Quality Screening: The Factors Behind the Tessera Rating

TL;DR

• Cheap does not equal good. A stock trading at 8x earnings is a gift if the business is durable and a trap if the earnings are about to halve.
• Two different numbers get confused constantly, so we separate them here. Tessera computes a 47-factor point-in-time panel for every US-listed stock every night. The Tessera Rating — the label you see in the screener and on every company page — is a weighted composite of 17 of those 47. This page is about the 17.
• The 17 are not equally weighted, and three of them currently carry a weight of exactly zero: PEG, debt/equity and the accrual ratio were each measured signal-free on the tradeable universe and zeroed rather than deleted, so the model reduces to 14 factors that actually move the score. They stay in the manifest because a zeroed factor is still computed, still visible, and can be re-weighted if the evidence changes; deleting it would silently rewrite the panel's identity.
• Among the fourteen that do carry weight, the calibration is on each factor's historical rank correlation with forward returns, and the largest is more than ten times the smallest. Return on invested capital carries the largest single weight.
• Normalization is per-factor, not uniform: 9 of the 17 are z-scored within their GICS sector; the other 8 are ranked as percentiles across the whole market.
• The composite is not the last word. Separate hard gates — accounting-manipulation and accruals flags in combination, a bankruptcy-risk floor, a leverage ceiling, a minimum on how many factors actually have data — demote a name out of buy-eligibility regardless of how good the score looks.

Why quality matters more than valuation

Quantitative value strategies have one consistent failure mode: the value trap. You screen for low P/E, low P/B, or high dividend yield, and you systematically end up long businesses whose earnings are in secular decline. The multiple looks cheap because the market is pricing in deterioration the screen can't see. When the next earnings cut arrives, the "cheap" stock re-rates lower — not because sentiment shifted, but because the denominator shrank.

The academic literature on this is not subtle. Piotroski's 2000 paper on the F-Score showed that within the cheapest quintile of book-to-market stocks, the ones with weak fundamentals underperformed the ones with strong fundamentals by roughly 7.5% annually over a twenty-year window. Similar work by Sloan on accruals and Fama-French on the "quality minus junk" factor lands in the same place: valuation alone is a noisy signal, and quality is the filter that separates real discounts from dying businesses.

The practical implication for Tessera is that our sector-relative P/E ranking — the engine that actually surfaces candidates — is necessary but not sufficient. A stock trading at a 35% discount to its sector median is interesting. A stock trading at a 35% discount to its sector median that also has weak returns on capital, heavy leverage, and earnings running well ahead of cash is a falling knife. Quality screening is how we tell those two apart before capital gets allocated.

Think of quality as the floor, not the ceiling. It will not make you rich on its own — a high-quality business at 50x earnings is still a poor risk/reward. But it keeps you out of the cheap stocks you should not own, which is most of them.

The panel and the model are not the same thing

This distinction matters enough to state twice, because conflating the two is how factor counts drift.

The panel is what gets computed. Every night, for every US-listed stock we cover, Tessera materializes a point-in-time record of 47 factors — fundamentals, valuation multiples, discounted-cash-flow outputs, momentum and technicals, earnings-surprise history, insider activity, institutional ownership, and a set of accounting-integrity scores. Point-in-time means each value is stamped with what was actually knowable on that date, so a backtest reading the panel cannot see a restatement that hadn't happened yet. The panel is what the rating, the backtester, and the portfolio rules all read from; a subset of it is exposed as screener filters alongside fields that aren't factors at all, like sector and price.

The model is what gets weighted. The Tessera Rating is a linear composite over 17 of those 47 factors, three of which are currently at zero weight — so fourteen do the work. That model — the one described below — is what produces the rating on every company page, what the nightly whole-market score cache stores, and what the free auto-managed book scores its candidates on.

The other 30 panel factors are computed and visible but carry no weight in this composite. Several of them still do real work as gates or as inputs to other views; that section is further down, and we name them rather than leaving them implied.

The 17 factors in the model

Grouped the way the factor manifest itself tags them.

Profitability and balance-sheet quality (5 factors)

The question this group answers: is this business earning good returns on the capital it employs, and is the earnings number honest?

• Return on invested capital (ROIC) — the headline metric, and the largest single weight in the model.
• Gross profitability — gross profit scaled to total assets.
• Asset turnover — revenue generated per dollar of assets.
• Debt-to-equity — total debt against shareholder equity.
• Accrual ratio — the gap between reported earnings and operating cash flow, scaled to assets; Sloan's measure.

Why ROIC rather than ROE. ROE can be inflated by leverage: a business with mediocre underlying economics can show a 20% ROE by loading up debt. ROIC is leverage-neutral — it asks how productively the total capital base is deployed. That is why ROIC, not ROE, is the profitability factor that carries weight here; return on equity is not in the model at all.

