Do Share Buybacks Beat the Market? What 227,219 Holding Periods Show
The Finding
Do share buybacks beat the market? Between 1998 and 2025, U.S. companies that reduced their share count by 5% or more over the trailing year beat their sector peers by 2.78 percentage points over the following twelve months, winning 70.3% of the 111 quarters measured. This was the pre-registered test, and it cleared.
The study covers 227,219 twelve-month holding periods across 17,031 tickers, 72.5% of which are delisted today. Delisted companies were kept in, and their returns run to the last available price. Returns are total return, equal-weighted, measured against the mean of all holdings in the same formation quarter and sector, so no sector tilt survives by construction. The test statistic discounts for overlapping windows, dividing the effective sample by four.
An exploratory cut, not pre-declared, found the strongest cell in the study was companies buying back 5–10% of their shares: +2.60 points versus flat-count peers, 68.5% of quarters, the only band clearing a discounted t of 2.0. A forward test on this band was pre-registered in October 2026.
What did not hold: the penalty for heavy share issuers nearly vanished once corporate-action artifacts were screened out. Worst case: every band lost roughly a third to a half on its worst formation quarter, all in the 2007–08 crisis; the 5–10% band's −35.9% was the shallowest drawdown.
Source: Sharadar; calculations by Brett Owens.
Disclosure
Pre-commitment. The primary test (share-count reduction of 5% or more versus all other holdings) was pre-registered in 00_PREREGISTRATION.md, committed 2026-09-09 at 14:19 (commit 7cb38e9). The pre-registration, pipeline scripts, and results file landed in a single commit, so the commit history cannot independently witness that the specification preceded the results; it is stated here that it did, and the reader may weigh that accordingly.
What was not pre-committed. The eight share-count bands (issued >5%, 2–5%, 0–2%; flat; bought back 2–5%, 5–10%, 10–20%, >20%) were descriptive bins drawn at report time, not before. The 5–10% "sweet spot" is therefore an exploratory finding. It is the strongest cell in the study, and it is exactly the shape a post-hoc bin would take. The forward test below exists to turn it from an observation into a test.
A discrepancy of record. The study's original results file stated that share-count moves beyond ±25% were excluded as corporate actions. No such screen existed in any script. The frozen event set carries reduction values from −67,015 (likely a reporting discontinuity) to +0.9997. This was found on 2026-09-14 during replication, annotated in place, and the outer bands were recomputed both ways. The recomputation counts as one additional look at the data. A replicator must not apply the 25% screen to reproduce the as-run numbers; applying one is a different, and arguably better, study.
Prior publication. A version of this finding appeared in a newsletter column on 2026-09-16, before this page existed, featuring one company as an illustration. That column is disclosed as prior work by the same author, not as independent evidence. This page carries the full specification the column did not, and reports the negative sub-finding the column did not have.
Data. Sharadar SF1 Core US Fundamentals, pulled 2026-09-09 under the author's own subscription on the author's own hardware. Raw vendor tables are not redistributed; a replicator pulls them under their own subscription and should expect small drift where the vendor has revised history since.
Companion study. A second pre-registered study (2026-09-09, commit 011a857) on the same event set tested whether dividend growth interacts with the buyback bands. Its results are reported in full below, including the cells that did not clear.
Why I Ran This Study
Most investors agree that buybacks are generally good for shareholders. They take shares off the market, and fewer shares make everything look better on a per-share basis: more earnings per share, more dividends per share, and a nice virtuous cycle going forward.
I was interested in two specific things I hadn’t seen checked:
Is the effect real once you keep the companies that died? Many buyback studies drop delisted names, so companies that went out of business and stopped trading are conveniently removed. Getting rid of them can skew the results.
Are more buybacks even better? A company buying back 20% of its shares is operating differently from one buying back, say, 5%. I suspected the difference mattered. The question was how.
The answer to the first question is yes. The effect holds with the dead companies kept in.
The answer to the second is that there’s a sweet spot for buybacks. Past it, the evidence gets very thin very quickly.
