Three Roads to ROE: What the Source of Profitability Predicts

00Thesis

Return on equity is the headline summary of corporate profitability, but it is not one signal — it is three economically different engines bolted together: operating margin, asset turnover, and the equity multiplier. The 1920s DuPont identity separates them. This piece does two things v1 did not: it builds the decomposition on average balances and separates the buyback-inflated equity multiplier from genuine financial leverage, and it tests what the source of ROE predicts — persistence, forward returns, risk, and rate sensitivity — rather than asserting that one source is higher quality. The sample is 59 large-cap US non-financials, every figure from 10-K/XBRL filings, 2010–2025.

• 01Within this basket, the highest headline ROEs are largely a denominator effect. For 11 names buybacks have shrunk book equity so far that ROE is mechanically meaningless; for these the equity multiplier (assets/equity) bears almost no relation to debt — Home Depot's 22.5× multiplier is just 1.95× net-debt/EBITDA.
• 02The equity multiplier is not financial leverage. Across the universe it correlates only 0.24 with true net-debt/equity, and a high-minus-low multiplier portfolio has no significant rate beta (+1.55%/100bp, t = 1.5). The claim that buyback-built ROE is “rate-sensitive” does not survive the test.
• 03A reproducible log-decomposition classifier and a formal quality score replace hand-labels. NVDA tops the quality ranking (score 9.31) — 93.7% ROIC, net cash, a 1.45× multiplier — the genuine article, not a leverage artifact.
• 04The forward tests are honest and nuanced: turnover-sourced ROE has the most stable margins and the best subsequent returns; leverage-sourced ROE is the most persistent — a buyback policy is stickier than a margin competition erodes — but that persistence is mechanical, not quality.

01Why ROE alone fails

Two firms, the same return on equity, opposite businesses. Meta Platforms earned a 30% ROE last year on a 32.3% net margin turning assets 0.64× — a pure margin machine. Costco earned an almost identical 30% ROE on a 2.8% net margin — a tenth of Meta's — by spinning assets 3.68× a year, a pure turnover machine. Same headline number; one is pricing power, the other is operational velocity, and they have completely different durability, capital needs, and rate exposure. A screen that ranks on ROE alone treats them as identical. The decomposition is the whole point.

02The identity, measured properly

ROE factors, by construction, into three multiplicative terms. The denominators are average beginning-and-ending balances — not ending-only, which mechanically overstates ROE and the multiplier for firms repurchasing stock into year-end:

The third term is where v1 — and most DuPont write-ups — go wrong. The equity multiplier $\overline{\text{Assets}}/\overline{\text{Equity}}$ is not financial leverage. It rises for three quite different reasons: real debt, operating liabilities (payables, deferred revenue, lease liabilities), and — dominantly for mega-caps — buyback-shrunk book equity. Only the first is a debt-service exposure. So alongside the multiplier I measure genuine leverage directly — net debt / equity and net debt / EBITDA — and a robust, denominator-safe profitability measure, ROIC = NOPAT / invested capital, plus operating profitability on assets (the Fama-French RMW lens). When the equity base is distorted, these are what survive.

03The mega-cap decomposition

Figure 1 ranks the 48 firms whose ROE is meaningful, annotated with true leverage (net debt / equity). The annotation is the point: the highest ROEs are not the most levered. Apple's 171% sits on net-debt/equity of just 0.7× and NVIDIA's 102% on net cash. Table 1 gives the full decomposition with each firm's dominant road and contribution share.

Table 1. DuPont decomposition on trailing-3y average balances, top 22 of 48 meaningful-ROE firms by ROE. Road = largest log-deviation from the universe median; share = that term's % contribution to the firm's log-ROE deviation. Source: SEC EDGAR company-facts (10-K / XBRL), Yahoo Finance; author's calculations.

The map separates the roads cleanly. The margin road runs up the right edge — NVIDIA, Microsoft, Meta, the software and pharma franchises earning 25–53 cents per revenue dollar at modest turnover. The turnover road sits top-left: Costco and Walmart, ~3% margins spun 2.6–3.7× a year. Bubble size — the multiplier — is the leverage road, but as the next section shows, that size is mostly buybacks, not borrowing.

04When the denominator breaks: the multiplier is not leverage

For 11 firms the equity base has been repurchased down so far that ROE — and the multiplier — stop measuring anything. Table 2 lists them. The headline is the gap between the two leverage columns: Home Depot's equity multiplier reads 22.5×, but its actual balance-sheet leverage is 1.95× net-debt/EBITDA — a thoroughly ordinary credit profile. Colgate's “multiplier” is 39.5×; its net-debt/EBITDA is 1.54×. These are not levered firms; they are firms that returned so much capital that book equity is a rounding error. Calling that “leverage,” as v1 did, is the central conceptual error — and note that for the negative-equity names even ROIC distorts (Philip Morris prints 344% because invested capital is near zero too); the robust read for this group is asset-based operating profitability.

Table 2. The 11 denominator-distorted firms (excluded from all ROE-based statistics). The equity multiplier overstates leverage by 5–20× versus net-debt/EBITDA. Source: SEC EDGAR company-facts (10-K / XBRL); author's calculations.

05A formal ROE quality score

“High-quality ROE” needs a definition, not an adjective. I define it as ROE that is high, operating-driven, well-covered by ROIC, lightly levered, and stable — and score it cross-sectionally:

The ranking (Table 3) is intuitive in a way the raw ROE ranking is not. NVDA and AAPL top it — high ROIC, fat margins, net cash, low margin volatility — while the bottom is populated by the genuinely levered and the cyclically volatile (AT&T, Deere, the levered staples). This is the list a quality factor should hold; it is almost the inverse of a naïve ROE screen, which would put the buyback-distorted names on top.

