Define a shock. Generate agent profiles. Watch the simulation. Read the thesis.

A geopolitical event. A commodity disruption. A policy shift.

Macro models explain the trend. What they don't explain is the shape, why certain sectors overshoot, others lag, some reverse without apparent cause.

That residual looks stochastic. It isn't necessarily.

That irregular response is the aggregated effect of many different investors, each with their own biases, mandates, time horizons, and conviction thresholds, all interpreting the same event differently.

Individually rational. Collectively irrational. This is emergent market behaviour: outcomes that arise from the interaction of agents, not from any single agent's decision. EMDOLA models this across 4 retail and 4 institutional investor types, each processing information differently, interacting over time, across many simulated iterations.

Agent-based modelling (ABM) is an established academic simulation framework; it predates what is currently understood as 'AI agents' entirely ("agent" here refers to a simulation construct, not an LLM-driven autonomous system). It was designed to study emergent behaviour of interacting individuals.

EMDOLA takes this framework and does something new: using LLMs to populate it with genuinely heterogeneous, behaviourally coherent investor types at a scale and fidelity that wasn't previously possible.

The key contribution of the LLM is not generating personas, it's generating internally consistent profiles across 21 behavioural dimensions simultaneously. Conviction threshold, herding sensitivity, recency bias, FOMO, memory window, and more, calibrated across 3 archetypes per type, in combinations that reflect how real investor types actually behave.

A parameter sweep cannot produce this. The LLM is used upfront only, once profiles are generated, the simulation runs entirely without it.

Each run draws a different population composition, some dominated by momentum traders, some by fundamentalists, some by passive institutional flows. This produces a dataset across the full space of plausible market compositions.

No single run needs to be right. What matters is what emerges across the distribution, the consistent signal beneath the variation. The distribution is the finding.

The output is a systematic mispricing signal pointing to where aggregated investor responses push prices independent of macroeconomic forces. This means pricing pressure, not necessarily price direction. They can differ, and both can be right, and together paint a fuller picture of market dynamics. Not a price prediction, a novel investment thesis to investigate.