Why consensus estimates differ from site to site

How it works

A consensus estimate is assembled in three layers, and each layer makes choices.

Analysts at banks and brokers publish forecasts of a company's earnings per share, revenue and other lines, inside research notes and models. Each analyst picks a basis — usually the company's own adjusted figure, sometimes GAAP — and revises on their own timetable.

Collectors gather those forecasts under contributor agreements and turn them into one number. Four names do most of this: LSEG's I/B/E/S, FactSet, S&P Global (its Capital IQ estimates, plus Visible Alpha, which it bought in 2024, as where research terminals get their data describes) and Zacks. Each decides which forecasts count, how old one may be, which accounting basis is the default, and whether the headline is a mean or a median.

Sites license one collector's output and then choose again: which basis to display, fiscal or calendar years, the reporting currency or the listing's, and how often to refresh. A broker's research tab is a site in this sense. Fidelity's research centre, for one, publishes a page on the Refinitiv I/B/E/S consensus it shows, describing I/B/E/S as fed by more than 900 contributing brokers.

Two sites showing different consensus figures for the same quarter are usually not wrong. They are answering slightly different questions.

Who supplies which site

What each product says about its estimates, read on 9 October 2026. Not documented means the vendor's public pages do not say.

Four of the seven resell S&P Global. That narrows the disagreement among them to basis, period and refresh timing — and still does not remove it. On 9 October 2026 Yahoo's Normalized consensus for Novo Nordisk's 2026 EPS was 22.53 kroner from 8 analysts, while Stock Analysis, also crediting S&P Global, showed 21.56 kroner on its adjusted basis. Neither page says which of the supplier's EPS measures it uses.

GAAP, adjusted and the company's own definition

The biggest gap is usually the basis, and the rule that creates it is Regulation G. A US company that publishes a non-GAAP measure — adjusted EPS, say — must present the most directly comparable GAAP measure alongside it and a reconciliation between the two. So both numbers are always public, and analysts are free to forecast either. Where a company guides only on its adjusted figure, as AbbVie does below, most analysts forecast that one. The non-GAAP entry covers the definitions.

The collectors resolve the split by majority. FactSet groups estimates into consensus classes by methodology and computes "one class at a time because creating an average across different classes can be misleading". LSEG says I/B/E/S actuals "align reported company results with the operating or comparable basis used by the majority of contributing analysts". Zacks adjusts estimates "to include or exclude selected non-recurring (NRI) items" to reach its consensus of EPS before non-recurring items.

AbbVie shows the size of it. Its release for the fourth quarter of 2025, furnished on 4 February 2026, reported diluted EPS of $1.02 on a GAAP basis and adjusted diluted EPS of $2.71, with the reconciliation between them, and gave 2026 guidance only for adjusted diluted EPS. On 9 October 2026, for the quarter to 30 September 2026:

For full-year 2026 Yahoo showed $7.89 GAAP from 11 analysts and $14.00 Normalized from 27, Stock Analysis $14.00 adjusted, and Zacks $14.02 from 10 estimates. The three adjusted figures agree to within two cents; the GAAP figure is a different quantity, forecast by a minority of analysts.

Yahoo's switch has a default, and it is not the same for every stock. The Apple page served on 9 October 2026 opened on GAAP; the AbbVie and Novo Nordisk pages opened on Normalized. For Apple it hardly matters — $8.82 against $8.83 for fiscal 2026. For AbbVie the default decides whether the headline is $7.89 or $14.00.

Which analysts, and how old their numbers are

Each collector has its own contributor list, so the panel differs before any rule is applied: that is why AbbVie's quarter counts 18 analysts on one site and 7 estimates on another. Then each decides when an estimate is too old to count.

• FactSet, which describes its point-in-time database as fed by over 800 contributors, builds its default consensus from estimates "validated via broker research within the past 100 days", and its point-in-time product takes the consensus as of local midnight for each company, so nothing entered after midnight counts toward that day.
• LSEG publishes a plain mean and, separately, StarMine SmartEstimates, which first exclude outliers, "estimates older than 4 months, and stale pre-revision figures", then weight what is left by recency and each analyst's track record. An "LSEG consensus" can be either.
• Zacks drops "all individual estimates made prior to the guidance that fall outside of its range" once a company issues guidance, and keeps everything made after it.
• Finnhub counts buy-side analysts and its own estimates alongside the sell side.

