Why two providers give different returns for the same stock

How it works

Sort corporate actions by what each does to a price series and there are four kinds.

Changes to the share count and nothing else — forward splits, reverse splits, stock dividends. The price gaps by a known ratio on the ex-date and no value moved.

Value leaving the company for its holders — ordinary and special cash dividends. The price gaps down by roughly the cash paid; the holder is whole and the series is not, unless the cash goes back.

Changes to what the security is — spin-offs, mergers, rights issues. Value leaves the parent for somewhere else, and adjusting the parent means knowing what that somewhere else is worth.

Changes to the label only — ticker, name and exchange changes. No arithmetic, and the kind that splices two unrelated companies together, because symbols get reused.

Two factors do nearly all the work. The split factor is the ratio: for an N-for-1 forward split every price before the ex-date is divided by N and every volume multiplied by N. It is exact, and the only thing to get wrong is the date. The dividend factor is 1 − D/P, where D is the dividend per share and P the closing price the day before the ex-date; every earlier price is multiplied by it, and successive dividends compound. Nothing in that second formula is observed: P is a choice — previous close is conventional, ex-date open is defensible — and D can be revised after the fact.

Then the choice almost nobody writes down: ratio or difference. A ratio adjustment multiplies prior prices by a factor; a difference adjustment subtracts the cash. Over five years of a large-cap stock the two land within pennies. Over thirty they do not: subtracting three decades of dividends from a price that was $3 in the 1990s drives the series through zero, which a chain of factors each between zero and one cannot do. Difference preserves dollar amounts exactly, ratio preserves percentage returns exactly, and you get one or the other.

For continuous futures that stops being academic, because roll gaps are large and recurring and a backtest wants both properties at once. CSI Data builds eight continuous contract variants, back-adjusted and proportionally back-adjusted among them, because the choice is genuinely open.

A worked example: NVIDIA, June 2024

NVIDIA announced a ten-for-one forward stock split in its first-quarter fiscal 2025 results on 22 May 2024. Each record holder as of the close on Thursday 6 June 2024 received nine additional shares, distributed after the close on Friday 7 June, and "trading is expected to commence on a split-adjusted basis at market open on Monday, June 10, 2024."

Three dates, and the one adjustment turns on is neither of the two you would guess. A split is a distribution of 25 per cent or more of the value of the security, and under FINRA Rule 11140(b)(2) "the ex-dividend date shall be the first business day following the payable date" — hence 10 June, not the record date of 6 June. For ordinary dividends, under 11140(b)(1), the ex-date "shall be the record date", true only since SR-FINRA-2023-017 amended the rule effective 28 May 2024 alongside the move to T+1 settlement. Code written before that derives the ex-date as record date minus one business day, and is a day out on every dividend in the market.

The same release sets a second trap: NVIDIA raised its quarterly dividend to "$0.01 per share on a post-split basis", payable 28 June to holders of record on Tuesday 11 June. So 10 June is the split ex-date and 11 June the dividend ex-date — two different factors on consecutive days.

Now look at Nasdaq's own historical quotes: a close of $120.888 on 7 June 2024, then a $120.37 open and $121.79 close on 10 June. The second pair are prints. The first is not — nobody traded NVDA at $120.888 on 7 June, because the split did not take effect until Monday's open. The tape said $1,208.88, and $120.888 is that divided by ten, carried to three decimals because the factor moved the price off the cent grid.

That third decimal is the tell. If the pre-split values in your series are always an exact tenth of a whole number of cents, the only factor applied is the split; if not, a dividend factor is in there too. The factor for that dividend, 1 − 0.01/121.79, applies to every price before its 11 June ex-date and would have moved the 10 June close to $121.78; Nasdaq reports an unrounded $121.79, so its series carries splits and not dividends. A defensible convention, and not written anywhere on the page.

Total return and price return drift apart, permanently

The gap is small per year and enormous per decade, and MSCI makes it easy to see because it publishes the same index three ways. For the ten years to 31 August 2026 the MSCI World Index returned 13.56 per cent a year gross of dividends, 13.01 per cent net of withholding tax, and 11.19 per cent on price alone. From 31 December 1987 to the same date, gross returns annualised at 9.08 per cent against 6.67 per cent — compounded over those thirty-eight years and eight months, the total return series ends at about 2.37 times the price return series. Three published answers for one index, all correct, differing by more than a factor of two over a career; if nobody tells you which one you hold, a Sharpe ratio computed from it is uninterpretable.