The Method
Universe. Sharadar tickers.category beginning "Domestic Common Stock"; exchanges NYSE, NASDAQ, NYSEMKT; ADRs and funds excluded. 17,031 tickers, 72.5% delisted as of the pull. A survivorship tripwire halts the run if fewer than 1,000 delisted names are present.
Formation. Each quarterly as-reported filing (SF1, dimension = 'ARQ') is a formation event. Share count is sharesbas, basic shares outstanding as reported, not diluted weighted average, which is null for some issuers, and not the cash-flow field. The year-ago pair is the ARQ row closest to 365 days prior, within ±45 days; no pair, no event. Formation date is the filing date; entry is the next trading day.
Measure. reduction = (shares_prior − shares_now) / shares_prior. Positive means the company shrank its count.
Filters at formation. Price ≥ $5 as traded; market capitalization ≥ $500 million.
Return. Forward 12-month total return on split- and dividend-adjusted close, exit at the last trading day within 365 calendar days of entry. Delisted holdings run to their last available price with no reinvestment and no substitution; 8,608 holdings (3.79%) end early, and 39.2% of all holdings sit on tickers delisted today.
Window. Formations 1998Q1 through 2025Q3, 111 quarterly cohorts. Last formation 2025-09-09, ensuring 365 days of forward prices. 227,219 completed holdings, of which 227,153 carry a sector tag.
Sector neutralization (primary). Each holding's return is demeaned on its formation-quarter × sector cell (11 Sharadar sectors). All spreads are computed on these residuals, so no sector tilt survives by construction. Reported alongside: cohort × sector cell matching, cohort × sector × size-quintile double matching, and a pre-registered size-decile match.
Weighting. Equal-weight within group within cohort throughout.
Inference. The unit is the formation quarter. A quarter is "won" when the cohort-level difference is strictly positive; cohorts enter with ≥5 treated holdings, and all 111 qualify. Two t-statistics are computed: naive and overlap-discounted, which divides the effective sample by four because adjacent quarterly cohorts share roughly nine of twelve forward months. The pre-registration fixed the discounted statistic as the honest one. No p-values are quoted.
The Pre-registered Test: Shrinkers ≥5% vs Everyone Else
Method | Quarters | Mean spread | Quarters won | t naive | t overlap-discounted |
Raw, no neutralization | 111 | +3.23 pp | 63.1% | +3.62 | +1.81 |
Size-decile matched | 111 | +3.02 pp | 63.1% | +3.32 | +1.66 |
Sector-neutral, demeaned (primary) | 111 | +2.78 pp | 70.3% | +4.57 | +2.28 |
Sector-matched cells | 111 | +3.03 pp | 62.2% | +4.12 | +2.06 |
Sector × size double-matched | 111 | +2.21 pp | 60.4% | +3.41 | +1.70 |
Excluding early-terminated | 111 | +3.28 pp | 64.9% | +3.58 | +1.79 |
1998–2000 removed | 99 | +3.03 pp | 64.6% | +3.85 | +1.93 |
2010 onward only | 63 | +3.16 pp | 68.3% | +2.91 | +1.46 |
The effect is positive under every method and in 10 of 11 sectors (Consumer Defensive the exception at −0.72 pp). The overlap-discounted statistic ranges from 1.46 to 2.28 depending on method: it clears the conventional 2.0 threshold on the primary sector-neutral cut and the sector-matched cut, and sits below it on the strictest double match and the post-2010 subsample. The result sits on the boundary across defensible methods. That is the honest reading, and it is stated here rather than by selecting the most favorable row.
The Exploratory Cut: Is There a Sweet Spot?
Bands on the trailing-year share-count change. Not pre-declared.
