Mapping the Deep Structure of Human Preference

The entertainment industry runs on taste. Hits and misses are determined by whether a property's emotional and cognitive offer matches what specific audiences want at a specific moment. The challenge: taste is not directly observable. You can only observe its outcomes — box-office receipts, stream counts, cancellation decisions — after the money is spent. This paper sets out the scientific case for a different approach: measuring the deep structure of human preference directly, before capital is committed, using the personality psychology that governs why people like what they like.

1. The taste problem

Entertainment investment is a fundamentally different problem from most capital allocation decisions. When a pharmaceutical company funds a drug trial, the target is biochemically defined. When an infrastructure fund finances a toll road, demand is modelled from traffic patterns and population data. Both involve uncertainty, but the uncertainty is about external factors — market growth, regulatory risk, execution quality.

Entertainment investment has a different uncertainty at its core: taste. Success depends on whether a large enough group of people will find a specific emotional and cognitive experience sufficiently compelling to pay for it — not once, in a test environment, but repeatedly, voluntarily, and competitively against every other option available to them that week.

Taste is not random. It has structure. People's preferences cluster in predictable patterns that persist across time, across domains, and across formats. A person who seeks out morally complex narratives in literature will tend to seek them in television. A person who values emotional warmth and resolution in film tends to value it in the books they read. These patterns are not coincidences of demographic overlap. They are expressions of stable underlying personality traits — traits that are measurable, mappable, and directly predictive of content preference.

The scientific literature on this relationship is now extensive. Yet the entertainment industry's primary measurement tools — box-office history, genre categorisation, demographic surveys, focus groups, collaborative filtering — treat taste as an emergent aggregate rather than a mappable structure. They measure taste's outcomes rather than its causes. This produces a systematic vulnerability: the tools work adequately when the historical data is abundant and the product is similar to previous products. They fail precisely when the stakes are highest.

The further you get from behavioural data, the wider our advantage.

2. Where current methods fail

2.1 The behavioural-data trap

Collaborative filtering — the recommendation engine architecture underlying most streaming platforms — works by finding users who behave similarly and assuming their future preferences will also converge. It is effective within a stable, well-populated catalog. Its failure mode is structural: it requires behavioural history to make predictions. When that history doesn't exist — new IP, new audience, new platform — it has nothing to work from.

This is not a data-quality problem. It is a category error. Behavioural data tells you what people have done. It cannot tell you what they will do when the product is new, the platform is new, or the audience profile has shifted. For the decisions that require the most capital commitment — greenlighting an original series, acquiring a catalog, licensing a gaming IP for screen adaptation — behavioural history is structurally inadequate.

2.2 The library-merger problem

When two large content catalogs merge, the recommendation and audience-modelling systems face a cold-start problem at scale. The engagement history of Catalog A's subscribers tells you nothing reliable about their response to Catalog B's content, because the behavioural signals were generated in a different product context, against a different competitive set, for a different price point. The merger that looked like a data asset turns out to be a data liability.

Psychographic models do not have this problem. Personality traits are stable across contexts. A subscriber's preference structure — their appetite for moral complexity, their threshold for emotional intensity, their response to ensemble-driven storytelling — does not change when the catalog around them changes. A personality-grounded model transfers across catalog mergers without degradation.

2.3 The gaming-to-screen translation failure

The adaptation of gaming IP to film and television has become the entertainment industry's most expensive recurring problem. The conventional explanation — “gaming audiences don't translate to theatrical audiences” — is accurate but unhelpful. The cause is a systematic confusion between platform audience and IP affinity. Completion rates, in-game purchases, and time-in-game describe the gaming audience's relationship with the game, not with the IP's underlying narrative and emotional features.

A game succeeds through agency, mastery, and reward loops. A film succeeds through catharsis, character arc, and emotional resolution. The audience that responds to an IP's story in game form may respond to the same IP in film — but not through the same emotional pathway, and not in the same demographic slice. Psychographic modelling addresses this directly, by scoring the IP's content features independently of format and mapping them against personality-driven audience segments for the target format.

