Prediction error: meaning and application

You have a product that is better than people think. The claim is on the pack and in the advertising, yet the perception barely shifts. Then someone tastes the product in the supermarket and says: "I didn't expect that." In neuroscience and behavioural science, that difference between expectation and perception is called a prediction error. Here you learn what it means, where it comes from, how expectations form and change, and what it can and cannot do for marketing and communication.

A prediction error is the difference between what a learning system expects and what it actually perceives or receives.

In learning models, that error sets how much an expectation is revised: the bigger the surprise, the more there is to learn. Predictive coding uses the concept as a building block of perception. For marketing it is a useful lens, but a surprise does not automatically revise an expectation, and a revised expectation does not automatically lead to behaviour or a purchase.

Learning theory: Rescorla and Wagner

The idea that surprise drives learning comes from research on classical conditioning. In 1972, Robert Rescorla and Allan Wagner published an influential model of this kind of learning (Rescorla & Wagner, 1972). In it, a learned expectation changes in proportion to the difference between the actual and the predicted outcome. Niv and Schoenbaum summarise the core as: make a prediction, observe what happens and update your knowledge if the prediction was wrong (Niv & Schoenbaum, 2008). The model also explained blocking: if a light already predicts food, an animal learns hardly anything about a sound added later, because the food is no longer unexpected (Schultz, Dayan & Montague, 1997).

Dopamine and the reward prediction error

Schultz, Dayan and Montague linked this principle to brain recordings. Dopamine neurons in monkeys fired briefly more strongly after an unexpected reward. After learning, that response shifted to the signal that predicted the reward. If an expected reward was omitted, activity dropped when the reward should have arrived (Schultz et al., 1997). This pattern resembles the prediction error in temporal difference learning, a computational model that updates predictions over time. Niv and Schoenbaum stress that this is a simplified and almost certainly incomplete model of dopamine (Niv & Schoenbaum, 2008). These were animal studies, not consumers in a shop.

Predictive coding as an overarching framework

Predictive coding applies the idea to perception. In the model by Rao and Ballard, higher visual brain areas send predictions to lower areas, which send back only the unpredicted remainder: the prediction error (Rao & Ballard, 1999). It was a computational model of the visual cortex. Friston developed the idea into a theory in which perception and learning both come down to minimising 'free energy', a statistical quantity (Friston, 2005). He later described free energy as an upper bound on surprise and named action as another way of reducing prediction errors (Friston, 2010). Philosopher Andy Clark popularised the broader framework as predictive processing and saw it as the best clue yet to a unified science of mind and action (Clark, 2013).

Neurofactor uses two working concepts to describe this in plain language. They are editorial working definitions, not established scientific terms.

• Encoding: repeated experience lays down an expectation. Taste a disappointing product a few times and you expect the next taste to disappoint too.
• Re-encoding: a new experience that clearly departs from that expectation and is credible can revise the expectation.

The Rescorla-Wagner model shows why repetition makes an expectation stable: the better the prediction fits, the smaller the error and the less changes with each experience (Niv & Schoenbaum, 2008). Change requires a deviation.

Why 'credible'? In predictive coding models, a prediction error carries more weight the more precise, and therefore reliable, the signal is estimated to be (Friston, 2010). Applying this to credible brand experiences is an interpretation of that model idea, not a measured law.

Memory research offers a related, cautious pointer. A retrieved memory can become temporarily changeable again; this is called reconsolidation. Sevenster, Beckers and Kindt found in humans that a prediction error during retrieval was a necessary condition for reconsolidation of a learned fear memory (Sevenster, Beckers & Kindt, 2013). This was a laboratory study of learned fear in which memory was influenced with a drug. It does not show that brand associations are rewritten in the same way, but it fits the idea that repeating what someone already expects changes little.

Predictive processing is influential but not settled. Clark wrote in 2013 that direct neuroscientific testing of hierarchical predictive coding was still in its infancy and that the evidence was mainly indirect (Clark, 2013). In 2020, Walsh and colleagues stated that predictive processing models had so far lacked the empirical support to justify their status, and assessed new neurophysiological studies aimed at that gap (Walsh et al., 2020). Work on the mismatch negativity, an EEG response to a deviation in a stimulus sequence, also weighs other explanations, such as neuronal adaptation to repeated stimuli (Garrido et al., 2009).

So use prediction error as a well-supported learning principle, and predictive coding as an influential but contested framework.

People often picture a category before ever using the product. The concept sharpens three points.

1. A claim is itself a prediction. "Full of flavour" on a pack is the sender's claim, easily dismissed as advertising by someone expecting "bland". A personal experience provides direct perception. That concrete, first-hand evidence can revise an expectation more strongly than a claim is plausible reasoning, not a general law; test it per product and situation.
2. Expectation colours experience. In an fMRI study with twenty participants, the same wine was rated more pleasant at a higher stated price, alongside more activity in the medial orbitofrontal cortex; primary taste areas showed no difference (Plassmann et al., 2008). The authors note possible effects of the research situation. A tasting is therefore not a neutral test: a strong expectation can pull the experience its way.
3. Surprise is not a goal in itself. An unexpected experience can also be negative, dismissed as an exception or quickly forgotten.

What neuroscience can and cannot do in marketing is explained in What is neuromarketing?.

1. Name the expectation concretely. Not "negative image" but "tastes bland" or "hard to install". Research the target group's associations with the category; see How do you measure associations?.
2. Establish whether the expectation is accurate. If an expectation is justified, the product needs to change, not the communication.
3. Choose an experience that addresses the expectation directly. One that targets a different aspect leaves the original expectation intact.
4. Define measures in advance: expectation before the experience, judgement straight after, expectation some time later.
5. Compare with a suitable control, such as a group that only sees the claim or tastes the product blind.
6. Draw conclusions about behaviour only from behavioural data. A revised expectation is an intermediate step, not a conversion.

This example is fictional. A plant-based dairy brand sees that people who never tried its products often link the category with "bland" and "watery". The brand is weighing a "full of flavour" campaign against supermarket tastings.

• Decision question: which approach revises the expectation "tastes bland" most, and does that last?
• Available information: associations from earlier qualitative research; no data yet on the effect of tasting.
• Suitable approach: three conditions among non-users: seeing only the claim, tasting blind and tasting with the brand visible. Each condition measures taste expectation beforehand, taste judgement directly afterwards and taste expectation again after three weeks. Outcome measures and analysis are fixed in advance. There are no results; this is a research plan.
• Possible interpretation: if the judgement rises with blind tasting but not with branded tasting, the existing expectation may be colouring the experience. If the expectation is unchanged after three weeks, there was a surprise, but no lasting revision.
• Next step: adjust the tasting experience or improve the product first, then test a larger roll-out on purchasing behaviour.

• Seeing prediction error as a 'buy button'. No brain signal forces a purchase. The concept describes learning, not a switch.
• Thinking you measure prediction error directly. A prediction error is usually inferred from a model and a task in which expectation and outcome are controlled. Without such a design, a brain measurement during an advertisement does not yield a measure of 'the' prediction error. What EEG does record is explained in What is EEG?.
• Trying to explain everything with predictive coding. A framework that fits any result after the fact makes no testable prediction.

Prediction error is a well-supported learning principle: expectations change mainly when reality departs from them. Predictive coding places that principle in a broader, influential but contested framework. For marketing and communication: name the expectation you want to revise, offer a concrete, credible experience that addresses it directly, and test whether the revision lasts and behaviour changes.