Semantic correlations and how to trade them

A price move in one prediction market can carry information for another.

Higher inflation odds may change the probability of tighter monetary policy. A recession signal in one country may affect growth expectations elsewhere. The economic connection can be real, but finding it systematically is difficult.

A statistical test can identify that one market tends to move before another. It cannot tell you whether the relationship has a plausible transmission mechanism or happened to appear in a noisy sample.

That distinction matters once money is attached to the signal.

In our recent research with collaborators from LinqAlpha, MIT, Kalshi, Arrowpoint Investment Partners, UC Berkeley, the University of Florida, and UNIST, we tested a simple division of labor: let statistics find candidate trades, then use a large language model to reject relationships that do not make economic sense.

The result was a better trading portfolio, driven mainly by smaller losses rather than unusually large wins.

The problem with statistically significant trades

Lead-lag analysis asks whether past movements in one time series help predict future movements in another. Granger causality is a standard tool for this job. Applied across prediction markets, it can identify relationships such as one inflation contract moving before a rate-decision contract.

But large pairwise searches create a selection problem. Test enough combinations and some will look significant by chance. Even legitimate relationships can break when policy regimes, market participants, or liquidity conditions change.

Prediction markets make this problem unusually interesting because every time series comes with a natural-language description. A price series is attached to a defined event, not just a ticker. That means a model can inspect the proposed relationship in economic terms.

The LLM does not need to discover the correlation. It can ask a different question: if the first event reprices, is there a coherent reason the second should follow?

A two-stage screen

The research separates discovery from judgment.

Stage one: statistical discovery

The first stage transforms daily prediction-market prices into log odds and applies Granger-causality tests. It evaluates both directions for each event pair and ranks candidate leader-follower relationships by statistical strength.

This produces the raw material: relationships supported by the historical price data.

Stage two: semantic filtering

The second stage gives the event titles and descriptions to an LLM. The model scores whether the proposed direction has a plausible economic transmission mechanism and assigns the expected sign of the relationship.

It does not search for new pairs or replace the statistical test. It re-ranks candidates that have already passed the quantitative screen.

The final comparison is therefore clean:

• A statistical portfolio selects the highest-ranked pairs using Granger evidence alone.
• A hybrid portfolio selects from the same candidate set after semantic re-ranking.
• Both portfolios follow the same mechanical entry, holding, and exit rules.

If performance changes, the difference comes from which relationships survived the second screen.

What the study tested

The experiment used 554 Kalshi markets in the Economics category from October 2021 through November 2025. Each evaluation used a 60-day training window followed by a 30-day testing window, producing 18 non-overlapping out-of-sample periods.

For every training window, the pipeline:

The default comparison used a seven-day holding period. The paper also tested holding periods from one to 21 days.

The LLM improved the left side of the distribution

Under the default configuration, semantic filtering improved the win rate from 51.4% to 54.5%. Total experimental P&L rose from $4,100 to $12,500.

The more important result was in the losing trades:

The hybrid portfolio did not win because the LLM found a few enormous trades. In fact, its average winner was smaller. It won because the model deprioritized relationships that later produced large losses.

That is why "semantic risk manager" is a better description than "AI trader." The LLM's contribution was selection discipline.

Semantic judgment mattered most when markets moved sharply

The benefit increased with the size of the leader market's move. For leader moves of five to ten probability points, the hybrid portfolio's win rate was 66.7%, compared with 57.1% for the statistical portfolio. For moves above ten points, the comparison was 71.4% versus 53.8%.

Win rate by size of the leader market's move

Seven-day holding period, both portfolios on the same execution rules

Statistical screen Hybrid semantic screen

All trades

Statistical 51.4%

Hybrid 54.5%

Leader move of 5–10 points

Statistical 57.1%

Hybrid 66.7%

Leader move above 10 points

Statistical 53.8%

Hybrid 71.4%

Large repricings are exactly where a fragile historical relationship can become expensive. A correlation that looked stable during quiet periods may have no mechanism connecting the two events when new information hits.

The LLM filter appears to help most when that distinction becomes consequential.

The loss reduction also persisted across every tested holding period. Depending on the horizon, average loss magnitude fell by 22.6% to 47.2%, with a mean reduction of 36.2%.

An example: economic mechanism over surface similarity

One relationship promoted by the hybrid approach connected a Japan recession contract to a U.S. GDP contract. The pair ranked only 71st by the statistical test but rose to fifth after semantic review.

The two contracts do not share many words. The proposed mechanism is still intelligible: a downturn in a major economy can affect U.S. growth through trade, financial conditions, and policy spillovers. The model could recognize that channel without relying on ticker similarity.

This is the advantage of combining structured market data with language. Statistics measure what moved first. Language models can test whether the proposed direction has an economic story strong enough to survive outside the original sample.

What this does not prove

The experiment does not establish causal ground truth between market pairs. It evaluates relationship quality through a fixed historical trading protocol. The work is a preprint, and the reported P&L should not be read as a live strategy result or a guarantee of future performance.

The study also leaves implementation questions for future work, including transaction costs, liquidity at scale, execution timing, and the effect of deploying a signal into the market.

Still, the design points toward a useful architecture for systematic trading research:

The strongest use of an LLM may not be generating more signals. It may be knowing which statistical signals deserve less capital.