Margin Inflection Backtest: the Signal Doesn't Hold — 2.52 vs. 15.02 Percent a Year
A company whose operating margin, after four quiet quarters, suddenly jumps by several percentage points: is that a better long-term buy than a company whose margin was already moving? We tested the question over 163 months (January 2013 to July 2026) across every US stock with an unbroken quarterly record — with delisted stocks kept in the portfolio, the filing date rather than the quarter-end as the trigger date, and a separate class for the jump from a loss into profit. The answer is uncomfortable: the main signal does not hold. Over 13.5 years it produces only 27 positions, compounds at 2.52 percent a year, and trails far behind the S&P 500 with dividends (15.02 percent) — and even the most favorable variant of the rule fails to beat the index. Against related signals, the margin inflection is the weakest member of the family, and it does not even lead the revenue inflection: among companies with both signals, the median time lead is zero months.
Thomas Mücke
Founder & Publisher
The question: is the margin inflection a buy signal?
A revenue jump is easy to spot — it sits on the top line of every income statement. A margin improvement is less obvious, but for many investors the more interesting news: it says a business is not just growing, it is getting more profitable while doing so. This backtest asks: is the moment a previously quiet operating margin jumps by several percentage points a better entry point than a company whose margin has already been in motion for a while?
We measure the TTM operating margin: the sum of operating income over the last four quarters divided by the sum of revenue over the same four quarters, calculated only when all four quarters are present without gaps and total revenue is positive. The change is expressed in percentage points — a margin that climbs from 1 to 2 percent has improved by 1 percentage point, not 100 percent. Two conditions must both hold:
- Acceleration. Each of the most recent at least two consecutive quarters improves the TTM margin by at least 3 percentage points year over year, and the margin in the signal quarter is greater than zero. This positivity condition separates the start of a profitable business model from a mere narrowing of losses: a company moving from −40 to −5 percent has improved by 35 percentage points and still is not making money.
- A quiet period before it. Each of the four quarters immediately before the first acceleration quarter moves the margin by less than 1 percentage point, measured in either direction. Without this condition, any company with an ongoing margin expansion that happened to have two strong quarters in a row would be counted too.
Companies meeting both conditions form class K. Companies meeting the acceleration condition but unable to document the quiet period because their history is too short form K-young; companies meeting acceleration but already in motion beforehand form L-old. A separate signal type sits alongside these: the margin turnaround, the jump from at least two quarters of non-positive TTM margin into at least two quarters of positive margin — without a quiet-period or threshold test, because here the sign change itself is the signal.
The result: the main signal does not hold
The answer is unambiguous. At twelve months' holding period — the main mode of this study, because among the fixed holding periods that is where the arm is most densely populated — class K produces exactly 27 positions over 163 months and a return of 2.52 percent a year. The conservative and full calculations are identical here, because none of these 27 positions carries the broken-price-series pattern described in the data-error section below.
| Arm | Positions | Return p.a., conservative | Return p.a., full | Largest drawdown | Hit rate |
|---|---|---|---|---|---|
| K (main signal, quiet period documented) | 27 | 2.52% | 2.52% | −62.79% | 66.67% |
| K-young (history too short) | 605 | 5.40% | 7.20% | −45.41% | 53.06% |
| L-old (already in motion beforehand) | 3,621 | 11.62% | 41.82% | −31.75% | 53.69% |
| all purchases in the backtest | 4,253 | 11.13% | 39.36% | −32.01% | 53.68% |
| S&P 500 Total Return | — | 15.02% (identical in both calculations) | −23.87% | — | |
| Universe, equal-weighted (trimmed mean) | — | 4.77% (identical in both calculations) | −53.23% | — | |
| Universe, median stock | — | −5.45% (identical in both calculations) | −60.68% | — | |
Twelve-month holding period, January 2013 to July 2026, equally weighted, 0.1 percent cost per side. Class K stays below the S&P 500 at every holding period:
| Holding period | Positions (K) | Stocks/month (median) | Months holding 1–2 stocks | Return p.a. |
|---|---|---|---|---|
| 3 months | 27 | 1 | 56 | −0.98% |
| 6 months | 27 | 1 | 85 | −2.83% |
| 12 months | 27 | 2 | 94 | 2.52% |
| as long as the margin does not fall twice in a row | 25 | 3 | 72 | 6.83% |
No holding period brings the main signal above the index — at three and six months the arm even loses money. And the arm is thinly populated: across 163 grid months, class K holds a median of 2 stocks, a maximum of 5, and holds at most 2 stocks in 108 months in total — 14 of those months being entirely empty. The 2.52 percent is not an equal-weighted portfolio; it is the compounded return of a handful of individual cases.
