Synthetic Time Series Generation for Decision Intelligence Using Large Language Models

Abstract

A model for generating synthetic time series data using pre-trained large language models is proposed. Starting with the Google T5-base model, which employs an encoder–decoder transformer architecture, the model underwent pre-training on diverse datasets. It was then fine-tuned using the QLoRA technique, which reduces computational complexity by quantizing weight parameters. The process involves the tokenization of time series data through mean scaling and quantization. The performance of the model was evaluated with fidelity, utility, and privacy metrics, showing improvements in fidelity and utility but a trade-off with reduced privacy. The proposed model offers a foundation for decision intelligence systems.