2.4 The demographic-proxy problem

Age and gender remain the primary segmentation axes for most entertainment audience research. They persist because they are measurable, legally defensible, and historically predictive at the population level. They fail at the IP level because the variance within demographic groups vastly exceeds the variance between them for most content decisions. A 28-to-35 female demographic contains people who prefer psychologically complex prestige drama, light romantic comedy, crime procedurals, and supernatural fantasy. Their demographic membership is identical. Their taste profiles are not.

2.5 The tracking era is ending

The critique of behavioural-data tools is not only methodological. It is structural — and it is now being enforced by regulatory and platform-level decisions the industry has spent a decade trying to avoid. GDPR (2016). Cambridge Analytica (2018). CCPA (2020). Apple App Tracking Transparency (2021) — cutting 55% of mobile traffic tracking and costing Meta alone an estimated $10 billion in lost annual revenue. In 2024, Google formally admitted defeat on cookie deprecation.

This is not a temporary headwind. The behavioural-data infrastructure — tracking pixels, third-party cookies, identity graphs, cross-site retargeting — is being dismantled by regulatory pressure, platform policy, and changing consumer expectations simultaneously. The question for every company that runs its audience intelligence on behavioural data is not whether this affects them. It is what they replace it with.

Psychographic modelling is, structurally, the answer. Personality traits are inferred from consented first-party signals. They are stable — unlike behavioural patterns, they do not degrade when cookies are blocked or device IDs are masked. They are portable across platforms and catalogs. As PII regulation tightens, psychographic prediction becomes more competitive, not less. The privacy tailwind that is destroying behavioural-data businesses is a direct advantage for trait-science infrastructure.

Stop studying ripples. Study the rock.

3. The scientific foundation

The case for psychographic modelling rests on a body of evidence in personality psychology that is now sufficiently large, sufficiently replicated, and sufficiently well-validated to constitute a reliable scientific foundation for commercial application. Since the 1930s, lexical and factor-analytic research has independently re-derived the same five stable trait dimensions across languages, cultures, and decades — arriving at the same Big Five from entirely separate methodological starting points.

Personality predicts content preference. Rentfrow and Gosling's 2003 study in the Journal of Personality and Social Psychology established the foundational mapping: Big Five traits reliably predict music preferences, with predictive accuracy ranging from 41% to 66% against independent behavioural observation. Schäfer and Mehlhorn's 2017 meta-analysis confirmed the pattern across genres. Nave et al. (2018, Psychological Science) extended it to active listening behaviour and Facebook Likes. The relationship is bidirectional and robust at scale.

Personality is inferable from behaviour at scale. Kosinski, Stillwell, and Graepel (2013, PNAS) demonstrated that digital behavioural traces reliably infer Big Five personality at scale; the underlying myPersonality dataset eventually grew past six million participants. Liu and Campbell's 2017 meta-analysis (113 samples, N=53,913) confirmed the pattern across platforms. The behavioural data streaming and social platforms already collect is not just a record of past preferences — it is a latent personality signal.

The replication question. The personality-behaviour literature has fared substantially better than many other areas of psychology under replication scrutiny — and the reason is structural. The LOOPR project, which conducted preregistered replications of personality-outcome associations, found that 87% replicated in the expected direction. Effect sizes are generally small to medium (r = .10–.30); the appropriate framing is probabilistic population-level prediction, not individual determinism. At commercial scale, that is exactly the framing that produces edge.

The five classical dimensions

For entertainment content prediction, each dimension maps to a distinct preference signal:

4. The Tiresias framework

The Tiresias Audience Simulation Engine (TASE) operationalises this scientific foundation as a five-stage simulation pipeline, converting IP feature profiles and personality segment data into affinity scores, audience maps, and commercial recommendations. Tiresias's proprietary technology extends beyond the classical Big Five with additional, scientifically validated traits — and proprietary machine-learning models custom-trained on consented user data to map preferences across entertainment domains.

Alongside this, we are developing a proprietary sub-trait classification under active internal validation — a layer beneath the Big Five that captures preference variance the classical model does not address. Findings will be disclosed in subsequent research notes once validation is complete.

4.1 Content feature scoring

Every IP is scored across thousands of content-feature dimensions on a 0–1 scale: narrative complexity, emotional intensity, violence intensity, visual spectacle, warmth, moral complexity, thematic novelty, humour, romance, character depth, world-building, horror tension, pacing intensity. The scoring is format-agnostic. The same source material — a novel, a graphic novel, a game — receives the same feature scores regardless of which adaptation format is under consideration. By separating the IP's content identity from its format identity, we predict audience response to an adaptation before the adaptation exists.