Why accruals matter. Sloan's 1996 paper is one of the most replicated results in accounting research: firms with high accruals — where reported earnings meaningfully exceed operating cash flow — underperform low-accrual firms in long-short portfolios. The intuition is simple: when net income is running ahead of cash, something is being capitalized, accrued, or timed. Sometimes it's benign. Often it's not. The model scores this the way Sloan's result predicts: lower accruals rank better.

One direction runs against the textbook, and we would rather say so. Each factor's sign is set by the calibration, not assumed from the literature, and gross profitability came out of that process scored in the opposite direction from the well-known published finding that more gross-profitable firms earn higher returns. We have kept the sign the calibration produced rather than overriding it to match the paper, but you should know it is there before you read our ranking as an implementation of that literature. It is one of the things the next re-calibration will re-examine.

Valuation (4 factors)

The question this group answers: what are you paying for it, relative to the peers it competes with?

• P/E — trailing earnings multiple, z-scored within sector.
• PEG — earnings multiple adjusted for growth.
• Free-cash-flow yield — free cash flow against market value.
• Shareholder yield — dividends plus net buybacks against market value.

Every one of these four is normalized within sector, which is the point. A 9x P/E for a regional bank and a 9x P/E for a software company mean opposite things. Ranking each against its own sector cohort is the only comparison that carries economic meaning. The same logic drives the sector-relative P/E ranking that surfaces candidates in the first place.

Intrinsic value (2 factors)

• DCF upside — the gap between a discounted-cash-flow estimate of intrinsic value and the current price.
• DCF value creation — whether the business is generating returns above its cost of capital, which is what makes growth worth anything.

Why value creation is separate from upside. High growth at low returns on capital is value-destructive: if the business reinvests a dollar and earns less than its cost of capital on it, growth is making shareholders poorer, not richer. Upside tells you the price looks wrong. Value creation tells you whether the underlying business deserves the re-rating.

Growth (3 factors)

• Revenue growth — most recent period, year over year.
• 3-year revenue CAGR — the multi-year compounding rate, which is much harder to fake with one good quarter than a single year-over-year print.
• EPS growth — bottom-line growth per share.

Both a one-period and a three-year revenue measure carry weight deliberately. Lumpy growth — 40% one year, -10% the next — is usually cyclical or project-driven, and a single annual print is misleading in either direction. The multi-year rate is the one that separates compounders from one-off years.

Momentum (2 factors)

• 6-month price momentum
• 12-month price momentum

Momentum is in the model because the cross-sectional evidence for it is about as robust as anything in the literature, not because we think price predicts fundamentals. Both are percentile-ranked across the whole market rather than within sector.

Earnings surprise (1 factor)

• Surprise magnitude — the average size of the last four quarterly earnings surprises: reported EPS against the consensus estimate that stood before the report, scaled by the size of that estimate.

This is the model's only analyst-derived input, and it is worth being precise about what that means. There are no analyst ratings in the composite, no price targets, and no estimate revisions. What is used is the historical fact of how far actual results landed from the number the street had published — a realized outcome, not a forecast. The estimate-revision factors that once carried weight here were removed in July 2026 after an audit found their historical inputs were not reliably point-in-time; they are still computed in the panel, and they carry zero weight until a clean forward history has accrued.

How the factors combine

Normalization is per-factor, not uniform. Nine of the seventeen are z-scored within their GICS sector — a z-score of +1 means one standard deviation better than the sector median. The other eight are percentile-ranked across the whole market. Which treatment a factor gets depends on whether the metric is structurally sector-dependent. Debt-to-equity, P/E, and accruals are: a 40% debt-to-equity ratio is aggressive for a software company, normal for a utility, and conservative for a regional bank, so those are compared within sector. Price momentum and revenue growth are not sector-structural in the same way, so they rank against the full market. Copy that says every factor is sector-z-scored is wrong, and we have said it ourselves in the past.

Weights are not equal. Each factor's weight is calibrated on its historical rank correlation with forward returns, measured across a training window and re-checked on a held-out window that the calibration never saw. The spread is wide — the largest weight is more than twenty times the smallest — and ROIC carries the largest single weight. We do not publish the individual weights on this page because they are re-gated periodically and a number printed here would go stale; the per-factor contribution to any individual stock's score is visible in the product itself, which is the version that stays current.

The normalized, weighted factors sum to a single composite. On stock pages that composite is presented on a 0–100 scale as the Tessera Quality Score.

