Methodology

Built on evidence.
Open to scrutiny.

Cadence reads public economic, consumer, search, advertising and news signals across eleven European markets. Every published figure traces to its source. The methods are here, and so are the tests that did not work.

One framework

Every market is read against its own history.

Cadence never ranks eleven markets on one scale. Each is measured against its own long-run mean, because a survey index has a different national baseline in every market and the levels are not comparable across them.

This quarter, Poland and Germany report almost the same consumer confidence: -14.0 and -14.5. Read as a league table they are the same situation. Read against their own norms they are opposite ones: Poland sits 2.8 points above its long-run mean, and Germany sits 5.1 below its own. One market is behaving better than it usually does and the other worse, and the two headline numbers do not tell you which is which.

Eurostat consumer survey · PL 2026-09 · DE 2026-09 · each against its own long-run mean

That is why the absolute number is never the headline. A market can look poor on the European scale while behaving normally for itself, and the opposite.

The signals

What Cadence reads, and what each one cannot tell you.

Eight families of measurement. Each is a specific instrument rather than a category, which is why the limits below are as much a part of the list as the descriptions.

Consumer surveys: Eurostat’s consumer survey: confidence, savings intent and major-purchase intent.; Limit Stated intent, not behaviour. Comparable to the same market’s own history, never to another market’s level.
Retail volumes: Eurostat retail turnover in volume, calendar adjusted and deflated.; Limit Goods moved, not money taken. A market-level series: it says nothing about any individual retailer.
Accommodation activity: Eurostat nights spent and arrivals recorded at accommodation establishments.; Limit Realised stays. Not spend, not inbound perception, and not the whole travel economy.
Search & app charts: Which brands are observable in a market and sector, and where they sit in the app charts.; Limit Coverage and position. How far a brand’s search attention moved is withheld this edition — see below.
Advertising activity: Creatives observed running, matched by verified advertiser domain.; Limit Capped at 40 per advertiser, so an absent count can mean an unmatched domain rather than a brand that is not advertising.
News tone: The salience and tone of news coverage, by theme.; Limit What was published, not what was believed. It is not public opinion and not social listening.
Household rates: ECB household mortgage and consumer-credit rates.; Limit Those series specifically. Not a general credit indicator, and not borrowing volumes.
Unemployment rate: Eurostat’s unemployment rate.; Limit The rate and nothing else. Not vacancies, wages or participation.

Two things we can measure and do not publish.

Branded search movement is withheld this edition. Repeated identical measurement did not reproduce the per-brand year-on-year figure at the precision Cadence publishes, and the archived production configurations show individual brands changing sign between runs. Neither the magnitude nor the direction is served. Which brands are observable is published; how far any of them moved is not.

Inflation is withheld in all 11 markets. Eurostat closed the series we used at 2025-12 for a rebase and has not published the replacement. A national consumer price index is not the same instrument, so we withhold the figure rather than substitute one that would not be comparable with what we published before.

Observation, interpretation, implication

Three different kinds of sentence, kept apart.

Most disagreements about market intelligence are really disagreements about which of these three a sentence is. Cadence labels them, so you can accept one and reject another.

Observed
What the instrument recorded, with its receipt. Checkable by anyone with the same public source.
Interpretation
What we think it means when two independent observations are read together. Ours, and labelled as ours.
Implication
What it might mean for how a brand should speak. A judgement you are entitled to disagree with, which is why it is never presented as evidence.

The separation is the reason a Cadence read can be argued with productively. If you reject an implication, the observation underneath it is unaffected.

Comparison and coverage

A signal we cannot publish is not a behaviour that is absent.

Cadence publishes a challenger-bank reading in 5 of 11 markets. Spain is one. Austria is not.

That does not mean Austria has no challenger banks, and it is not a reading of zero. It means the evidence needed to publish that comparison was not there this quarter. The behaviour may be identical in both markets; what differs is what we can observe well enough to put our name to.

Reading an absence as a finding is the most common serious error in this field, and it runs in both directions: a market with no published signal is not quiet, and a sector with no published movement is not stable. Cadence distinguishes we looked and nothing moved from we could not look, and says which one applies.

The same discipline applies inside a signal that does publish. In all 5 markets where the challenger-bank pairing is published, the movement is still withheld: we can tell you which challenger and which incumbent are being compared, and not how far either moved.

