompletely, which is why it is half the gap and not all of it. The annualised contribution is clamped at ±8%, because a company that looks ten times cheap is usually a broken business or a broken data point, not a 60%-a-year tailwind. A sector with fewer than ten names gets no reversion target at all rather than reverting toward a two-company median.
• Dilution — share-count growth is a direct per-share drag and buybacks are the same thing in reverse. Halved and clamped at ±3%, because one year’s buyback is a weak predictor of the next five.
• Market drag — a uniform haircut derived from the level of CAPE against its long-run median of 16.2, assuming a quarter of the gap closes over ten years. It is identical for every company, so it cannot change the ranking; it exists so that the published level is not misleading while CAPE sits near 41.

Then it is shrunk, on purpose

Each term above is individually bounded, but they stack. A cheap insurer can legitimately collect the cash-yield cap, a full growth term, the re-rating cap and a buyback credit all at once and arrive somewhere north of 30% a year. No liquid large-cap earns that for five years running. The statistical reading is that a single-stock expected return estimated from trailing data is a very noisy measurement of a quantity with low genuine cross-sectional dispersion — the classic case for shrinking an estimate toward the mean (James–Stein; Blume did the same for betas in 1971, Vasicek formalised it in 1973).

So the published number keeps 55% of the company’s own estimate and takes 45% from a market prior, where the prior is the risk-free rate plus a deliberately modest 3.5% equity risk premium, plus the same market drag. Because the identical shrinkage is applied to every name, it does not change the ranking at all — it changes only the levels. Two companies in the same order before shrinkage are in the same order after it. What it changes is how confident the number looks, and that is the point.

Why the ranking is a ratio, not a return — and why it is a Sortino

The board is ordered on (E[r] 5y − risk-free) ÷ downside deviation, then tilted by durability. Ranking on raw expected return would put the most violent names on top every single night, because the terms that produce a high estimate (a large cash yield, a big gap to the industry median) are exactly the terms that come with a wide dispersion of outcomes. Dividing asks a more useful question: how much expected compensation is there for each unit of the discomfort you have to hold to collect it.

Downside deviation rather than total volatility, because a Sharpe ratio punishes a company for rising sharply exactly as hard as for falling sharply. Only returns below the target enter the denominator here, so upside surprise is no longer treated as evidence of risk. That single change is most of what stopped the screen being structurally hostile to growth.

The durability tilt is ±25%, and it moves the order only. A name scoring 100 on durability has its ranking statistic multiplied by 1.25; one scoring 0 is multiplied by 0.75. A negative statistic is divided instead, because multiplying a negative number by 0.75 would move it up the board — a fragile company rewarded for being fragile. Dividing keeps the tilt monotone across the whole range and continuous at zero, so nothing can be moved by a discontinuity rather than by its score. The E[r], σ, band and Sortino printed in every row are the measured figures; the tilt is never folded into them.

σ is annualised realised volatility computed from 251 daily closes — the standard deviation of daily log returns, scaled by √252. It is measured, not modelled, and it is entirely backward-looking: a company that has been quiet for a year and is about to stop being quiet will look safer here than it is. One trailing year is a short window and it is the least stable input on the page.

The one-sigma band, and why every number carries one

A point estimate published on its own is the single most misleading way to present this kind of arithmetic. Every expected return on the board therefore ships with a one-standard-deviation band around it, on the annualised return. The band narrows as the horizon lengthens — averaging over more years disperses the annualised figure by √h — and that narrowing means only that arithmetic, not that the model knows more about year five than about year one. Even the narrowest band on the board spans outcomes a reasonable person would describe in completely different words.

The sector-diversified basket

The panel above the board walks the ranked list from the top and takes the best-placed name in each sector until it has twelve, one per sector, then weights them by inverse volatility so each contributes a comparable share of standalone risk. Equal capital weighting would hand a 70%-volatility name several times the risk budget of a 20%-volatility one.

Its volatility is computed as √(w^TΣw) from the actual covariance of the holdings’ daily returns — not as a weighted average of their individual volatilities. Those two numbers are very different, and the gap between them is precisely the diversification benefit: the weighted average is what you would get if every holding moved in lockstep. Publishing the average would silently overstate the risk and make the exercise pointless. The covariance is estimated from a single trailing year and names without enough overlapping history are dropped rather than padded, since padding with zeros would fabricate a period of perfect calm and understate correlation.

The basket is a mechanical illustration of the model under a diversification constraint. It is not a recommendation, it is not a portfolio anyone is being told to hold, and no view about any company in it is expressed by its presence there. Nothing selected these twelve companies except their position on a ranked list and the sector constraint.

What is not in it

There is exactly one forward-looking input, and it is thinner than it sounds. Robinhood's earnings endpoint publishes consensus EPS for the next one or two quarters without a key, and the PEG metric uses it — so the flat claim this page used to make, that no forward estimate appears anywhere, was wrong for most of the board. No guidance, no revision data and no price target is used; everything else is a trailing figure from a filing or a price series.

The thinness is the part worth knowing. Across the universe 1,542 companies carry a single estimated quarter against 411 with two, and a lone quarter is a fragile thing to rank on — some of those rates are large enough to move a position by themselves. So consensus growth is credibility-weighted toward the company's own trailing three-year earnings CAGR: w = n ÷ (n + 1), where n is the number of estimated quarters. One quarter carries half the weight, two carry two thirds. A genuine acceleration still moves the number; it simply has to be larger to move it as far. Where there is no trailing record to temper it, a single quarter is refused outright and the metric is left missing rather than guessed. Rows using consensus are marked peg_forward, and those resting on one quarter also carry peg_thin_consensus.

Every other input is a trailing figure from a filing or a price series. There is also no backtest, and there will not be one: with keyless data there is no point-in-time universe and XBRL figures are restated after the fact, so any historical result would mostly measure hindsight. The parameters were chosen from published research and stated reasoning rather than fitted to outcomes.

What exists instead is a forward record. One basket was frozen on the day it was published, is held without rebalancing, and is measured every night against the S&P 500 total return — dividends reinvested on both sides, so the comparison is like for like. It cannot be curve-fitted, because it was fixed before the outcome existed; its weakness is the opposite one, that it is short. Below twelve months this site publishes cumulative totals only and withholds every annualised and risk-adjusted figure, because on a record that length they are arithmetically defined and meaningless — a good first fortnight annualises into the thousands of percent. Until that record is long, read the board as an ordering device under stated assumptions, not as a prediction.

1. Screen. Robinhood’s own instrument flags decide what is tradable; then a $500M market-cap floor and a $5 price floor. Roughly 2,500 companies survive.
2. Direction. Metrics where lower is better are sign-flipped, so from here on higher is always better.
3. Rank to z. Each metric is percentile-ranked across the universe and converted through the inverse normal, clamped to ±3. Ranking is the outlier control — no separate winsorisation is applied or needed.
4. Sector-neutralise. For sectors with at least 30 companies, the sector median is subtracted. Small sectors and “Unknown” are left alone, because a twelve-name median is noise. Momentum is never neutralised.
5. Assemble pillars. Missing reward metrics are dropped and their weight redistributed; the three penalty metrics are imputed at neutral and keep their weight. A pillar is invalid if under a third of its real weight survives.
6. Weight by regime. Valuation’s weight is multiplied by the macro multiplier; the rest are rescaled to keep the total at one.
7. Score. The composite is percentile-ranked among all ranked companies and expressed 0–100. Ties break on coverage, then market cap, then ticker. The top 200 are published.