# LLMs for Time Series: an Application for Single Stocks and Statistical Arbitrage

**Authors:** Sebastien Valeyre, Sofiane Aboura  
**arXiv:** [2412.09394](https://arxiv.org/abs/2412.09394)  
**Date:** 2024-12-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2412.09394v1/

## Key points

- Large Language Models (LLMs) have gained traction in various fields, including time series prediction.
- A study utilizing the Chronos model showed that LLMs can effectively handle time series that appear almost random and generate alpha in financial markets.
- The performance simulation indicated that LLMs can identify inefficiencies amidst randomness.
- Results were compared with traditional Short Term Reversal (STR) or trend-following approaches.
- The study used the "amazon/chronos-t5-tiny" version of the Chronos model with 11 million parameters pretrained on various datasets but not specifically on financial time series.
- Three different datasets of residual daily returns were utilized, focusing on liquid stocks with significant market capitalization.
- This research aimed to demonstrate how LLMs can adapt and display intelligence in trading without being explicitly trained for that purpose.

## Summary

In recent years, Large Language Models (LLMs) have gained traction in various fields, including time series prediction. While there is a common belief that LLMs are not suitable for predicting financial market returns due to their perceived randomness, a counterexample challenges this misconception. By utilizing the Chronos model developed by Ansari et al. (2024), both pretrained configurations and fine-tuned supervised forecasts were tested on the largest American single stocks using data from Guijarro-Ordonnez et al. (2022). The study focused on constructing a long/short portfolio based on the predictions generated by the LLM model. Surprisingly, the performance simulation indicated that LLMs can effectively handle time series that appear almost random, showcasing their ability to identify inefficiencies amidst randomness and generate alpha in financial markets. Furthermore, the results were compared with traditional Short Term Reversal (STR) or trend-following approaches as described in Jegadeesh (1990) and Jegadeesh and Titman (1993). The methodology of the empirical backtest involved using the "amazon/chronos-t5-tiny" version of the Chronos model with 11 million parameters pretrained on various datasets but not specifically on financial time series. The study utilized three different datasets of residual daily returns from Guijarro-Ordonnez et al. (2022), focusing on liquid stocks with significant market capitalization. These datasets included IPCA factors, PCA factors, and FF factors, each with specific rolling windows for analysis. Overall, this research aimed to demonstrate how LLMs can adapt and display intelligence in trading without being explicitly trained for that purpose. By fine-tuning the Chronos model and analyzing its performance against established strategies, significant potential was identified for enhancing LLM predictive capabilities in financial markets.