Sensitivity: even the best cell falls short of the index
Does the signal hold under a different setting of the rule? We turned three dials — threshold, quiet-period limit, number of acceleration quarters — one at a time, holding everything else constant; that gives six variants. All figures are in the full calculation — for class K the conservative and the full calculation coincide in every variant anyway:
| Variant | What was changed | Positions (K) | Return p.a. | Largest drawdown |
|---|---|---|---|---|
m3_r1_f2 (default) | threshold 3 pp, quiet period 1 pp, 2 acceleration quarters | 27 | 2.52% | −62.79% |
m2_r1_f2 | threshold lowered to 2 pp | 80 | 10.40% | −39.35% |
m5_r1_f2 | threshold raised to 5 pp | 8 | 2.49% | −33.25% |
m3_r05_f2 | quiet-period limit tightened to 0.5 pp | 2 | 1.45% | −12.86% |
m3_r2_f2 (best cell) | quiet-period limit loosened to 2 pp | 209 | 12.93% | −29.97% |
m3_r1_f3 | three acceleration quarters required instead of two | 23 | 10.00% | −57.28% |
The best cell — a quiet-period limit loosened to 2 percentage points — raises the case count nearly eightfold, from 27 to 209 positions, and lifts the return to 12.93 percent a year. That is more than any other variant of the main signal, but still 2.09 percentage points below the S&P 500 Total Return (15.02 percent). The quiet-period limit does not change how many companies trigger a signal overall — it only reshuffles positions between classes K and L-old. Two of the variants leave too few cases to prove anything: 8 and 2 positions are individual companies, not a strategy.
The margin turnaround: the jump out of the loss zone
A separate signal type asks whether the switch from a non-positive to a positive TTM margin — at least two quarters on each side — is a buy signal. Here too the index remains unbeaten:
| Holding period | Positions (K) | Return p.a., conservative | Return p.a., full | Largest drawdown | Hit rate |
|---|---|---|---|---|---|
| 3 months | 1,859 | 11.24% | 19.97% | −35.94% | 50.30% |
| 6 months | 1,859 | 13.05% | 24.29% | −34.78% | 50.89% |
| 12 months | 1,850 | 10.64% | 53.91% | −40.91% | 49.24% |
| as long as the margin does not fall twice in a row | 1,851 | 12.13% | 57.05% | −47.87% | 48.57% |
At 1,850 positions (twelve months), the margin turnaround is far more densely populated than the main signal — no surprise, since it only requires the sign change, not four documented quiet quarters. Even so, conservatively calculated, it returns 10.64 percent a year, below the S&P 500. The gap between the conservative and full calculation (53.91 percent) is notable — part of that gap sits in the same broken price series described in the data-error section.
Family comparison: only the revenue inflection holds up
TickerGuard has already tested the same question — "is the fresh inflection a better buy than the established move?" — for revenue and for earnings. All three backtests use the same convention: twelve-month holding period, 2013–2026, the conservative figure excluding positions under jump suspicion. One difference remains: the margin inflection has no organic-growth segment — a margin cannot be bought, and an acquisition tends to dilute it rather than lift it. Its main arm is simply called K, while the revenue-inflection arm is reported as K|org (organically grown companies only).
| Signal | Variant | Arm | Positions | Return p.a., conservative | Return p.a., full |
|---|---|---|---|---|---|
| Revenue inflection | k2_30_70_r15 | K|org | 277 | 16.56% | 26.28% |
| S&P 500 Total Return | — | — | — | 15.02% | 15.02% |
| Margin turnaround | wende | K | 1,850 | 10.64% | 53.91% |
| Earnings inflection | g2_30_r15_m2 | K|org | 258 | 6.76% | 6.76% |
| Margin inflection | m3_r1_f2 | K | 27 | 2.52% | 2.52% |
The revenue inflection remains the only family member that holds up — it is the only one of the four that beats the index on a conservative basis. The margin inflection trails at 2.52 percent, well behind the margin turnaround and the earnings inflection.
Does the margin inflection come earlier? No.
An obvious defense of the margin inflection: even if it does not hold as a standalone buy signal, it could still be an early hint of an upcoming revenue inflection — margin picking up before revenue visibly does. We measured this directly: for each company, the gap in months between the first signal months on each side.
| Selection | Companies with margin inflection | Companies with revenue inflection | Both | Share both | Median lead | Revenue first | Margin first | Same month |
|---|---|---|---|---|---|---|---|---|
| all freshness classes | 2,555 | 2,157 | 1,269 | 49.67% | 0 months | 624 | 575 | 70 |
| documented inflections only (K) | 27 | 513 | 6 | 22.22% | −47 months | 1 | 5 | 0 |
In the broad selection (all freshness classes), nearly half of all companies with a margin signal (1,269 of 2,555) also show a revenue signal at some point — the median gap between the two first hits is exactly zero months. Revenue came first in 624 cases, margin first in 575, and both landed in the same month in 70. This broad, meaningful sample does not support the claim that "margin is the earlier signal."