4.2 Psychographic segmentation — TriggerMap™

Audience segments are derived via TriggerMap™, the proprietary T-Score segmentation system. Boundaries overlap; segment sizes are non-equal — accurately reflecting the real structure of personality distributions. Segments are defined by their T-Score profiles, content preferences, and predicted behaviour: seek-out, word-of-mouth, wait-for-recommendation, or skip. They are not demographic proxies. A “Cultural Seeker” segment may span ages 24 to 58. An “Immersive Escapist” may be evenly distributed across gender. The segmentation reflects the structure of preference, not the convenience of external characteristics.

4.3 Affinity simulation

For each segment, the simulation computes a 0–1 affinity score — the probability-weighted likelihood of positive reception. Overall affinity is the population-weighted average across all segments, with a 95% confidence interval derived from segment variance. The interval is a substantive data point, not a disclaimer: a wide interval indicates genuine audience polarisation, which is a different commercial picture from a property with a narrow interval around the same mean.

4.4 Text-to-personality extraction (Augur)

Tiresias's Augur LLM was custom-built for our methods. It infers Big-Five personality-relevant features directly from user-generated text. In testing against ground-truth personality data, Augur explains roughly 30% of the variance in user personality (R² ≈ 0.30, Spearman r > 0.50) — a coefficient that places it within the upper range of validated psychometric instruments and at or above the typical agreement between trained human raters on Big-Five scoring.

Why that figure matters commercially: at population scale the unit of decision is a segment, not an individual. Even moderate per-person inference accuracy aggregates into decisive segment-level signal, because individual error averages out across thousands of inferred profiles. This is why Augur enables cold-start analysis on genuinely new IP — new novels, unpublished scripts, graphic-novel runs, game narratives — without behavioural history, box-office comps, or prior adaptation.

4.5 Benchmark performance

Performance is measured against two baselines. The primary commercial benchmark is industry-standard collaborative filtering — the architecture underlying most major streaming platforms. Tiresias outperforms it by +39% for film, +57% for TV, and +193% for books in top-K precision against held-out preference declarations (N=6,000 U.S. representative sample). Music shows +287% lift — reflecting the particular weakness of behavioural models in a format where personality is the primary predictor. The secondary benchmark is MBTI-based segmentation; our agent-swarm simulation architecture beats it by 30%+ on internal validation runs.

For external scale: in published research, Netflix's own personalised recommender produces a 12% engagement uplift over a non-personalised popularity baseline (a top-10 list). [Aridor, Zielnicki, Bibaut et al., 2025. The Value of Personalized Recommendations: Evidence from Netflix.] That is the industry reference for what state-of-the-art collaborative filtering buys in production. Tiresias's lift over the same family of methods — multiples larger across film, TV, and books — is the gap psychographic modelling opens between describing past behaviour and predicting future preference.

5. Three problems, one framework

5.1 New IP with no behavioural history

When a studio is evaluating rights to a novel published eighteen months ago, or a game from an independent studio with no theatrical precedent, collaborative filtering and box-office-history approaches return the same answer: insufficient data. Psychographic simulation has no such dependency. Augur extracts the IP's personality-relevant features from the text or narrative directly. TASE maps those features against the target audience's personality segments. The affinity score and segment breakdown are available within days, before any audience has encountered the property in its adaptation form.

5.2 Cross-domain adaptation

The gaming-to-screen problem is the most commercially consequential version of this. TASE scores the IP's content features independently of format, then simulates affinity against the personality segment distribution relevant to the target format. A gaming IP's theatrical audience simulation uses the theatrical audience's personality segment model, not the gaming audience's behavioural data. The output is a prediction of who will show up to the film — and what the film needs to deliver to retain them — rather than a description of who plays the game. The same logic runs in reverse for film-to-game and any other cross-domain extension.

5.3 Catalog M&A and post-merger integration

Catalog acquisitions are valued, in practice, on two assumptions: that the target's historical engagement reflects what the acquirer's subscribers will engage with, and that the recommendation systems of the combined entity will serve both libraries with equal fidelity from day one. Both assumptions fail under the structural problems described in §2.2. The behavioural signals that drove the target catalog's historical engagement were generated inside a different product, against a different competitive set, at a different price point. They do not transfer. The recommendation system that inherits them does not perform.

