Economics explained
What a nowcast is, and why it disagrees with the official print
Private inflation and activity estimates arrive weeks before the government statistic and are built from different raw material. Understanding what they measure explains most of the gaps, and the gaps are where the trading opportunity sits.
In this story
Illustrative. Bank Season is an editorial prototype. This story, its sources, the issuers named in it and every figure it quotes are invented to demonstrate the publication. Nothing here is reported fact or investment advice. Read the disclosure.
Official economic statistics arrive late by design. A consumer price index for August is collected through the month, processed, seasonally adjusted and published in the middle of September, by which time the economy has moved on. Nowcasting is the practice of estimating the number before it exists, using data that is available immediately, and the private research houses that do it well have become a routine part of how rates markets are traded. This is what the method involves and where it goes wrong.
What is actually being estimated
A forecast predicts a future value. A nowcast estimates a present one that simply has not been measured yet. The distinction sounds pedantic and is not, because it determines how the estimate should be judged. A forecast can be wrong because the world changed. A nowcast can only be wrong because the estimation was poor, since the underlying reality was already fixed when the estimate was made.
The inputs are whatever exists at higher frequency than the official series. For inflation that means retail scanner data covering tens of millions of transactions a week, posted prices scraped from online listings, transaction level data from payment processors for services categories, and administrative records for rents and medical billing. For activity it means payroll processing files, freight bookings, electricity load and card spending.
Typical nowcast inputs against the official equivalent
| Component | Official collection method | Nowcast input | Lead time gained |
|---|---|---|---|
| Food and household goods | Price collectors sampling outlets monthly | Retail scanner transactions, weekly | 3 to 5 weeks |
| Rents | Panel of units surveyed every six months | New lease signings from listing platforms | 9 to 14 months |
| Medical services | Billed charges sampled quarterly | Claims level payment records | 6 to 10 weeks |
| Airfares and lodging | Fare and rate sampling on set dates | Booking level transaction data | 4 to 6 weeks |
| Employment | Establishment survey with partial response | Payroll processing files | 2 to 4 weeks |
Lead time is the typical interval between the nowcast input becoming available and the corresponding official figure being published.
The rent row explains most of the disagreement between private and official inflation estimates over the past four years. Official shelter measures rents across a panel of housing units surveyed twice a year, so the published series reflects the average of leases signed over the preceding twelve months. A nowcast built on new lease signings measures what a tenant signing today would pay. Both are correct measurements of different things, and the gap between them can exceed two percentage points for a year at a time.
How accurate they actually are
The honest way to judge a nowcast is to compare each published estimate against the figure that eventually arrived, across a long enough sample to include a few surprises. The Marlowe Institute core services nowcast has a mean absolute error of 0.14 percentage points on the three month annualised measure over the past three years. That is small against a series whose month to month variation is frequently four times larger, which is why the estimate is worth watching even when it is wrong.
0.14 pts
Mean absolute error, Marlowe core services nowcast
Against the eventual published figure, three year sample
Nowcast error against the published figure
Nowcast estimate minus the eventual official reading, percentage points, core services three month annualised.
Two features of that chart matter. The errors alternate in sign rather than accumulating, which means the model is not biased in one direction. And the two largest errors, in the third quarter of last year and the first of this one, both came in quarters when a category with thin private data moved sharply. Nowcast accuracy is not uniform across components, and the weakest coverage is always in the categories that are hardest to observe from transactions.
Where they reliably fail
- Categories with no transaction footprint, such as imputed rent on owner occupied housing, which has to be modelled rather than measured and is roughly a quarter of the core basket.
- Seasonal adjustment, which the official agencies revise annually using methods a private house cannot replicate exactly, and which accounts for a large share of small discrepancies.
- Quality adjustment in electronics and vehicles, where the official treatment records a better product at the same price as a price decline and private estimates often do not.
- Turning points, because a nowcast is calibrated on the relationship between inputs and outputs during a period when that relationship held, and a structural break invalidates the calibration before it shows in the error.
- Revisions, since the official figure a nowcast is judged against is itself revised, sometimes twice, and a nowcast that looked accurate on publication day can look poor a year later.
The turning point problem is the one that costs money. Every nowcast performed well through the goods disinflation of 2024 and 2025, because the relationship between scanner prices and the published index was stable. The June crossover in core goods, when the year on year rate moved from negative to flat, was picked up by the private estimates about six weeks before the official series confirmed it. That six weeks was worth a great deal to anyone positioned in the front end, and it is the entire commercial case for the product.
The nowcast earns its keep in the six weeks before a turn and is redundant for the two years in between. Clients who subscribe for the monthly print rather than for the turns have misunderstood what they are buying.
Marguerite Paz, senior economist at the Marlowe Institute
Reading one sensibly
Treat the level as approximate and the direction as informative. Check whether the gap to the official series is stable before drawing any conclusion from a single month. Look at which components drove the estimate rather than the headline, because a nowcast moved by airfares carries far less signal than one moved by medical services or rents. And remember that the same research house publishes the payroll and quits estimates that the labour side of the debate rests on, so the inputs are not independent of each other.
The current disagreement is a good example of all of this. The Marlowe nowcast has core services excluding housing at 3.4 per cent annualised, essentially unchanged since March, while the front end of the Treasury curve is priced for easing that requires that figure to fall. One of those is wrong and the nowcast will be the first to say which, at 13:30 New York time on the day it publishes, about two weeks before the official statistic settles the matter.