Restricted to the documented margin inflections (class K), only 6 companies with both signals remain: here margin led revenue in 5 cases, revenue led in only 1, with a median lead of 47 months. The picture appears to flip in margin's favor — but 6 companies are too thin a base to draw anything from, and they do not contradict the finding from the much larger sample above (1,269 companies).
Revenue base: the pattern repeats
As with the related revenue-inflection backtest, the size of the business matters here too. Split by the sum of the last four reported quarterly revenues (full calculation, twelve months, all purchases regardless of freshness class):
| Revenue base | Positions | Return p.a. | Largest drawdown | Hit rate |
|---|---|---|---|---|
| under $10 million | 49 | −8.84% | −87.60% | 36.73% |
| $10 to $100 million | 609 | 10.64% | −36.41% | 47.29% |
| over $100 million | 3,595 | 41.99% | −34.40% | 54.99% |
Small revenue bases are no more workable here than for the revenue inflection: below $10 million in revenue, the backtest loses money on average. Unlike the revenue inflection, however, this segment fully flows into the headline numbers above — there is no live scanner for the margin inflection that could enforce a revenue floor.
The data error and why the conservative figure is the headline number
While hand-checking individual positions, one price series stood out. It runs under the ticker SRX and sits at $4.75 in June 2015, jumps to $0.057 in July, to $165.74 in August, back to $4.22 in September, and back up to $154.67 in October — with an adjustment factor that never changes across any of these months. This is not a price move; it is a series missing a reverse-split adjustment. That the prices and the reported figures of this position do not belong together is clear from where each one ends: the quarterly figures come from SRA International, which filed its last report in November 2015 and deregistered with the securities regulator afterwards — the price series under the same ticker runs on until April 2020. The affected position books a monthly return of +290,672 percent in August 2015 as a result. That single position lifts the "all purchases" arm's monthly return in August 2015 to +1,157 percent (with 250 stocks held) and, with it, the main variant's twelve-month annual return from 11.13 to 39.36 percent. Seven of 4,253 positions in the main variant exceed the applied threshold (monthly return above 200 percent) and are removed in the conservative series — which is why that series is the headline figure throughout this study.
SRX is not the only affected series: CMRF ($5.76 to $0.0014 and back to $5.00 after eight months of placeholder pricing) and other positions carry the same pattern, sometimes broken downward instead of upward. The applied filter only checks for jumps above 200 percent. Recomputed with a two-sided filter, the reported returns actually rise: all purchases from 11.13 to 12.01 percent, K-young from 5.40 to 8.41, L-old from 11.62 to 12.12 — class K stays unchanged at 2.52 percent, because no such series sits in it. The known error therefore pushes the figures down rather than up: the direction is conservative, not overstated.
A second, smaller caveat concerns the total-loss calculation: of the 27 positions in class K (twelve months), exactly one carries a forced sale because its price series ends mid-study (position "FI," entered April 2025, series ends in November 2025). In the full calculation that exit is booked at the last available price; only in the total-loss calculation running alongside does it cost −100 percent. There, that single position pulls the arm's return from +2.52 to −2.35 percent a year — an individual case, not a robust metric for the whole signal.
Litmus test: known companies behave as expected
Three positions were fully hand-verified, matching the report to the decimal: UFPT (UFP Technologies) triggers a clean main signal in March 2023 — four quiet prior quarters moving by at most 0.98 percentage points, then two acceleration quarters of +4.39 and +5.37 percentage points. ANET (Arista Networks) meets the acceleration test in May 2023 but not the quiet-period test — its margin was already improving by up to 2.89 percentage points a quarter for four quarters beforehand — and correctly lands in class L-old. ABNB (Airbnb) triggers a margin turnaround in May 2022: two quarters of negative TTM margin, then two of positive, entering on the filing date of the second profitable quarter.
Well-known efficiency stories confirm the rule rather than break it: META's "year of efficiency" in 2023 only triggers a signal in retrospect as of April 2024 (class L-old, +13.82 percentage points) — because the TTM margin measures 2023 against a still-high 2022 base all year and only reaches two consecutive acceleration quarters in the first quarter of 2024. AMZN behaves the same way: signal in May 2024, also L-old. The negative controls KO (Coca-Cola) and PG (Procter & Gamble) never trigger a single K signal across the entire period in any of the six tested variants; their few L-old hits are explained by base effects from one-off charges (KO October 2025, PG e.g. October 2020), not genuine margin inflections.
How we calculated this
Calculations use only quarters that had been filed with the securities regulator by the relevant month-end — a TTM margin only counts as known from the latest filing date among all eight figures involved (four operating-income figures, four revenue figures, each with its own filing date). In the data, these two dates diverge for 7,408 of 230,528 quarterly rows, in extreme cases by more than four years; a backtest using the quarter-end instead of the filing date buys systematically too early.
