CryptexLLM: How LLM Generalizability Forecasts High Volatility

We present CryptexLLM, an approach to time series forecasting that adapts large language models (LLMs) for predicting high volatility data. We extend the TimeLLM framework with an adaptive weighted loss function, feature engineering, and sentiment analysis. Our experiments show that the approach we took outperforms traditional LSTM models and statistical methods, with our best performing model being Llama 3.1. The adaptations we made improve directional accuracy, which is particularly useful for financial applications, while still maintaining computational efficiency. Our results suggest that LLMs have the potential to effectively generalize to volatile time series domains.