Table 3. ROE quality score, cross-sectional z-scores over the 48 meaningful-ROE firms. Sector-neutral z-scores (margin/turnover/multiplier within GICS sector) are also computed in the data. Source: SEC EDGAR company-facts (10-K / XBRL), Yahoo Finance; author's calculations.

06Is quality priced?

Quality is only an edge net of price. Figure 3 plots the quality score against free-cash-flow yield. The fit slopes down (β = -0.14, t = -0.9): higher-quality names do trade richer — the market is not asleep. But the relationship is loose (R² = 0.02), and the residuals are the opportunity. The marquee compounders — NVIDIA, Apple, Costco — sit bottom-right (top quality, ~2% FCF yields: quality you pay full freight for). The interesting quadrant is top-right: Adobe (+4.54 quality, 9.2% FCF yield) and Accenture (9.4% yield) offer top-quartile quality at mid-teens P/Es. And the cheap, high-yield names on the left — AT&T, Verizon — are cheap precisely because the score flags their ROE as low-quality and levered. Cheapness without quality is a value trap with a DuPont signature.

07Does the source predict anything? The forward tests

This is the upgrade from description to research. I classify every firm-year 2010–2024 by road (log-decomposition vs that year's cross-section), then measure forward outcomes — the panel is 696 firm-years. The results (Table 4) are real but resist a tidy morality tale. Turnover-sourced ROE looks best on the metrics that matter: the highest median forward 12-month return (+16.4% vs +13.4% and +13.8%), the lowest forward volatility, and by far the most stable fundamentals — forward margin volatility of just 1.2 points, against 5.1 for the high-margin names, whose fat margins swing the most. Risk (forward drawdown) is similar across all three.

Table 4. Median forward outcomes by ROE-source road, panel of 696 firm-years, 2010–2024. Returns measured from ≈4 months after fiscal-end (filing). Source: SEC EDGAR company-facts (10-K / XBRL), Yahoo Finance; author's calculations.

Persistence is where the data overturns the intuition. Regressing next-year ROE on the road dummies, controlling for the current ROE level and sector (Table 5), ROE is highly persistent overall (autoregressive load 0.8, t = 30.5). But conditional on level, margin- and turnover-sourced ROE mean-revert significantly more than leverage-sourced ROE (−9pp, t = -4.9; and −6.09pp, t = -2.9). Read carefully: leverage-sourced ROE is the most durable — because a capital-structure choice is stickier than a fat margin, which competition erodes. But that durability is mechanical, not economic: it is the persistence of a buyback policy, and (§06, §08) it earns no valuation premium and no excess return. Durable is not the same as good.

Table 5. Pooled OLS, n = 676, R² = 0.65; sector fixed effects included. Leverage-driven is the omitted category. Source: SEC EDGAR company-facts (10-K / XBRL), Yahoo Finance; author's calculations.

08The equity multiplier is not rate risk

v1 claimed the buyback-built ROE of names like Apple is rate-sensitive — “the cost of the borrowed denominator rises with rates.” It tests false, for the simple reason that the denominator was never borrowed. Three pieces of evidence. First, the equity multiplier correlates only 0.24 with true net-debt/equity across the universe — it explains 6% of the variation in actual leverage. Second, 5 of the top 15 firms by ROE carry net cash; their multiplier reflects buybacks and operating liabilities, and rising rates help them (more interest income). Third, the direct test: a sector-neutral long-short portfolio sorted on the equity multiplier has a rate beta of +1.55%/100bp that is statistically zero (t = 1.5); the portfolio sorted on true net leverage is also insignificant (t = -0.1) at monthly frequency in these net-cash-rich mega-caps. The honest conclusion: the multiplier is a profitability-accounting artifact, not a financing exposure. Where these companies do carry rate sensitivity, it is in cash-flow duration, not the balance sheet — which is the subject of the rates piece.

09The secular-margin question, scoped honestly

v1 closed with a macro flourish — “a decade of equity returns underwritten by widening margins.” The decomposition does not support that as stated. Decomposing each firm's stock return since 2015 into $\Delta\log P=\Delta\log m+\Delta\log\text{rev}-\Delta\log\text{shares}+\Delta\log(\mathrm{P/E})$, the median name's price gain came more from revenue growth and multiple re-rating than from margins (Fig. 4): of the total, 40% revenue, 36% multiple expansion, only 17% margin expansion, and 7% buybacks. Margins mattered — but they were the third-largest driver, not the headline.

Blended net margin across the basket did rise — from 9.2% in 2015 to 15.4% in 2025 (Fig. 5) — but two caveats neuter the macro claim. Survivorship: this is a fixed basket of firms that remained large-cap, so margin expansion is partly selection, and partly weight drift toward the high-margin tech names that re-rated. Scope: it is a statement about 59 survivors, not “the index.” The defensible version is narrow — within this basket, blended margin widened — and even there, the return attribution says the re-rating did more work than the margins.

10Implications

• →Never screen on ROE alone, and never read the equity multiplier as leverage. Decompose, and check net-debt/EBITDA for the real balance-sheet risk; for the buyback-distorted cohort, switch the denominator to ROIC or operating-profit-on-assets.
• →Quality is a defined, scorable object — and partly priced. Hold the residual: top-quartile quality at non-trivial FCF yields (the Adobe/Accenture quadrant), not quality at any price.
• →Source predicts character, not a clean return premium here: turnover-sourced ROE is the most fundamentally stable, leverage-sourced the most mechanically persistent. Use the taxonomy to understand durability and risk, not as a standalone alpha signal in this survivor sample.

AData & method