Does a Dividend Safety Score Actually Predict Cuts? We Tested Ours
A dividend safety rating is a claim about the future. Almost nobody who publishes one goes back to check whether it came true.
We did. Our Dividend Health Score produced 32,167 point-in-time readings on 2,797 companies between 2021 and 2025. Companies rated very safe cut their dividend within the next eighteen months 1.2% of the time. Companies rated very risky did so 20.0% of the time.
What was actually measured
The hard part of any backtest is not the maths. It is making sure the score could not see the future.
Every reading here is reconstructed as of a knowledge date, meaning the date the market could first have known that quarter's results. Financial statements are restricted to those already filed. The dividend record is truncated to fiscal years that had already ended. Analyst estimates are withheld entirely, because we hold no historical record of them and inventing one would be the whole error.
Each reading is then paired with what happened next: did this company cut its dividend in the following eighteen months? Only companies actually paying at the time are counted, since a company paying nothing cannot cut.
That reconstruction is not free. A rebuilt reading and a live one are not quite the same measurement, because the live score sees analyst estimates and the rebuilt one does not. The effect is to make the test slightly harder than the live product, which is the direction an honest test should err in.
The result
| Rating at the time | Readings | Cut within 18 months |
|---|---|---|
| very risky | 100 | 20.0% |
| risky | 2,056 | 16.5% |
| borderline | 7,755 | 13.0% |
| safe | 16,366 | 7.4% |
| very safe | 5,890 | 1.2% |
The ordering is clean and the spread is wide. A company at the bottom of the scale was roughly seventeen times more likely to cut than one at the top.
One objection arrives immediately. A single company contributes many readings, so these rows are not independent observations. Keeping only the last reading per company per calendar year cuts the sample to 11,104 and barely moves the answer: 23.7% at very risky, 15.5% at risky, 12.7% at borderline, 7.6% at safe, and 1.3% at very safe.
Why lists of dividend cutters disagree
Any count of dividend cuts rests on a definition, and the obvious definition is wrong. It is worth knowing why, because it is the reason two sites can give you different answers about the same company, and the reason a screen you build yourself will return names that do not belong.
The obvious test compares a payment against the same payment a year earlier. Two entirely ordinary events break it, and both break it in the same direction: they flag large, stable companies that cut nothing.
The first is a change of payment schedule. ASML moved in 2023 from one interim and one large final instalment to three interims and a final. Each individual payment got smaller. The annual total rose. On a payment-against-payment test the company appears to have halved its dividend, and ASML has never cut it in its life. It has raised it for thirteen straight years.
ASML Holding N.V. (ASML.AS)Technology · Semiconductor Equipment & MaterialsData as of 08 Sept 2026high confidence
The second is a special dividend. Some companies pay an occasional extra many times the size of the ordinary quarterly one, and data feeds routinely label that extra as a regular payment. The year it does not repeat then reads as a collapse, when nothing has happened at all.
Both problems have the same shape, so both have the same fix. A change of schedule leaves the annual total intact, while a special inflates it. So the test here compares annual totals with specials stripped out, and decides what counts as a special using a threshold drawn from earlier years rather than from the year being judged. That last part matters more than it sounds: a threshold taken from the year under test deletes the larger pre-cut payment and hides the cut you are looking for.
One distinction survives all of it. A company that reduces its dividend and a company that stops paying altogether both appear as a fall in the annual total, and they are different events. A reduction is usually about affordability. A stop is more often a decision.
The objections, and whether they hold
Is it just the pandemic? The window opens in 2021, so the 2020 cut wave is already outside it. Removing 2020 and 2021 readings entirely leaves 20.6% at very risky against 1.1% at very safe. If anything the gap widens.
Is it just property companies? Removing every REIT and every business development company leaves 12.7% against 1.2%. The bottom band thins to 79 readings and gets noisier, but the ordering survives.
Is the score just yield wearing a different hat? This is the serious one, because high yields do predict cuts and a safety score that merely rediscovered that would be worth nothing. It does not. Inside the lowest yield band, under 2%, companies rated risky cut at 20.0% while those rated very safe cut at 1.0%. Inside the highest band, above 8%, the equivalent figures are 26.3% and 5.6%. The rating separates outcomes within a yield band, which is the only test that matters here.
Does it work everywhere? The ordering does. The levels do not. Companies rated safe cut at 2.8% in the United States and 10.8% outside it, and the same pattern holds at every rating. That gap is large enough to deserve its own article. Part of it is real: a European board typically resets the dividend each year against that year's earnings, while an American one defends a quarterly figure it has committed to. Part of it is measurement, because a variable annual dividend produces falls that a fixed quarterly one does not. Either way, a rating is better read as a ranking within a market than as an absolute probability of a cut.