Band | Holdings | Raw 12-mo return | Sector-neutral vs flat | Quarters won | t overlap-discounted |
Issued >5% | 42,480 | +8.43% | −1.66 pp | 39.6% | −0.66 |
Issued 2–5% | 25,251 | +13.47% | +1.66 pp | 50.5% | +1.12 |
Issued 0–2% | 81,474 | +12.13% | +0.29 pp | 50.5% | +0.29 |
Flat (0–2% buyback) | 34,588 | +12.93% | — | — | — |
Bought back 2–5% | 25,098 | +13.01% | +0.62 pp | 61.3% | +1.12 |
Bought back 5–10% | 13,405 | +14.03% | +2.60 pp | 68.5% | +2.52 |
Bought back 10–20% | 4,202 | +11.51% | +2.07 pp | 55.0% | +1.66 |
Bought back >20% | 655 | +15.87% | +5.26 pp | 53.2% | +0.59 |
The 5–10% band is the only cell in the study that clears a discounted t of 2.0. The dose curve from 2–5% to 5–10% to 10–20% runs +0.62 → +2.60 → +2.07: coherent, but not "more is better." The >20% band's +5.26 is carried by 655 holdings, as few as one to a dozen in many quarters, and its t of 0.59 says as much.
Reading this honestly: a bin drawn at report time that turns out to be the best bin is what a data-mined result looks like. It may also be true. The forward test below is the only thing that distinguishes the two.
We go looking for the findings other people assume. Get the next one straight to your inbox.
What Did Not Hold: The Heavy-issuer Penalty
The original results emphasized the penalty on companies issuing shares, the "printers." Two things happened to that finding on replication.
First, sector composition. Raw, heavy issuers trailed flat-count peers by 4.50 points. Sector-neutral, the gap shrinks to 1.66 points. Two-thirds of the raw penalty was that issuers cluster in sectors that underperformed.
Second, the corporate-action screen. When the ±25% screen the original text described is actually applied, 12,392 of the 42,480 heavy-issuer events are removed, a mix of genuine mega-dilutions and likely reporting discontinuities or corporate actions. What remains:
Band | Run | Holdings | Sector-neutral vs flat | Quarters won | t overlap-discounted |
Issued >5% | as-run | 42,480 | −1.66 pp | 39.6% | −0.66 |
Issued >5% | screened | 30,088 | −0.51 pp | 44.1% | −0.09 |
Bought back >20% | as-run | 655 | +5.26 pp | 53.2% | +0.59 |
Bought back >20% | screened | 325 | −0.25 pp | 47.0% | −0.29 |
Bought back 5–10% | either | 13,405 | +2.60 / +2.58 pp | 68.5% | +2.52 |
The heavy-issuer penalty, once screened, is indistinguishable from zero. The >20% shrinker premium halves its sample and turns slightly negative. The 5–10% band does not move.
This is reported as a negative finding, not buried. A claim that "companies printing shares lag their peers three quarters in five" appeared in the prior publication; on properly screened data, that claim does not survive at any conventional threshold.
Worst Cases by Band
Band | Worst formation quarter | Worst four-quarter run | Best formation quarter |
Issued >5% | 2008Q1: −47.9% | 2008Q3: −37.9% | 2009Q1: +78.5% |
Issued 2–5% | 2008Q1: −38.8% | 2008Q3: −30.9% | 1999Q1: +97.0% |
Issued 0–2% | 2008Q1: −42.1% | 2008Q3: −31.5% | 2020Q2: +68.1% |
Flat | 2008Q1: −39.5% | 2008Q3: −29.4% | 2020Q2: +68.8% |
Bought back 2–5% | 2008Q1: −36.7% | 2008Q3: −28.3% | 2020Q2: +79.0% |
Bought back 5–10% | 2008Q1: −35.9% | 2008Q3: −28.6% | 2020Q2: +98.7% |
Bought back 10–20% | 2008Q1: −40.2% | 2008Q2: −28.8% | 2020Q2: +93.4% |
Bought back >20% | 2007Q4: −36.3% | 2008Q3: −31.9% | 2020Q3: +421% (a handful) |
Every band's worst formation quarter falls in the 2007–08 crisis, and every band lost roughly a third to a half on it. The 5–10% band's −35.9% is the shallowest loss; the heavy-issuer band's −47.9% is the deepest. A 5–10% buyback did not protect anyone from 2008. It lost less.
Companion: Does Dividend Growth Change the Picture?
A second pre-registered study on the same event set asked whether companies that grow their dividend and buy back 5–10% do better than either alone. Dividend growth is a proxy (the data does not distinguish regular from special dividends), defined as the median per-payment amount in the trailing 12 months exceeding the prior 12 months' median by 0.5%, with at least two payments in each window. 113,669 holdings classified; 113,484 unclassified and held out.
