The Layered Cake: what if we looked at the noise through a user’s eyes?

A backstage note on how the model came about, and what it turned out to explain.

I want to walk back through a moment that, in hindsight, split our monitoring practice into a before and after — not a formula, but a mental image that made the formula necessary.

01

The single-number problem

For a long time we treated a media peak the way most monitoring dashboards still do: as a single number. A brand gets 20,000 mentions in a month, sentiment comes out to +0.25, and that is the story we tell about the event. Somewhere in the fintech corpus, though, that story kept refusing to add up. A crisis would produce a mild, almost forgiving aggregate score. A neutral corporate update would produce a swing that seemed too large for what had actually happened. Comparing the size of a peak to its overall volume explained some of this — but not all of it, and not the cases that mattered most.

02

The reframe: one person, one moment

The reframe came from a fairly simple exercise: stop looking at the whole month of mentions the way an analyst does, laid out on a chart, and instead imagine looking at it the way one ordinary person would — someone who opens their feed at a single moment in time, with no dashboard, no aggregation, no idea that “30% of this is positive and 70% is negative.” That person does not experience a distribution. They experience a moment. If the positive material happens to sit on top — because it was posted more recently, because the platform’s ranking surfaced it, because it simply arrived later — that is the entire signal they get. The other 30%, or 70%, buried underneath, might as well not exist for them right now.

Once I pictured that — a person standing in front of the wall of numbers we build, but seeing none of it, seeing only whatever layer happens to be on the surface at that instant — the aggregate tone we report started to look less like a fact and more like an artifact of measurement. We had been averaging across a stack, when the stack itself was the phenomenon worth explaining.

That is where the anchor to the previous period came from. If what sits underneath is simply what was already there — the ordinary, ongoing conversation about a brand that never really commented on the new event at all — then its tone should look like last month’s tone. Not a new assumption bolted onto the model, but a direct consequence of the mental picture: the buried layer is buried precisely because it is old.

03

The two layers

That gave us the split we needed: an inertial layer, roughly equal in size to the brand’s average monthly baseline volume, carrying forward the tonality of the month before; and a reactive layer — the actual increase in volume — carrying whatever the audience actually thought about the event. Once you have both, you can solve for the reactive tone directly, instead of reading it off the aggregate and hoping the aggregate was telling the truth.

04 · The case that convinced us

Klarna, February 2024

The case that convinced us this was not just a tidy idea was Klarna’s AI-related workforce reduction in February 2024. The observed sentiment for that peak was −0.31 — negative, but only moderately so, the kind of number that might suggest a bruised but manageable reputational hit. The reactive sentiment, once we stripped out the inertial layer, came out to −0.66 — more than twice as negative. The difference was not noise. It turned out that 71% of the positive mentions counted inside that month’s aggregate were inertial residue: leftover enthusiasm from January, when Klarna’s CEO had been actively promoting a positive narrative about AI-driven efficiency. Almost three-quarters of the “positivity” the dashboard reported in February had nothing to do with the layoffs at all — it was an echo of a different story, one that had already ended.

That is the moment the imagined user turns out to matter again. Someone opening their feed during that February peak, at almost any given moment, would have seen mostly the layoffs conversation — sharp, current, negative. The +0.31-to-−0.31 aggregate swing we reported was true in a statistical sense and false in an experiential one. The person the aggregate was supposedly describing was not actually there.

05

The pattern

Ten of the eleven fintech events we later ran through this decomposition showed the same pattern — the inertial layer dilutes the true reactive signal, understating negative events and, less often, inflating mild positive ones. One case, Wise, ran the other way: a genuinely lukewarm reaction to weak earnings guidance got dressed up by an unusually positive previous month into something that read as “fairly well received,” right before the stock dropped.

None of this changes what the underlying event was. It changes what we are allowed to claim the audience felt about it — and, just as important, it brings the method back to where it started: not with a chart, but with a person standing in front of one layer at a time, never seeing the whole cake.

How we know this

Methodology & scope

This is a “behind-the-method” note, not a peer-reviewed paper. Figures cited (Klarna −0.31 / −0.66, 71% inertial residue; 10 of 11 events) come from ICDS internal analysis of the fintech corpus. Replace bracketed placeholders with your final dataset details before publishing.