ADP vs Nonfarm Payrolls: three years of monthly comparisons, a regression and a direction-agreement rate

What this page does

ADP payrolls, widely known as the "mini nonfarm", are released two days before the Bureau of Labor Statistics payrolls report, so the market has long treated them as a trailer for the main event. But the two are built by different organisations, from different samples, and they are revised on different schedules. This page lines up the 36 months from 2023-09 - 2026-08 side by side, recomputes the gaps, direction, correlation and linear regression, and labels the provenance and verification status of every row.

Every figure comes from the same pipeline as the site's daily dashboard: payrolls from the East Money datacenter-web endpoint, and ADP from the same family of sources the site's ADP fetch uses. Both were pulled on 2026-09-20, and the payroll side already includes subsequent BLS revisions.

Definitions and sourcesRead this first: it governs how the numbers below should be used

Headline statistics

All 36 pairs in the window. ADP averages 9.7 and payrolls average 7.8 (10k persons per month). Regression: NFP = 0.40 × ADP + 3.96, in units of 10k persons.

Monthly job gains over the last 36 monthsunits: 10k persons

Scatter and regression: can ADP predict payrolls?

Regression result (36 pairs, units of 10k persons): slope beta = 0.40, intercept alpha = 3.96, so NFP = 0.40 × ADP + 3.96. The correlation coefficient is r = 0.27, the coefficient of determination is R2 = 0.07, and the residual standard deviation is 9.1.

How to read it: an R2 of 0.07 means ADP explains only about 7% of the monthly variation in payrolls. The other 93% comes from government hiring, sample differences, seasonal adjustment and the birth-death model. The slope is far below 1, which is the statistical statement of a familiar fact: a move in ADP does not pass through one for one into payrolls.

Subsample robustness check

The contrast is the finding. Over the first 24 months (2023-09 to 2025-08) the two series agreed on direction 91.7% of the time with r = 0.28. Over the last 12 months (2025-09 to 2026-08) agreement falls to 58.3% (7 of 12) with r = 0.03 and R2 = 0.00; the linear relationship has essentially disappeared. One caveat matters here: 10 of the 12 ADP values in that recent window sit in the segment that could not be re-verified online, while the ADP values in the first 24 months all passed live verification, so the two blocks are not equally trustworthy. With only 36 observations, subsample results are a robustness check rather than a stable rule.

Full data table2023-09 - 2026-08

"Agree" means ADP and payrolls were both positive or both negative. The consensus column comes from the same historical series for 2023-09 to 2025-08 and from the TradingEconomics calendar for 2026-07 and 2026-08 (its July figure is 7.0 against 6.8 in the previous version, reflecting different consensus panels), with the earlier published values carried over for 2025-09 to 2026-06. The ADP data status column marks each month as fetched live for this update or carried from the previous version. Neither 2026-09 figure has been released yet.

Payroll revisions: the first print is not the final number

The payroll "current value" moves as the BLS revises. The site's own revision tracker records the latest drift (units of 10k persons):

July 2026 is the single most important change in this update. The first print of -2.3, widely read at the time as employment turning negative, now stands at +2.1, a drift of +4.4. The previous version of this article used July 2026 as its flagship failure case, "ADP +4.4 against payrolls -2.3, the two pointing in opposite directions". On the current values both are positive and the month now agrees on direction. That example has therefore been dropped in favour of the widest gaps below.

The four widest gaps by |NFP - ADP|:

Of the four, 2024-10 agreed on direction but missed badly on magnitude (ADP +23.3 against payrolls +3.3, a gap of -20.0); the other three pointed in opposite directions. A miss can therefore mean either a magnitude error or a sign disagreement, and the two are not the same kind of failure.

The months that disagreed on direction (7 of 36)

5 of these 7 disagreement months fall inside the ADP segment that could not be re-verified online (2025-09 to 2026-06), including four consecutive months from 2025-09 to 2025-12. The agreement rate and correlations for that stretch should be discounted accordingly.

How the two relate, and why ADP is only a reference

• The lead relationship: ADP arrives two days early and the market uses it to calibrate expectations for payrolls. The sequence is ADP on Wednesday, payrolls on Friday.
• Direction broadly matches: across all 36 months the two agreed 29 times, a rate of 80.6%. But over the last 12 months agreement drops to 58.3% (7 of 12), so even the directional hint has weakened sharply in the most recent year. Note that 10 of those 12 ADP values could not be re-verified online; see the robustness check above.
• Magnitude often misses badly: r = 0.27, R2 = 0.07, and the mean absolute gap is 8.1. The largest single-month gap is 23.1 (2025-01: ADP +18.3 against payrolls -4.8). On average payrolls print 1.8 below ADP, so the two are not related by a simple shift either.
• Why it fails: ADP excludes government payrolls, so swings in public hiring push the two apart. Its sample skews to mid-sized and large firms, under-covering start-ups and small businesses. Both series are extrapolated from samples and both are revised, and the birth-death model is prone to systematic bias around turning points. Strikes and seasonal hiring in government and education appear only in payrolls.
• What it means for gold and for trading: a big ADP beat usually hardens tightening pricing and lifts the dollar and real-rate expectations, pressuring gold intraday, with the reverse for a big miss. But the statistics here show a very weak quantitative link to payrolls (R2 = 0.07) that has weakened further over the last 12 months, so ADP should not be used on its own to bet on a payrolls move. Combine it with the surprise versus consensus and with rate pricing.

What changed versus the previous versiondid the new data move the conclusion?

The previous version was generated on 2026-09-02 with 35 pairs and payrolls on a first-print basis. This update was fetched on 2026-09-20 with 36 pairs and payrolls on a current-value basis that includes revisions.

What moved: the core qualitative conclusion is unchanged, in that ADP is a directional reference only and cannot be used to forecast payrolls quantitatively. But the correlation and the fit are weaker, not stronger: r falls from 0.30 to 0.27 and R2 from 0.09 to 0.07. Direction agreement edges up to 80.6% over the full window, yet over the last 12 months it falls to 58.3% with an R2 close to zero, which would say that even the directional claim does not hold in the most recent year. That said, 10 of the 12 ADP values in that recent window could not be re-verified online, so the claim of a recent breakdown should be treated as a live hypothesis rather than an established fact. The previous version's flagship divergence case, July 2026, has been overturned by the payroll revision, and the page has been corrected accordingly. Every statistic here was recomputed from the new data; no data was selected to preserve the earlier conclusion.

Frequently asked questions