Psychographic modelling addresses the diligence and integration problems directly, because personality traits are stable across the conditions that break behavioural data. Subscribers' psychographic profiles — derived from their engagement with Catalog A — apply directly to predicting their affinity for Catalog B's content. The profile transfers; the behavioural context does not need to.

Three concrete applications follow:

• Pre-deal audience-affinity diligenceMap the acquirer's subscriber base against the target catalog before close. Identify which segments will engage, at what affinity, and against which titles in the target library. The output is a defensible, segment-by-segment view of subscriber retention and incremental engagement risk — priced into the bid rather than discovered after it.
• Post-merger programming strategyIdentify the content gaps and commissioning priorities that serve the combined audience's psychographic profile, rather than re-running either predecessor's historical programming pattern. The merged library is a new product; programming should be planned for the new audience, not the legacy one.
• Cross-format and library-extension diligenceScore acquirable IP — whether a catalog of novels, a games portfolio, or a music library — against the personality segment distribution of the target adaptation format. The same framework that resolves greenlight diligence on individual properties resolves rights-portfolio diligence at scale.

For diligence rooms, the speed advantage in §6 is not an abstract feature. It is the difference between an audience-affinity view that arrives within the negotiation window and one that arrives after the deal closes.

6. The speed imperative

The traditional content development timeline — option, develop, test, greenlight, produce, release — is measured in years. Streaming has compressed the release cycle at the back end while increasing the pressure at the front end: more content decisions, faster, with higher consequences for wrong calls.

Traditional audience research tools have not kept pace. A theatrical test screening program takes months. A focus group study for a major IP takes a month. A collaborative-filtering engine requires months of engagement data to produce reliable signal. You can have all of them in a month. Feature scoring, segment simulation, affinity modelling, and commercial recommendation — the full pipeline from IP input to actionable output — is a computational process, not a data-accumulation process. It does not wait for audience engagement to accumulate. It derives predictions from the stable psychological architecture of preference, which is already known.

In rights negotiation contexts — where an IP is available for a limited window, where competing bids are being evaluated, where a library acquisition is time-sensitive — the ability to generate a rigorous audience-affinity analysis before the window closes is the difference between an informed decision and a gut call.

De-risk content transactions with audience intelligence built on proprietary psychometric models — before a dollar is spent.

7. What this changes

Psychographic modelling does not replace creative judgment. A simulation cannot tell a writer's room what to write, or tell a director how to shoot a scene, or tell a marketing team what the campaign line should be. These are human decisions, and the quality of their execution is the primary determinant of whether a well-positioned property succeeds or fails.

What psychographic modelling changes is the quality of the decision framework within which those judgments are made. It answers four questions that conventional methods address poorly:

• Who is the core audience for this property, defined at the level of personality-driven preference rather than demographic proxy?
• How large is that audience, and how strongly does the property's feature profile match what they want?
• What are the specific features that drive their affinity, and which features create risk?
• Which adaptation decisions — format, sequencing, tone, campaign positioning — are most consequential for audience alignment?

These are the questions that determine whether a content investment is priced correctly, positioned correctly, and developed correctly. They are answerable with psychographic data. They are not answerable with box-office history, genre categorisation, or demographic surveys alone.

The entertainment industry makes billion-dollar decisions about what people will like before they have had a chance to like it. The scientific tools to make those decisions with dramatically more precision than current methods allow have existed in the academic literature for decades. The gap has been in operationalising them at the speed and scale that commercial decision-making requires.

That gap is now closed. And the timing is not coincidental.

The regulatory and platform-level dismantling of behavioural-tracking infrastructure has created a forced migration. Every company that built its audience intelligence on third-party cookies, cross-device identity graphs, and behavioural retargeting is now in the market for something to replace it. The replacement has to work without PII. It has to be privacy-compliant by design, not by exception. It has to be stable across platform changes and regulatory motions. And it has to predict preference rather than merely describe historical behaviour.

Trait science is that replacement. It was the right answer before the tracking era ended. It is the only durable answer now that the tracking era is ending. The infrastructure window for building the new substrate is open. It will not stay open indefinitely.