What it got wrong
The interesting failures are not the low-rated companies that cut. They are the highly rated ones.
3M counts as a miss and a save at the same time, depending on which reading you start from. In February 2023 it was rated very safe, and it cut its dividend on 23 May 2024, inside the eighteen-month window. Counted from that reading, it is a miss. But the score did not sit still. By July 2023 it had dropped to safe, and by January 2024 to borderline, four months before the cut went ex. Every reading after the first one called it. Today, with the rebased dividend comfortably covered, it carries a score of 66.
3M Company (MMM.US)Industrials · ConglomeratesData as of 08 Sept 2026high confidence
The genuine failure looks different. Progress Software was rated safe or better throughout, was never downgraded below safe, and simply stopped paying a dividend after August 2024. No warning appeared in the rating at all. It now carries no score, for the reason set out in our piece on why some stocks have no dividend safety score: a company with no current dividend has nothing to rate.
That failure mode is worth understanding, because it is structural rather than bad luck. The score reads a company's ability to keep paying. It cannot read a decision to stop paying something the company could still afford. Affordability is in the accounts; a change of policy is not.
How to read a rating like this
The practical use of the table is not the individual percentages. It is the shape.
Moving from safe to borderline roughly doubles the historical cut rate. Moving from borderline to risky adds about a quarter again. Those are meaningful differences in odds across a portfolio of twenty holdings. They are much weaker guidance about any single company, because a 7.4% chance of something over eighteen months is very hard to feel.
The other use is as a filter on your attention. A portfolio of thirty companies cannot be researched evenly, and the ratings are lopsided in a helpful way. The bottom three ratings covered 31% of the readings and accounted for 52% of the cuts. The top rating covered 18% of the readings and 2.7% of the cuts. Reading the weakest third of a portfolio first therefore finds about half the trouble.
That is a smaller claim than prediction, and it is the one the numbers support. Once a holding is flagged, the five-step check is what turns a rating back into a reason.
What this does not prove
Three limits, and none of them is small.
The first is survivorship. Only companies still in our catalogue carry a score history, so a company that cut, collapsed and delisted is absent. That removes the worst outcomes, and it removes them disproportionately from the low-rated end.
The second is the period. Outcomes here run from 2022 to 2026, which contains no systemic dividend shock. The test shows the score separating companies in ordinary conditions. It says nothing about how it behaves in another 2020.
The third is residual measurement error. About one detected cut in fifteen could be dated only to a year rather than to a payment, because no single payment came in low enough to mark the moment. A handful of detections are artefacts we cannot remove automatically. Those errors inflate measured cut rates rather than deflating them, so the real separation is probably a little wider than the table shows. That is the direction we would rather be wrong in, but it is still being wrong.
A rating is a probability, not a promise. The most useful thing in the table is not the 1.2%. It is that the 1.2% is not zero.
Want to see where your own holdings sit? The Dividend Health Score is on every stock page, and the screener sorts by it.
Asking whether a metric has ever been tested is the last of five checks worth running on any dividend figure. The other four are in how to check a dividend figure before you rely on it.
Frequently asked questions
- Does a high safety score mean the dividend is guaranteed?
- No. Across 5,890 readings on companies rated very safe, 1.2% still cut within the following eighteen months. That is a low rate, not a zero rate, and the rating is a statement about odds rather than a promise.
- What counts as a dividend cut in this test?
- A fall of more than 10% in a company's annual dividend per share, after removing one-off special payments, confirmed against its fiscal-year record and dated to the ex-date of the first payment that came in lower. A change in payment schedule is not a cut, because it leaves the annual total intact.
- Is the score just measuring yield in disguise?
- No. Inside a single yield band the rating still separates outcomes by an order of magnitude. Among payers yielding under 2%, those rated risky cut at 20.0% and those rated very safe at 1.0%.
- Does the test work outside the United States?
- The ordering holds everywhere, but the levels differ. Companies rated safe cut at 2.8% in the United States and 10.8% elsewhere, so a rating should be read against local norms rather than as an absolute rate.
- What are the limits of this measurement?
- Three. Only companies still in our catalogue have a score history, so the worst outcomes are missing. The window opens after the 2020 cut wave, so it contains no systemic shock. And a residual number of detected cuts are measurement artefacts we cannot automate away.
DividendAtlas provides data and research for informational purposes only. Nothing here is investment advice or a recommendation to buy or sell any security. Always do your own research.