Spin-offs are the case one price series cannot solve

Everything above computes from the parent's own history. A spin-off does not, because the size of the departure is set by the price of a security that did not exist the day before.

General Electric distributed GE Vernova on 2 April 2024, one GEV share for every four GE shares held on the record date of 19 March. The number a price series needs is in GE's attachment to Form 8937, the filing that reports organizational actions affecting the basis of a security. Using the New York Stock Exchange opening prices that morning — $140.53 for GE and $142.85 for GEV — 79.74 per cent of a holder's basis stayed with GE and 20.26 per cent went to GEV. The back-adjustment factor for GE's pre-spin history is that 0.7974.

Two things to take from it. The factor required GEV's own opening price, which is not derivable from GE's series by any method. And GE's own wording: "U.S. federal income tax law does not specifically prescribe how you should determine the fair market values", and "other approaches to determine fair market value may also be possible." The issuer declines to fix the convention. Vendors pick one silently, and many skip spin-offs altogether — FirstRate Data says so plainly, which is rare and right.

Where the events are published, and why none of it is a feed

There is a chain, and it produces events rather than factors. It starts with the issuer: under 17 CFR 240.10b-17, failing to give notice of a dividend, a stock split or a rights offering is itself a manipulative device, and notice must reach FINRA "no later than 10 days prior to the record date" or go to the exchange under comparable procedures. FINRA or the exchange then designates the ex-date under Rule 11140. Nasdaq publishes its own Daily List of dividends, splits, listings, delistings and symbol changes, archived back to 1999 — behind a data feed request form, a Nasdaq Global Data Agreement, a price list and a separate S&P licence for the CUSIPs. FINRA publishes its own daily list for OTC equities, and DTCC distributes announcements to its participants as ISO 20022 messages, against its own data dictionaries, message specifications and file layouts. Four publishers, four shapes, and not one of them is a price series.

A backtest needs none of those. It needs one adjustment factor per symbol per day, joined to prices, complete across delistings and renames, and stable when a figure is corrected weeks later. Turning heterogeneous notices into that table for every listed name is the work being sold, and why corporate actions handled correctly is among the first things the free layer does not reach.

How to test a series you already have

In rising order of effort, and stop as soon as one fails.

Look for an unexplained single-day move near a known ex-date. Scan for daily moves beyond, say, 30 per cent and check each against the issuer's filings. A phantom 90 per cent drop is a split on the wrong date, or no split at all.

Reconstruct one split by hand, terms from the issuer's release and ex-date from the exchange: the adjusted close for the day before the ex-date should be the raw print divided by the ratio, exactly.

Compare a long-horizon return against a published index, total return and price return both. Whichever it matches is what you hold.

Take a high-yield name and compute both returns. If they agree over five years on a stock yielding four per cent, there are no dividends in your data at all.

What you can do about it

Store unadjusted prices plus a separate corporate actions table, and adjust at query time. This is the one architectural decision worth getting right on day one, because it is the only one that is expensive to reverse. A store holding only adjusted prices cannot be re-derived when a vendor changes convention, cannot absorb a late dividend correction, and cannot answer what the price was on the day — which is what position sizing, tax lots and any broker reconciliation need. Rebuilding it means refetching the whole history from a vendor who may no longer sell it on the same terms. Choosing a database for tick data covers the engines, and tick data storage the products.

Ask your vendor which convention it uses, in writing, and expect silence. Four questions: are dividends in the adjusted series; is the factor multiplicative or additive; what price does it divide by; are spin-offs handled. Tiingo ships end-of-day history adjusted or raw; FirstRate Data ships unadjusted, split-adjusted and split-and-dividend-adjusted side by side; Norgate Data ships unadjusted and back-adjusted continuous futures. Shipping the variants separately is the honest shape.

Never compare two sources until you have reproduced one known event by hand in both. Otherwise you get a disagreement and no diagnosis, and you keep the source whose numbers you liked.

Know whether your series is total return or price return before quoting a Sharpe ratio at anybody, and write it next to the number, the way a survivorship-bias caveat belongs next to a backtest — same reasoning as for historical index constituents. Check the benchmark is the same kind: a total return strategy against a price return benchmark manufactures about two percentage points a year of alpha out of nothing.

Re-run the backtest you already trust on the corrected data. The rest of the field is in market data APIs, and every engine in backtesting frameworks will run happily on a series with the dividends missing and report a confident number for it.