The Weather Forecaster Who Only Looks Out the Window.
Brands treat last quarter's campaign data as insight into next quarter's customer. Baruch Fischhoff named the psychological mechanism that explains why this feels right — and why it almost always isn't.
Vilhelm Bjerknes was a Norwegian physicist who believed, in 1904, that weather could be predicted mathematically.
This was not a popular position. Meteorology in 1904 was almost entirely observational — forecasters looked at what was happening now and made educated guesses about what was likely to happen next, based on experience and pattern recognition rather than any formal model of how atmospheric systems actually behaved. The forecasters who were respected were the ones who had been watching the sky for decades. Their value was their accumulated observations. Their limitation was identical to their value: they could only tell you what they had already seen.
Bjerknes argued that this was structurally insufficient. Weather is not static. Atmospheric systems are moving, interacting, changing — and any forecast based solely on current conditions without a model of how those conditions evolve will always be a forecast of the present, not the future. You cannot predict tomorrow's weather by describing today's accurately enough.
He published a paper in 1904 outlining a set of mathematical equations that could, in principle, be used to calculate future atmospheric states from present ones — the first systematic attempt to turn meteorology from retrospective observation into genuine forecasting.
The insight was already complete before any computer existed to run it: observation tells you what has happened. A model tells you what is likely to happen next. These are not the same thing, and no amount of additional observation closes the gap between them.
The Mechanism That Makes Historical Data Feel Like Insight
Creeping determinism was named and documented in 1975 by Baruch Fischhoff — then a researcher in Daniel Kahneman's lab — in a paper published in the Journal of Experimental Psychology: Human Perception and Performance. It has since become one of the most cited findings in behavioural psychology.
Fischhoff's experimental design was simple. He gave participants descriptions of real historical events along with four possible outcomes. One group was told which outcome had actually occurred. The other group was not. Both groups were asked to estimate the probability of each outcome as though they did not yet know the result.
The group that knew the outcome rated it as substantially more probable than the group that did not. When people know how something turned out, they perceive the outcome as having been more predictable than it was. They look back at the evidence that preceded it and find the pattern that led to it — because they already know where to look.
Fischhoff called this the creeping sense that the outcome could not have been otherwise. In retrospect, we seem to perceive the logic of events which unfold themselves in a regular, linear fashion according to a recognisable pattern with an alleged inner necessity. The path from cause to effect appears logical and somewhat inevitable — because we are tracing it backwards from a destination we already know.
Now look at what happens in every campaign debrief.
The campaign performed well. The marketing team reviews the targeting, the creative, the channel selection, the timing. Everything looks coherent. Everything looks like it led, logically, to the outcome. The team concludes: we understood our customers.
What they have actually done is look at a map of a journey they already completed and concluded that they knew the route in advance.
The customer insight that feels most solid — the understanding that seems most reliable — is the understanding built from data about what already happened. And creeping determinism ensures that understanding will always feel more certain, more predictive, and more transferable to future decisions than it actually is.
What the Weather Forecaster Who Only Looks Out the Window Actually Sees
Most brand data infrastructure is observational rather than predictive.
A customer data platform collects what customers have done. An ESP tracks what they opened and clicked. A CRM records what they purchased and when. An analytics dashboard shows the channel and content and timing that produced the last conversion.
All of this is retrospective. All of it describes the customer as they were when the data was collected. None of it — without a model of how those conditions evolve — tells you anything reliable about who the customer is today or what they are likely to do next.
The weather forecaster who looks out the window and reports that it is currently raining is providing accurate, useful information. The problem begins when that observation is used as the primary input for tomorrow's forecast. Not because the observation is wrong — but because current conditions and future conditions are related but not identical, and the relationship between them requires a model, not just more observation.
Bjerknes' insight applies with complete precision to every brand communication programme treating last quarter's campaign performance as insight into what next quarter's customer will respond to.
The customer who bought during last year's Diwali campaign is not the same customer sitting in front of next year's Diwali campaign. Their context has changed. Their relationship with the category has changed. Their relationship with the brand has changed. Their competing demands have changed. None of that change appears in the retrospective data. Only the purchase does.
And creeping determinism ensures that the purchase looks, in hindsight, like the natural consequence of the strategy — rather than what it may also have been: a customer who was ready to buy regardless of the campaign, or a customer whose behaviour in that moment was driven by a contextual factor the data never captured.
The Diwali insight that drives the next three Diwali campaigns is looking out the window.
The Four Ways This Shows Up in the Meeting Room
"Last Year's Strategy Worked So We're Scaling It This Year"
The most common expression of creeping determinism in brand planning.
Last year's strategy worked. This is known because the outcome is known. Because the outcome is known, the strategy appears coherent and well-targeted in retrospect — the decisions look like they led logically to the result. The team is confident because they have seen the outcome.
