Adaptive LLM Factor Mining for Hedged Crypto Perpetual Portfolios

Summary

This proposed system uses a large language model to generate and revise candidate price-and-volume factors for a cross-sectional long-short portfolio of USDT perpetual contracts. It screens liquid instruments, classifies market conditions using Bitcoin volatility, and evaluates candidate factors with Rank IC using lagged factor values to predict subsequent returns. Correlated factors are deduplicated, factors with weak or declining IC may be removed, and surviving factors are combined using recent IC weights. Scores determine which instruments enter long and short baskets, followed by periodic rebalancing.

The design also specifies position limits and leverage parameters, alongside fixed stop loss, fixed take profit, and a dynamic trailing exit. However, the document presents a system description and workflow rather than empirical performance evidence. Its stated IC tests use hourly data, and the factor-generation process can overfit historical patterns or lose predictive value as conditions change. The workflow source is incomplete in the supplied text, so the full implementation and its operational behavior cannot be independently established from this document.

Key ideas

• The workflow generates candidate factors with an LLM and screens them using lagged Rank IC tests.
• It combines cross-sectional standardized factor scores with recent IC weights to rank instruments.
• The described portfolio selects long and short baskets from liquid USDT perpetual contracts.
• Position monitoring includes fixed loss and profit thresholds plus a dynamic trailing exit.
• The document supplies no performance results, and the workflow excerpt is incomplete.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.