Cell | Holdings | Sector-neutral mean | Quarters won | t overlap |
Grower + bought back 5–10% | 5,513 | +2.23 pp | 63% | +1.28 |
Bought back 5–10%, non-grower | 1,915 | +0.71 pp | 54% | +0.24 |
Grower alone | 69,972 | −0.24 pp | 47% | −0.27 |
Neither | 36,269 | −0.65 pp | 38% | −1.28 |
Heavy issuer, non-grower | 6,110 | −2.53 pp | 25.5% | −2.92 |
What this licenses: the combination ranks first but does not clear significance. Dividend growth alone is flat against sector peers on nearly 70,000 holdings, which is worth sitting with. The one cell that does clear is heavy issuers that also don't grow their dividend, at −2.53 points and a quarter of quarters won. On the secondary definition of "grower," that issuer split does not reproduce.
What I Take From This
Three takeaways:
The pre-registered result held. Companies that shrink their share count beat their sector peers, and they do it even when we keep the companies that went bust. That is the key finding.
The sweet spot is real in the data, but it’s a little sketchy to make a prediction from it. The best band was 5% to 10% a year. Unfortunately, I drew that band after looking at the data, which compromises it. That’s a case where I could be fitting the data conveniently, so we need to test it going forward. The forward test is the arbiter. It starts with this quarter’s filings as a pilot, and the formal count begins in 2027.
I got something wrong: the issuer penalty. In a September 16 column, I said companies printing shares lag their peers. Once we neutralize for sector and screen the data properly, that penalty is statistically indistinguishable from zero. I wanted to make that clarification here.
Forward Test: Pre-registered October 2026
What we're testing. Companies that shrank their share count by 5–10% over the prior year beat "flat" companies (0–2% shrink) in the following 12 months, measured against sector peers. Historically, from 1998 to 2025, that gap averaged +2.6 percentage points and was positive in 68.5% of quarters. The forward test asks whether it holds on companies we had not yet seen when we published.
What the test can and can't tell us. This is a forward-monitoring record, not a promise of a decisive answer. Quarterly cohorts overlap: neighboring quarters share about 9 of their 12 months. So 20 cohorts carry roughly the information of 5 independent ones. Even if the historical effect is real and unchanged, there is only about a 1-in-5 chance the first formal checkpoint will confirm it statistically. A result that is "not statistically confirmed" is not evidence that the effect is gone.
Null hypothesis (H0). For cohorts formed after this test was locked, the average sector-adjusted 12-month total-return gap between the 5–10% band and the flat band is zero or negative.
Method. This is the same calculation as the historical study, with one stated difference (the sector tag, below).
Data: Sharadar as-reported quarterly fundamentals. Shares are basic shares outstanding, as reported.
Formation and entry: formation is on the filing date; entry is the next trading day.
Universe: US common stocks on NYSE, NASDAQ and NYSE American, priced at least $5 (as traded) with at least $500M market cap at formation. The universe includes companies that later delist.
Bands: change in share count over the trailing four quarters. The 5–10% band is a reduction above 5% up to 10%. The flat band is a reduction above 0% up to 2%.
Return: 12-month total return (dividends and splits included), equal weight. Each holding's return is measured against the average of all holdings in the same formation quarter and sector.
Delistings: a holding that delists keeps its return through its last available price. There is no reinvestment and no substitution. A holding with no price after entry is excluded and counted.
Sector tag: the vendor's tag, recorded at formation. The historical study used each company's tag as of 2026, so this forward test is an operational analogue of that study rather than a literal copy.
Frozen inputs: each cohort's inputs are frozen and fingerprinted when it forms. This covers the as-reported fundamentals, the formation price and market-cap screen, and the sector tag. Later vendor revisions cannot change a recorded cohort, and any change in vendor version or method is logged.
Which companies count.
Lock date: October 14, 2026, or the actual publication date if later. Only filings dated on or after it count.
Pilot: filings from the lock date through December 31, 2026 form a partial-quarter descriptive pilot. It is reported separately and is not part of the formal count.
Formal count: starts with the first quarter of 2027. Each calendar quarter forms one cohort.
