Modelling Human Action

Behavioural simulation models what a particular actor may do next under a defined set of circumstances. It combines evidence about the actor's past behaviour with the information they could observe, the choices available to them, and the state of the surrounding system. The aim is not to produce a persuasive imitation of a person. It is to estimate a distribution over plausible actions—including no action—and make the evidence and uncertainty auditable.

What behavioural simulation means

Behavioural simulation—spelled behavioral simulation in US English—is the computational study of how an actor responds to a changing environment. The actor might be a trader reacting to an earnings surprise, a company responding to a competitor, a regulator addressing a new risk, or an online community member deciding whether to share a claim.

The prediction target should be observable: post, reply, buy, sell, announce, delay, abstain, escalate, or choose another well-defined action. This makes the system testable. A description such as “the actor becomes more cautious” may be useful context, but it is not an outcome until caution changes something that can be observed.

Behavioural simulation operates at a different level from our guide to social simulation. An actor-level model asks what one entity may do given its history and context. A social simulation connects many such actors so their actions alter one another's environment. The two disciplines complement each other, but weak actor models do not become reliable merely because hundreds of them interact.

Behaviour is not a persona

Language models make it easy to generate a detailed persona: biography, values, occupation, temperament, and writing style. That profile can look psychologically complete while containing traits that were never observed. When the same model invents the persona and predicts the action, stereotypes can become self-confirming evidence.

Research increasingly supports a more disciplined approach. A 2026 WASSA study on operational validity in simulated social-media users found that authentic behavioural histories were more useful than generated biographies for next-action prediction. Its form–content experiments also showed that linguistic similarity and behavioural fidelity are separable. A model can sound like somebody without choosing as they do.

This distinction changes the product question. Instead of asking “does the agent feel real?”, ask “did the system recover the right action set, assign probability to the observed outcome, and remain calibrated on actors and events it did not see during development?” Style can be evaluated later, after the behavioural decision has been made.

The evidence hierarchy

The strongest signal about future behaviour is usually relevant past behaviour. The hierarchy below is a practical starting point rather than a universal law.

1. Authentic behavioural traces. Use time-stamped actions, choices, and responses made by the actor in comparable settings. Preserve the original source and distinguish an explicit action from an interpretation of one.
2. Current observable context. Reconstruct what the actor could have known at the prediction boundary. Later events, revised records, and hindsight summaries must not leak into the prompt or features.
3. Stable public attributes. Role, institutional constraints, mandate, or stated policy can narrow the action space when they are documented and relevant.
4. Inferred traits and generated summaries. These can compress long histories, but they should be labelled as inference and tested with ablations. If removing the biography improves accuracy, the biography was decoration or noise.

Rich self-reports can also be useful when collected carefully. Park and colleagues created agents for 1,052 participants from interviews and surveys, then compared simulated responses with held-out answers and each participant's own test–retest consistency. Their general-purpose simulation study is important because it treats human reliability as a ceiling and uses withheld measurements. The lesson is not that any long interview guarantees fidelity; it is that grounding and evaluation must be designed together.

Exposure, uncertainty, and no action

An actor cannot respond to information they did not receive. Yet exposure is often the least observed variable in social and market data. Seeing a public post does not prove that a particular person saw it. Following an account makes exposure more plausible, but still does not guarantee attention. A robust system records exposure as observed, proxied, or unknown rather than quietly treating all public information as universally seen.

This matters most when the outcome is silence. Non-action is a real behavioural outcome, but its cause is ambiguous. The actor may have ignored the event, missed it, considered it irrelevant, waited strategically, or acted through an unobserved channel. A model should include no action in the candidate set without claiming more certainty about its cause than the evidence permits.

Activity bias produces the opposite error. If every simulated actor is asked “how do you respond?”, most agents will manufacture a response. Real systems contain far more ignored stimuli than public reactions. Calibrating the action rate—and testing it against observed base rates—is essential before analysing the content of the generated actions.

Separate discovery, scoring, and rendering

Open-ended generation is good at proposing plausible actions but poor at expressing comparable probabilities. Fixed classification is easier to score but can miss an action that was never put on the list. A useful architecture separates three jobs.

1. Discover a high-recall action set

Retrieve actions from comparable historical moments, domain rules, and bounded model generation. Include no action and an “other” route. Candidate recall should be measured independently: if the observed action is absent, even a perfect ranker cannot recover it.

2. Score the same options

Give every candidate the same actor history and timestamp-safe context, then estimate relative likelihood. Keep outcome probability separate from model confidence. A system may assign 60% probability to an action while having low confidence because the actor's exposure is unknown or little comparable history exists.