Regulation FD is what makes a guidance-date rule workable. A company that tells a broker, an adviser or a fund material nonpublic information has to make it public at the same moment if the disclosure was intentional, or promptly if not. Guidance is therefore a public, dated event, and a collector can sort estimates into before and after it.

The moment matters as much as the rule. Yahoo's EPS trend table shows its consensus as it stood 7, 30, 60 and 90 days earlier; AbbVie's current quarter moved from $3.86 to $3.81 over those 90 days. A site that refreshes daily and one that caches for a week will disagree by however far the estimates moved in between, most of all in the weeks around a report, when revisions bunch.

Fiscal years, currencies and what "per share" means

Fiscal years. "Current year" is the company's fiscal year unless a site says otherwise. On 9 October 2026 Yahoo labelled Apple's year ending 30 September 2026 and Broadcom's ending 31 October 2026 both "Current Year (2026)". A source that converts to calendar years splices quarters from two fiscal years into one figure. Koyfin lets you choose "between fiscal and calendar periods"; check which one a site is on before setting two side by side.

Currency. Collectors store an estimate in the currency it was made in and convert on request: FactSet's manual lists a currency parameter and a separate estimate-currency flag for securities whose local currency differs from the estimates'. Novo Nordisk, listed in New York as an ADR, shows how far apart that leaves the sites. On 9 October 2026 its 2026 EPS consensus was 22.53 on Yahoo, labelled "Currency in DKK", and 21.56 on Stock Analysis, which notes "Financial currency is DKK". Zacks showed 3.53, with no currency label on the table, beside the ADR's $38.18 dollar quote; its revenue consensus of 46.10 billion, against Stock Analysis's 300.79 billion kroner, can only be in dollars. A forward P/E computed from a dollar price over a krone EPS is wrong by the exchange rate.

Mean, median, and how a beat is scored

The headline consensus is a mean on most sites. Yahoo calls it "Avg. Estimate"; Koyfin says its main figures are "the average of consensus" and offers median, high and low as extra columns; LSEG measures "beats and misses using comparable actuals against consensus mean estimates". FactSet's surprise report lets the user choose mean, median, high or low. With twenty forecasts the mean and the median are close; with four, one outlier moves the mean a long way.

The actual is a choice too. LSEG adjusts actuals "to the consensus basis where reported figures contain unusual, one-time or non-comparable items". FactSet scores US, Japanese and Australian surprises against the company's reported actual, and the rest of the world — mainly Europe — against the median of the post-event consensus, the "Broker Actual". That manual was last updated in June 2015; FactSet's point-in-time page, read on 9 October 2026, still describes the same 100-day default consensus window.

So a beat is a comparison of two chosen numbers. On Yahoo's AbbVie page, the fourth quarter of 2025 was a 2.16% beat on Normalized — $2.71 against $2.65 — and a 64.51% miss on GAAP — $1.02 against $2.87. Percent surprise misbehaves near zero as well: for the third quarter of 2025 the GAAP consensus was −$1.19, the GAAP actual $0.10, and the page printed a 108.40% surprise. What any surprise means for the share price is outside what these numbers describe.

What you can do about it

Find the supplier before comparing. Look for a data-providers page, a footnote or a help article. If two sites resell the same collector, a difference is basis, period or timing; if they resell different collectors, the panels differ as well and the figures may never match.

Put the bases side by side. Where a site has a GAAP / adjusted switch, read both. Otherwise take the company's own reconciliation — the table Regulation G requires in the earnings release — and check which of its lines the consensus is closest to. An adjusted consensus should be compared with the adjusted actual in that release, and a GAAP one with the GAAP actual.

Check the period end, not the year label. Compare the fiscal period end dates; if one source uses calendar years, convert the other or do not compare.

Check the currency and the share. For a foreign company, look for the currency label on the estimates table rather than the quote, and for an ADR, the ratio to ordinary shares.

Look at the analyst count and the spread. Four estimates with a wide high-low range is a different kind of number from thirty in a tight one. A GAAP consensus from a handful of analysts on a company that guides on an adjusted basis is a minority view, not the market's expectation.

Record which consensus you used. If you keep estimates or surprises in a spreadsheet or a model, store the supplier, the basis, the period end, the currency and the date you read it with each number. Getting an earnings calendar as data covers the API fields, several of which document none of these.