Qualifying: a cohort qualifies if it has at least 5 companies in each band. Cohorts that don't qualify are reported and not counted.
When a cohort is "mature." A cohort is mature 12 months after its last formation date, plus 30 days for data to settle.
When we report.
Schedule: once a year, each February, starting February 2028. The pilot first appears in February 2028; the first formal cohort appears in February 2029.
Contents: each report shows every mature cohort's gap, the running average, the share of cohorts that were positive, and the test statistic below.
Before 20 formal cohorts: reports are descriptive only, with no verdict in either direction.
The test statistic (fixed now).
Statistic: t = (average cohort gap) ÷ (standard deviation of cohort gaps ÷ √(n/4)), where n is the number of mature, qualifying formal cohorts. The n/4 is the same overlap adjustment the historical study used.
Threshold: the one-sided 5% critical value of Student's t with n/4 − 1 degrees of freedom. At 20 cohorts that is 2.13. A small-sample threshold is used because 20 overlapping cohorts are worth only about 5 independent observations.
Formal checkpoint at 20 mature formal cohorts (expected with the February 2033 report). The label is exactly one of:
Forward evidence positive and statistically confirmed: t is at or above the threshold.
Forward evidence negative and statistically confirmed (reversal): t is at or below minus the threshold.
No difference: t is between −0.5 and +0.5.
Not statistically confirmed — positive direction: the average is above zero, and none of the labels above applies.
Not statistically confirmed — negative direction: the average is below zero, and none of the labels above applies.
After the checkpoint, reporting continues every February as a descriptive record. Any further formal test would be registered separately before its data exist.
What changes the headline.
Reversal: only a statistically confirmed reversal relabels this study's headline, as "Did not replicate in our forward test," dated.
Any other result: labels only this section.
The historical record: the original historical results stay on the page exactly as published, with their date. Nothing is deleted or rewritten.
Changes outside the test: later changes in tax rules or buyback policy are recorded as context for interpreting results. They are not a reason to change or pause the registered method.
Suspension: if the method can't be run as written (for example, a required vendor field disappears or the data subscription ends), the test is labeled "Suspended: method could not be run as registered," with the reason. A suspension is not a failure. The test resumes only if the identical method can be restored.
Source. Sharadar; calculations by Brett Owens.
Every study here tells you what held and what didn't, the way this one did. Enter your email for the next one.
Replication
Every figure on this page is reproducible from the committed artifacts. Commits: 7cb38e9 (pre-registration, pipeline, results, 2026-09-09 14:19) · 6ca5d0c (sector-neutral cut, 15:25) · 534a73b (band table, 16:04) · 011a857 (companion study, 16:43). File hashes (SHA-256, first 16): pre-registration 2dd5c62dc74aa758 · build events 32252905dc4ea1c7 · returns 5b710905f7936958 · report 1a645943fa9745e4 · sector-neutral b7948249c25aaa85 · results 8d2e7922c32ba082 · frozen event set bc24035982db0d99. Complete list in file_hashes.json.
Chart data: figure1_bands_raw_and_sector_neutral.csv · figure2_cohort_spreads_raw_and_sector_neutral.csv · figure3_band_by_cohort_raw_and_sector_neutral.csv · worst_case_periods_by_band.csv · outer_bands_asrun_vs_screened.csv.
Source: Sharadar; calculations by Brett Owens.
Listed on the kill register. Changelog: v1, October 21, 2026, initial
Limitations
The primary result is borderline: the "discounted t" runs from 1.46 to 2.28 depending on method.
"Beat the market" means these stocks outperformed their sector peers, not the index at large.
The 5–10% band is exploratory; it has not yet been forward tested.
The pre-registration and results share one "commit," so we can't independently verify the plan was written before the results.
Returns are gross, equal-weighted, and cover $500M+ market cap US stocks from one data vendor.
Data Sources and Required Attributions
Source: Sharadar; calculations by Brett Owens.
This is research, not individualized investment advice. Nothing here recommends buying or selling any security. Yields, prices, and relationships shown are historical; past results do not predict future ones. Figures are as of the dates stated and will not be updated in place without a changelog note.
© 2026 How to Retire on Dividends
Comments