3. Render only after deciding

Once an action is selected, a separate stage may express it in the appropriate channel and style. This prevents a model's ability to write an eloquent post from influencing whether the actor was likely to post at all. It also allows behavioural accuracy and linguistic resemblance to be tested as distinct properties.

How much history should an agent remember?

More context is not automatically better. Long unfiltered histories increase cost, introduce stale evidence, and can dilute the few precedents that matter. The FineRob benchmark contains 78,600 longitudinal records from 1,866 users and reports that expanding memory can reduce next-behaviour prediction quality. Its results favour organised retrieval and reasoning over indiscriminate accumulation. See the FineRob paper for the benchmark design.

Retrieval should therefore be conditional on the decision. Recency, situation similarity, relationship to the other actor, channel, and prior response pattern may all matter. The retrieved evidence should remain inspectable so a researcher can see whether the system relied on a genuine precedent or an irrelevant but semantically similar event.

Summaries can compress history, but they should not replace the underlying ledger. A claim such as “usually responds defensively to criticism” needs linked examples, counterexamples, and a window. Otherwise yesterday's model-generated interpretation becomes tomorrow's supposed fact.

How behavioural simulations should be evaluated

Randomly splitting rows is rarely enough. The same actor, event, or conversational thread can leak across training and test sets, making memorisation look like generalisation. Evaluation should hold out what the deployment must actually handle.

• Actor-held-out tests measure transfer to entities without a tuned history.
• Time-held-out tests preserve causality and reveal drift across regimes.
• Event-held-out tests test whether the model can handle a new type of stimulus instead of recalling a near-duplicate episode.
• Simple baselines—majority action, actor base rate, recency, nearest precedent, and a conventional classifier—show whether the complex system adds value.

Accuracy alone hides rare actions and inflated certainty. Report macro F1 or per-class recall for imbalanced choices; top-k recall for candidate discovery; log loss or Brier score for probability quality; and expected calibration error or reliability plots for confidence. Review errors by actor, action, exposure quality, history length, and event type.

The BehaviorChain benchmark, containing 15,846 behaviours from 1,001 personas, found that strong language models still struggle with continuous behaviour simulation. Work on simulating social-media engagement similarly found that zero-shot language models underperformed a supervised BERT baseline on action prediction, even when language models were useful for semantically aligned responses. These are healthy reminders that generative sophistication does not remove the need for baselines.

Calibration turns scores into usable probabilities

A ranker can put the observed action first and still be dangerously overconfident. Calibration asks whether outcomes assigned roughly 70% probability occur roughly seven times in ten across comparable cases. This matters operationally: a research team may monitor a 20% tail risk even when another action ranks first, while a supposed 95% result should be rare to contradict.

Calibration should be fitted on held-out development data and tested on a later untouched set. It should also be inspected within meaningful groups. A model may look calibrated overall while overstating certainty for new actors, rare actions, low-exposure cases, or periods of market stress. Where data are sparse, probabilities should move toward defensible base rates and the system should expose the reason for its uncertainty.

Version changes require re-evaluation. A new language model, retrieval policy, candidate generator, or context template can change the distribution even when top-line accuracy appears stable. Store the model and prompt version with every forecast, preserve the full candidate set, and compare calibration before promoting a new system. This creates a record that can be audited after the real action becomes known.

Where behavioural simulation is useful

Behavioural simulation is most useful for bounded, decision-relevant questions: which actors are likely to respond to a catalyst, which action types should be monitored, what context changes the ranking, where evidence is missing, and which scenarios deserve a human analyst's attention. It can support crisis rehearsal, organisational planning, policy analysis, market research, and the actor models inside larger social simulations.

It should not be used to claim privileged access to a person's private intention or to automate consequential decisions about people without appropriate consent, oversight, and recourse. Public behavioural traces are still personal data; simulation outputs can reproduce bias and create false confidence. Report assumptions, data provenance, uncertainty, and known blind spots with the result.

In market intelligence, behavioural models work best alongside live measures of attention and context. Our guides to stock sentiment, crypto mindshare, and social-sentiment research explain the evidence layer that a simulation must not invent. Browse the research library for the wider methodology.

Sources and further reading

• Schwager et al.: operational validity in simulated social-media users
• Park et al.: LLM agents grounded in self-reports
• Qu et al.: FineRob longitudinal behaviour benchmark
• Li et al.: BehaviorChain continuous behaviour benchmark
• Törnberg et al.: LLM simulation of social-media engagement