What they cannot know, because the counterfactual was never run, is how much of last year's outcome was produced by the strategy and how much was produced by market conditions, competitive dynamics, or customer readiness that would have produced a similar outcome regardless of the specific tactical choices made.
Scaling a strategy that worked in a market condition that no longer exists is not optimising. It is forecasting tomorrow's weather from a photograph of last year's sky.
Bjerknes did not look at yesterday's weather and call it a forecast. He built a model of how atmospheric conditions evolve.
"Our Best Customers Look Like This — So We'll Target More People Who Look Like This"
Lookalike targeting based on historical purchase behaviour is the most widely deployed expression of observational data treated as predictive insight.
The customers who bought are known. Their characteristics are known. The logical inference appears to be: find more people who share those characteristics.
What creeping determinism obscures is that the customers who bought were ready to buy at that moment — and the model cannot distinguish between the characteristics that caused the purchase and the characteristics that merely happened to be associated with it. A lookalike audience built on last year's converters describes who converted, not who was ready to convert. These are not the same population.
The forecast assumes the weather patterns that produced last year's result will produce the same result in conditions that have already shifted.
"The Post-Purchase Survey Tells Us What Customers Want"
Post-purchase surveys are conducted after the outcome is known — by the respondent.
The customer who has just purchased is in a specific psychological state: they have made a decision and their cognitive system is actively working to reinforce that decision as correct. The survey data they produce reflects this state, not the actual decision process that preceded the purchase.
This is creeping determinism operating at the individual customer level. The customer who bought describes their motivations in a way that makes the purchase seem inevitable and well-reasoned — because they are tracing the path backwards from a destination they already reached.
The insight produced by asking customers why they bought is a description of how customers want to believe they buy. Not how they actually buy.
"Our Data Shows Our Best Channel Is Email — So We're Doubling Email"
The attribution model that identifies email as the best-performing channel is built entirely from historical conversion data. It knows which channel the customer interacted with before converting. It does not know whether the customer would have converted anyway through a different path, or whether the email interaction is a coincidence of timing rather than a cause of conversion.
Doubling the budget of the channel that appears in the most historical conversion paths is a logical application of observational data. It is also a strategy built entirely from data that describes what happened and almost nothing about what caused it.
Bjerknes' window, showing yesterday's clear sky, being used to justify leaving the umbrella at home.
Three Inputs That Shift From Observational to Predictive
Bjerknes solved meteorology not by collecting more observations but by building a model — a forward-looking framework that used current conditions as inputs rather than as conclusions.
The equivalent for brand communication is not more data. It is a model of how customer conditions evolve. As a translation of that principle — not a derivation from Bjerknes' equations, but an application of the same structural logic — three inputs shift a customer data framework from retrospective to predictive.
Input 1 — Behavioural velocity, not behavioural history
Not what the customer did, but how their behaviour is changing. A customer whose purchase frequency has been increasing for three months is a different signal from a customer whose purchase frequency has been flat for three months at the same absolute level. The velocity is the leading indicator. The absolute value is the lagging one.
Most analytics dashboards show absolute values. The model requires velocity.
Input 2 — Contextual signals, not demographic proxies
The contextual factors that predict purchase readiness — a life event, a seasonal trigger, a competitive displacement — are more predictive than demographic characteristics because they describe the customer's current state rather than their historical category membership.
A 35-year-old who just had a second child is not well-described by their age and gender. They are well-described by the life event that just changed their purchase priorities, channel preferences, and attention availability simultaneously.
Demographic data is a proxy for context. Context is the real input.
Input 3 — Decay-adjusted engagement, not raw engagement
An email open from three weeks ago is a different signal from an email open from yesterday. An engagement signal that has been declining for ninety days is a different prediction from one that has been stable for ninety days.
Most engagement scoring models treat all engagement history as equally weighted regardless of recency. A predictive model applies a decay function — the signal from yesterday is worth more than the signal from last month, which is worth more than the signal from last quarter.
The observation tells you that the customer engaged last year. The decay-adjusted model tells you whether they are likely to engage tomorrow.
The Forecast That Was Right
Vilhelm Bjerknes died in 1951. The first weather computer at Princeton ran its first atmospheric forecast in 1952 — one year after the man whose equations it was solving was gone.
But his framework had already changed meteorology permanently. Not by replacing observation — observation remained essential — but by establishing that observation is the input to a model, not the model itself. The observation tells you where the atmospheric system is. The model tells you where it is going.
Every brand treating its customer data as insight rather than as input to a model of how customer behaviour evolves is standing at Bjerknes' window in 1904.
The data is right. The customer who bought last Diwali is real. The channel that converted them is documented. The creative that worked is recorded.
None of it is a forecast.
The window tells you what is. The model tells you what comes next. Only one of them is a forecast.
If your brain is already triaging this page for a 5-second window, skip the reading—the complete narrative is perfectly laid out in the infographic below.
Published by Hetvabhas — independent analysis of brand communication
No vendor agenda. No sponsored content. No false reasoning.
