Does real-time inflation data lead the CPI? The Lead Test — 93 prints graded, every input audited

Two claims hide in "leads CPI". The first — that a real-time gauge's reading on the eve of a print is a better guess of the print than the previous print — is not established: the difference is 0.03pp in the gauge’s favour, under the 0.05pp margin this page requires before calling either the better guess: since 2019 the previous print misses the next one by 0.32pp on average and the Nowflation CPI Gauge by 0.30pp (93 prints; PCE-weighted 0.27pp). The second — that the gauge's reading today carries information about CPI months ahead that CPI's own history does not — is the testable one: since 2019 the Gauge's monthly YoY correlates 0.74 with official CPI six months later against 0.74 for CPI predicting itself over the same 84 months; under the lead-test rule (clear persistence by at least 0.03 at some horizon from 5 to 12 months) the Gauge ties it, at best 0.24 vs 0.23 at 12 months; in a joint regression of CPI six months ahead on both, the Gauge's t-statistic is 0.91 and CPI's own lag 0.37, so neither coefficient reaches the |t| of 2 this page treats as significant (R² 0.549 on the Gauge vs 0.546 on CPI itself, 0.550 with both, n 84). Since July 2023 — a flat stretch of CPI with little variance to explain — the six-month figures are 0.42 vs 0.19 on 30 months (the Gauge clears persistence under the same rule there, on a sample too small to lean on). Last print: Sep 2026 CPI 3.7% (released 2026-10-14); the Gauge read 3.43% the evening before (gap -0.27pp), the previous print was 3.4%.

Correlation with official CPI k months ahead — 2019-01 to today (paired months: 91 at k = 0, 80 at k = 12)

Since July 2023 (37 paired months at k = 0): Gauge k0 0.89, k1 0.61, k2 0.31, k3 0.15, k4 0.12, k5 0.21, k6 0.42, k7 0.36, k8 0.20, k9 -0.01, k10 -0.29, k11 -0.56, k12 -0.65; CPI's own persistence k0 1.00, k1 0.79, k2 0.46, k3 0.21, k4 0.14, k5 0.10, k6 0.19, k7 0.22, k8 0.22, k9 0.15, k10 -0.07, k11 -0.40, k12 -0.66.

Every live input against the BLS subindex it feeds — lead-lag on monthly YoY (Pearson r)

The rent lead: the tenure-weighted all-tenant rent model (chain-linked Zillow ZORI / Apartment List new-lease blend, 4.5% monthly turnover over 48 months) leads CPI rent of primary residence by 5 months at r 0.98 (0.86 unshifted; 0.89 for CPI rent's own persistence over the same horizon); the raw new-lease blend peaks at a 13-month lead (r 0.89). Shifted forward 5 months, the model's latest reading of 1.94% YoY (Sep 2026) is where it says CPI rent (2.78% in Sep 2026) is headed by Feb 2027 — a reading, not a graded forecast.

The graded table — each variant on the eve of the release vs the print, last 24 prints

Print-eve summary since 2019 (93 prints): The Gauge MAE 0.30pp, r 0.986; PCE-weighted MAE 0.27pp, r 0.989; previous print MAE 0.32pp, r 0.980. Last 12 prints: The Gauge 0.21pp, PCE-weighted 0.20pp, previous print 0.30pp.

Method: Print eve: each variant's YoY on the calendar day before every CPI release since 2019-01 (reference month = the month before the release; the annual seasonal-factor revision that shares February with the January print is dropped), graded against the print as BLS quotes it — the NSA CPI-U 12-month change at one decimal — with the previous print as the naive benchmark; k=0 is that print (the month before the eve, the frozen standard), k=1 the print for the month the eve sits in, both with beats-naive counts and an autocorrelation-adjusted effective sample. Lead: the monthly average of each variant's daily YoY correlated with official CPI YoY k months later, k = 0–12, scored on the same months as two naive benchmarks — CPI's own persistence corr(CPI_t, CPI_t+k), and the real-time naive corr(CPI_t−1, CPI_t+k), the last print actually published when the gauge month closed; regressions of CPI(t+k) on the gauge, on CPI itself, and on both, with the joint t-statistic. Components: Pearson r on monthly YoY between each live input and the BLS subindex it feeds, the input shifted forward k months (rents 0–18, pump price 0–3, Manheim 0–6, farm gate 0–9); the naive columns are the subindex's correlation with its own value k months later and with its last published month. New-lease rent = chain-linked new-lease blend (Zillow ZORI 0.8 / Apartment List 0.2, renormalized over the sources present — the shelter model's own blend); all-tenant rent = tenure-weighted average of that blend with 4.5% monthly turnover over 48 months (the shelter model's own parameters). Latest vintage of every series; no smoothing, no window search, no parameter tuned to the result. Nothing here is tuned to look good; the naive benchmark is on every chart. Related: The Gauge, Exposed, Why the Gap?, Scoreboard, methodology.

Zillow data: Data Provided by Zillow Group.

Related: How the Gauge is built · Compare the forecasters · The scoreboard · Methodology