What a financial chatbot could learn from the questions traders ask

Could a chatbot learn more about an asset by listening to investors than an exchange learns from their trades? Wolfgang Kuhle’s theoretical preprint, Information Aggregation in Markets with Analysts, Experts, and Chatbots, explores that possibility. Its central mechanism is not superior stock-picking intelligence: it is direct access to information scattered across users’ questions.

The paper uses no data and does not test an existing chatbot. Instead, it constructs three mathematical economies: a market with privately informed traders, one with an analyst report, and one with a dominant chatbot. The comparison identifies conditions under which a platform can know more than market prices reveal.

Why prices leave something out

In the starting model, each trader receives a noisy clue about an asset’s underlying return, called its fundamental. Traders combine that clue with information inferred from the price, then choose how much to buy or sell. They are risk-averse and update their beliefs using Bayesian reasoning.

Prices bring these decisions together, but they also reflect an unpredictable supply shock representing noise-trader activity. A price movement therefore does not reveal whether traders have learned something about the return or whether the supply shock has changed. The equilibrium price remains an imperfect guide to the fundamental.

Kuhle derives prices that depend linearly on the fundamental and the relevant disturbances, checking that traders’ resulting demands clear the market. These are equilibrium calculations, not forecasts fitted to observed prices.

One report, many private readings

The second economy introduces an analyst report with two distinct sources of uncertainty. The report contains an error shared by all readers. Each trader also interprets it with an individual reading error.

Crucially, traders do not directly observe the report’s underlying meaning. They observe their own imperfect reading. Across the model’s population, individual reading errors average away, but the shared report error does not. That common error enters the market price alongside the noise-trader shock, and traders cannot simply remove it using their private readings.

This does not mean adding an analyst report necessarily makes prices worse than having no report. The paper finds that prices become more informative than in the no-report benchmark. The limitation is that shared report error remains an additional source of uncertainty. If everyone interpreted the report identically, traders could instead filter its contribution out of the price.

Learning from the incoming questions

The chatbot changes the route by which information travels. Before trading, everyone submits a question based on their private clue. The baseline assumes the chatbot knows how to recover that clue from the question.

The model contains a continuum of traders with independent, zero-mean private errors. Averaging their recovered clues therefore cancels those errors exactly, allowing the chatbot to infer the fundamental without uncertainty. Unlike a market price, this information channel is not contaminated by the noise-trader supply shock.

A further calculation relaxes the requirement that every question reveal an individual clue. Suppose traders choose between just two questions depending on whether their clue falls above or below a known threshold. With the assumed normal signal distribution, the exact proportion choosing either question still identifies the fundamental. Even coarse questions can carry aggregate information in this idealized setting.

When the chatbot publishes its perfectly inferred answer, traders still read it imperfectly. But there is now no shared answer error. Their uncertainty falls, and equilibrium prices become more informative than in either of the other economies. Prices nevertheless remain imperfect: the platform knows the fundamental exactly, while individual traders do not.

Better information, more trading

The chatbot economy also produces the highest expected trading volume. Traders put more weight on their individual readings of an answer that lacks common error, generating greater dispersion in their positions.

Thus, within this model, more trading accompanies more informative prices. Heavy turnover alone need not indicate that a market is poorly informed.

Where the result could break

Perfect learning rests on demanding assumptions: one dominant platform, an effectively infinite population, independent signal errors, known statistical relationships and instantaneous processing. Query submission is assumed rather than explained through traders’ incentives. The model is static, uses normally distributed signals and does not capture fat-tailed returns.

The chatbot is an ideal Bayesian aggregator, not a tested language model. If its answers contain shared errors, including hallucinations, the paper places it back in the noisy-analyst framework.

This arXiv preprint therefore offers a conditional mechanism, not evidence that today’s chatbots outperform exchanges. Its broader question is who gains access to dispersed knowledge. Free advice could improve market information while concentrating an informational advantage in the platform supplying it. The paper discusses manipulation risks, but does not establish that such behavior occurs.