Synthetic Personalities: How Well Can LLMs Mimic Individual Respondents Using Socio-Economic Microdata?

Researchers demonstrate that large language models can effectively create detailed digital twins of individual consumers using existing socio-economic panel data, achieving 78.8% accuracy on held-out questions. The study maps construction decisions across model types, information depths, and embedding methods, showing that market research scalability is now limited by data volume and model selection rather than data collection design.

This research addresses a practical gap between theoretical LLM applications and real-world marketing implementation. Rather than building personas from scratch or relying on purpose-collected survey data, the authors leverage pre-existing CRM and panel data that companies already possess, making their findings immediately relevant to business operations. The systematic evaluation across 2.1 million synthetic responses reveals that detailed individual twins are achievable with open-weight models, democratizing access to this capability beyond proprietary solutions.

The study's key insight—that diminishing returns occur past 75% data entropy—creates genuine operational value. Marketing teams can achieve near-optimal performance (relative to using all available data) while reducing computational costs and privacy exposure. The shift from narrative persona summaries to raw dialog history consistently improved accuracy, suggesting that unstructured historical interactions encode behavioral patterns more effectively than curated summaries.

For the broader AI industry, these findings validate the twin-based market research paradigm as technically sound and economically viable. The 78.8% accuracy and 0.590 Fisher-z correlation demonstrate sufficient performance for actionable insights. This addresses a significant bottleneck: companies accumulate vast behavioral data through loyalty programs and CRM systems but lack efficient methods to extract predictive intelligence.

The work signals that LLM-driven personalization and behavioral prediction are moving from research prototypes to practical business tools. However, success depends on model selection, data preprocessing, and engineering choices rather than breakthrough algorithmic advances. Organizations should expect rapid commercialization of these techniques, with competitive advantage shifting toward those who optimize the construction parameters mapped in this research.

• →LLM digital twins achieve 78.8% accuracy on individual consumer prediction using existing panel data, enabling practical market research scaling.
• →Information depth shows diminishing returns past 75% entropy quartile, creating cost-efficient optimization points for real-world deployment.
• →Raw dialog history embedding outperforms narrative persona summaries across all model configurations, suggesting behavioral data encoding matters more than curation.
• →Open-weight LLMs prove sufficient for detailed individual-level twins, reducing dependence on proprietary models and enabling broader adoption.
• →Twin quality is now constrained by item volume and construction decisions rather than data design, shifting implementation focus to engineering optimization.