- Universe and period. US stocks, January 2013 to July 2026, 163 months. Delisted stocks remain in the portfolio; if a price series ends mid-period, the position is force-sold at the last available price — a total-loss calculation with −100 percent for this case runs alongside throughout.
- Portfolio. Equally weighted, monthly rebalancing, 0.1 percent cost per side, minimum price $1 on the entry date (checked on the raw price, calculated on the adjusted price). Purchases happen on an episode's first signal; a running position is not added to. The "as long as" holding period ends once two consecutive quarters show a year-over-year falling margin, and at 120 months at the latest.
- Metric. Only operating income (us-gaap:OperatingIncomeLoss) is read — never net income, which carries interest, taxes, and disposal gains and therefore measures something other than the ongoing business.
- Continuity. Calculations run on the unbroken trailing segment of the quarterly sequence, measured by label rather than calendar date.
- Threshold comparisons run with a tolerance (1e-9) so floating-point representation does not randomly push a value above or below a limit.
- Cross-check. The monthly series re-aggregated in the report deviate from those stored in the portfolio run by 0 percentage points.
Sources. Operating income, revenue, and filing dates come from the US securities regulator's public mandatory filings (annual reports on Form 10-K and quarterly reports on Form 10-Q, machine-readable via EDGAR). Monthly prices and the S&P 500 Total Return benchmark come from our own database. No network calls during the run.
What this study does not say
- The main signal is statistically thin. 27 positions over 163 months, a median of 2 stocks held at a time, empty in 14 months. That is a handful of individual cases, not a portfolio.
- 15.14 percent of quarters in the dataset have no reported revenue and drop out of the measurement entirely — one reason this backtest sees fewer companies than the revenue inflection.
- 15.26 percent of quarters are derived by subtraction (fourth quarter = annual figure minus the first three quarters), because the regulator does not require it to be filed separately. These quarters are calculated, not reported.
- Banks and insurers are excluded entirely, because they do not report an operating-income line in this sense.
- Taxes and trading spreads are not modeled. Only 0.1 percent cost per side is included.
- Only monthly jumps above 200 percent were checked. Smaller, similarly uncorrected reverse splits may still sit in the conservative figure.
- A backtest is not a forecast. This study is market research, not investment advice and not a buy recommendation.
Consequence: no scanner
Neither the main signal (2.52 percent a year), nor its most favorable sensitivity variant (12.93 percent), nor the margin turnaround (10.64 percent), nor established margin expanders (11.62 percent) beat the S&P 500 Total Return of 15.02 percent a year. This backtest therefore does not become a live scanner. Unlike the revenue-inflection backtest, where the rule beat the index on a conservative basis and became a scanner with a revenue and timing filter, there is no viable basis here for an automated buy signal. We publish this study anyway, because an honest backtest also has to show which plausible ideas did not work out in hindsight.
Frequently Asked Questions
A two-part buy rule: acceleration (in each of the most recent at least two quarters, the TTM operating margin improves by at least 3 percentage points year over year, with a positive margin in the signal quarter) and a quiet period before it (each of the four prior quarters moves the margin by less than 1 percentage point). Purchases happen on the filing date of the second acceleration quarter, equally weighted, with 0.1 percent cost per side, held for 3, 6, 12 months, or for as long as the margin does not fall year over year in two consecutive quarters (capped at 120 months).
Because over 13.5 years it produces only 27 positions and, at 2.52 percent a year, trails far behind the S&P 500 (15.02 percent). Even the most favorable of six tested rule variants — a loosened quiet-period threshold with 209 positions and 12.93 percent a year — falls short of the index. The neighboring classes, established margin expanders (11.62 percent) and margin turnarounds (10.64 percent), also stay below it.
No, it is merely less bad. Companies whose margin was already moving before the acceleration quarter return a conservative 11.62 percent a year on 3,621 positions — more substance than the main signal, but still below the S&P 500. None of the three freshness classes in this backtest beats the index.
No. Among the 1,269 companies showing both signals, the median time gap between the first hits is zero months; revenue came first in 624 cases, margin first in 575, both in the same month in 70. The common assumption that a margin improvement precedes a revenue acceleration is not supported by this data.
Because the price history contains individual series in which a reverse stock split was never carried into the adjusted price. One documented position books a +290,672 percent monthly return in August 2015 alone, lifting the whole arm's annual return from 11.13 to 39.36 percent. It is not the only affected series, and the applied filter — excluding any position with a monthly jump above 200 percent — is conservative in direction rather than overstating results.
No. Neither the main signal, nor its most favorable variant, nor the neighboring classes beat the index — a scanner would automate a rule that did not hold up historically. We only publish scanners for signals that pass this test, and the margin inflection is not one of them.
