[Paper Review] LLM as a Risk Manager: LLM Semantic Filtering for Lead-Lag Trading in Prediction Markets

Sumin Kim, Minjae Kim+8 more|arXiv (Cornell University)|Feb 4, 2026

Stock Market Forecasting Methods0 citations

TL;DR

A two-stage framework combines Granger causality with an LLM-based semantic filter to rank lead–lag relationships in prediction markets, improving trading performance and reducing downside risk on Kalshi data.

ABSTRACT

Prediction markets provide a unique setting where event-level time series are directly tied to natural-language descriptions, yet discovering robust lead-lag relationships remains challenging due to spurious statistical correlations. We propose a hybrid two-stage causal screener to address this challenge: (i) a statistical stage that uses Granger causality to identify candidate leader-follower pairs from market-implied probability time series, and (ii) an LLM-based semantic stage that re-ranks these candidates by assessing whether the proposed direction admits a plausible economic transmission mechanism based on event descriptions. Because causal ground truth is unobserved, we evaluate the ranked pairs using a fixed, signal-triggered trading protocol that maps relationship quality into realized profit and loss (PnL). On Kalshi Economics markets, our hybrid approach consistently outperforms the statistical baseline. Across rolling evaluations, the win rate increases from 51.4% to 54.5%. Crucially, the average magnitude of losing trades decreases substantially from 649 USD to 347 USD. This reduction is driven by the LLM's ability to filter out statistically fragile links that are prone to large losses, rather than relying on rare gains. These improvements remain stable across different trading configurations, indicating that the gains are not driven by specific parameter choices. Overall, the results suggest that LLMs function as semantic risk managers on top of statistical discovery, prioritizing lead-lag relationships that generalize under changing market conditions.

Motivation & Objective

• Motivate the challenge of identifying robust lead–lag relationships in event-level prediction markets.
• Propose a two-stage framework that combines statistical discovery with semantic validation.
• Evaluate whether LLM-based re-ranking improves trading performance over Granger-only screening.

Proposed method

• Stage 1 uses Granger causality on log-odds-transformed market prices to identify candidate leader–follower pairs across multiple lags (p∈{1,2,3,4,5}).
• Stage 2 re-ranks these candidates with an LLM (GPT-5-nano) using event titles/descriptions to assess plausible economic transmission mechanisms and assigns a plausibility score.
• Trading protocol: for each leader–follower pair, trigger follower trades when leader price changes exceed a threshold, hold for h days, and measure out-of-sample PnL.
• Final portfolio selects top M directed pairs (M=20) from the Stat.; Hybrid approaches for comparison.
• Evaluation uses rolling windows with 60-day training and 30-day testing on Kalshi Economics markets.

Figure 1: Two-stage framework for leader–follower pair discovery in prediction markets. Stage 1 produces a candidate set of Top K directed pairs (K=100) ranked by Granger significance, and Stage 2 applies LLM-based semantic re-ranking to select the final Top M portfolio (M=20).

Experimental results

Research questions

• RQ1 Can LLMs differentiate mechanistically plausible lead–lag relationships from brittle statistical correlations identified by Granger causality?
• RQ2 Does semantic re-ranking improve trading performance and reduce downside risk beyond Granger-based screening?
• RQ3 Are the gains robust across different holding horizons and market conditions?

Key findings

• LLM-based semantic filtering yields substantial total PnL gains (+205%) over the statistical baseline.
• Downside risk is reduced, with average loss dropping from $649 to $347 (46.5% reduction).
• Win rate rises from 51.4% to 54.5% under default settings.
• Loss reduction persists across same-event and different-event pairs (40%+ improvements).
• Semantic filtering provides larger gains during larger leader moves (5–10pt and 10pt+ moves).
• LLM-selected pairs can include economically meaningful links outside Granger Top-M, illustrating qualitative gains beyond statistical rankings.

Figure 2: Signal-triggered trading protocol used to evaluate ranked lead-lag relationships from Figure 1 : leader price moves trigger follower trades, with direction determined by the Granger-induced trade sign and out-of-sample PnL used to evaluate the ranked pair list.