Gates that run independently of the score

A composite is an average, and averages hide catastrophes. So several checks run outside the weighted sum and can pull a name out of buy-eligibility no matter how strong its score is:

• Corroborated accounting-manipulation flag. An earnings-manipulation score and the accruals ratio must both flag before this fires. Either one alone is already partly reflected in the composite, and acting on one alone double-counts it.
• Bankruptcy-risk floor. A distress score below its threshold removes a name from buy-eligibility. It is a crude, dated formula, but it works well enough as a screen for names that deserve more scrutiny.
• Leverage ceiling. Debt-to-equity above the ceiling, or negative equity, blocks new buys.
• Minimum factor coverage. If too few of the factors actually have data for a stock, the score is not trustworthy enough to act on and the name is not buy-eligible. A thin score is a missing score, not a neutral one.
• Share-issuance and percent-accruals screens. On the portfolio side, candidates that have been issuing shares heavily over five years, or whose earnings run far ahead of cash, are removed from the candidate set before a book is built. Both screens block new entries and never force an exit.

Note the asymmetry: gates block buys, they do not force sales. Existing holdings are handled by the rotation and exit rules described in Position Sizing and Stop Losses.

The Tessera Rating in practice

The composite and the gates resolve into a five-level rating:

• Exceptional — buy-eligible
• Strong — buy-eligible
• Fair — not buy-eligible; existing positions are not force-exited
• Weak — not buy-eligible
• Avoid — not buy-eligible

Only Exceptional and Strong are treated as buy-eligible. Any gate above that fires on a name rated Exceptional or Strong pulls it down to Fair — which is what "the gates run independently of the score" means in practice.

Pairing this with the rest of the stack: a candidate has to be rated Exceptional or Strong, clear the gates, trade at a meaningful sector-relative discount, survive regime filtering, and win a competitive-rotation comparison against current holdings before capital gets committed. Even then, position size is capped, and the auto-managed books screen against a $1 billion market-cap floor.

Computed but not weighted

Twenty-seven of the 44 panel factors carry no weight in the model above. They are computed nightly and available in the screener, and some of them do real work elsewhere in the system — but none of them moves the Tessera Rating. Naming them is the honest version of "transparent methodology":

• Accounting-integrity scores — the Piotroski F-Score, the Altman Z-Score, and the Beneish M-Score. Two of the three are gate inputs, as described above. None is weighted in the composite.
• Income quality, ROIC trend, ROIC trajectory, FCF growth, revenue acceleration, margin trend — quality and trajectory measures. Some of these carry weight in other scoring presets available on the paid plan, which is exactly why a factor count quoted without naming the preset is meaningless.
• Technicals — RSI, 50- and 200-day moving-average ratios, distance from 52-week high, 1- and 3-month momentum, breakout score.
• Analyst-derived reads — recommendation counts, buy percentage, a fundamentals grade, and the estimate-revision factors described above. None weighted.
• Ownership and insider activity — institutional ownership percentage, 90-day insider buy count, 90-day net insider value.
• Five-year share issuance and percent accruals — both are gate-only by design, not composite inputs.
• Market cap, which is a universe filter rather than a quality signal.

If you are comparing us to a competitor's composite, the count that matters is 17, weighted, and named above. The 44 is the substrate.

Known limitations (honest caveats)

Quality screening is a filter, not an oracle. A few things it does not do well:

• It's lagging. Financials release quarterly and often on a 45-90 day delay. A business whose quality is deteriorating in real time will still score well on stale data for a quarter or two. Price momentum and regime detection partially compensate, but the lag is real.
• Fraud detection is weak. The manipulation and accruals gates would have flagged some accounting frauds and would have missed others. No factor model reads a 10-K for intent.
• Sector classification shapes everything. GICS has known quirks — conglomerates, recently reclassified names, platform businesses that don't fit a bucket cleanly. A misclassified stock gets compared to the wrong peers, and the nine sector-normalized factors become noise for that name.
• Revenue recognition is industry-specific. Aggressive bookings in subscription software, deferred revenue dynamics in ad-tech, percentage-of-completion accounting in long-cycle industrials — these can all look clean on the standard earnings-quality metrics while being economically aggressive. The model does not catch this.
• Small-cap coverage is thinner. Sector-relative comparisons need enough peers to be meaningful, and below roughly $1 billion in market cap data quality degrades and filings may be months stale. The auto-managed books screen at that floor for exactly this reason; the screener will show you smaller names, and those scores deserve more skepticism.
• Weights are calibrated on history. Factors go through periods of strong performance and periods of underperformance, and a weight fitted on the past has no mechanism to know which factors will keep working. We re-gate periodically; that is a mitigation, not a solution. More on this in Backtest Disclosure.

None of this makes the screen useless — it makes it a screen. The goal is to eliminate the bottom tail systematically. The top tail still requires everything else: valuation, regime, position sizing, and the discipline to exit when the thesis breaks.

Related reading

• Relative P/E Sector Analysis
• Backtesting Pitfalls
• Position Sizing and Stop Losses