Every cross-market claim carries its denominator. Three of eleven, never “several”. A pattern that reduces to one entity is reported as that entity.

A country state is not a European state. A finding that holds in ten markets separately is still not a European pattern until it has been tested as one, and it is reported as untested until then.

Evidence and provenance

Every figure carries its receipt.

Not a bibliography at the end of a report. On the figure, where the figure is used, naming the source, the series, the market and the period.

Here is one as published, from the Austria edition. It reports a gap, so it carries a receipt for each side of it:

Food, beverages and tobacco: +1.9 % YoY volume; Eurostat sts_trtu_m · EUROSTAT_RETAIL_FOOD · 2026-08 vs 2025-08
Non-food excluding fuel: -0.4 % YoY volume; Eurostat sts_trtu_m · EUROSTAT_RETAIL_NONFOOD_EX_FUEL · 2026-08 vs 2025-08

Comparison basis both legs Eurostat sts_trtu_m VOL_SLS, same month, opposite signs

That last line is doing the work. Comparing two numbers drawn from different series is the most common way to produce a confident wrong answer, so when Cadence reports a gap between two things, both sides come from the same instrument and the same period — and the edition says so in its own words rather than leaving you to assume it. Where two figures cannot be compared on the same basis, we report that instead of reporting the gap.

Eligibility travels with the claim too. When a finding rests on three of seven observable entities, the published claim says three of seven — because a reader deciding whether their own market is typical needs the denominator, and it is the first thing a sceptic looks for.

What we tested

We tested whether the signals predict demand. They do not.

Cadence is a read, not a forecast. That is not caution and it is not positioning. It is what two pre-registered tests returned, and both are published in full.

Phase 1 — April 2026

We pre-registered a test of whether music sentiment forecasts French consumer confidence at one to three month lags across 2021–2025, controlling for confidence’s own autocorrelation and for macro covariates. Thirty-two tests, eight features at four lags, Bonferroni corrected.

It failed across all three components. Music features did not improve the forecast. Adding them made it worse.

Phase 2 — June 2026

Test B, the brand and category backtest — the one a predictive product would rest on — does not pass. Both arms fail. No locked feature-lag survived correction, and none of twenty-eight exploratory cells survived either. Against a three per cent improvement bar, nought of six categories reached it, and music degraded the forecast in four of the six.

Test A, the nowcast, is a non-confirmation. Run first on France alone at n=18 it was inconclusive, so we powered it across France and Germany at roughly n=45 each, which engages the generalisation bar we had locked in advance. The point estimates were consistently positive, between +1.7% and +9.9%, but none cleared significance and the bar was not met. France did not clear on its own. Powering up did not rescue it.

The careful word is “non-confirmation”. Not “validated”, which is reserved for clearing the bar. Not “disproved”, because the estimates lean positive. A properly powered test asked the question and did not answer it.

So no predictive, forecasting or “predicts demand” claim is made anywhere in Cadence. League tables are descriptive share of attention. That is the result, not a style of writing.

The null we nearly faked

Worth telling, because it is the reason to trust the result above.

The first implementation of the backtest had a silent bug. A prediction call raised an error that an over-broad exception handler swallowed, so both the baseline and the music model quietly fell back to predicting the training mean. Every category came back with the two models identical to five decimal places.

That is a null, and a null was the result we expected. It would have been the easy story: we tried, it did not work.

It was caught because the numbers were too tidy. Two genuinely different models do not agree to five decimal places. We fixed it and re-ran, and the real result is stronger than the fake one: music degrades the forecast in four of six categories.

The rule we wrote for ourselves afterwards: if a result looks suspiciously identical, suspect a swallowed exception rather than a clean null.

The boundary

What Cadence does not claim.

Not a forecast. A read of how a market behaves now, in the published quarter. Ask for a later one and Cadence declines rather than extrapolating.
Not causation. Two measurements moving together is a relationship, not a mechanism, and Cadence does not assert why.
Not individuals. Every series is market-level or brand-level. There is no person in the data and no profile of one.
Not commercial outcomes. No revenue, units, market share, footfall or conversion. Cadence does not know what you sold.
Not opinion. News tone is what was published, and consumer surveys are what people said. Neither is what a population believes.

If the evidence cannot support a sentence, we decline it and say why. That rule is the product.

Field Notes, including the failures → See a market read →