TU Wien CAIML

CAIML Seminar: “Stock and Crypto Price Forecasting by Generative AI”

CAIML Seminar with Helmut Prendinger.

helmut-prendinger.jpg

On July 30, 2024, CAIML Seminar with Helmut Prendinger took place.

Abstract

The objective of our work is to accurately predict the closing price of financial instruments such as stocks or crypto. We focus on short-term forecasting, for instance, the next 3 to 7 hours (on hourly chart) or the next 3 to 7 days (on daily chart). Recently, large language models (LLMs) have driven rapid progress in natural language processing (NLP) and computer vision (CV). Chat-GPT is a well-known example of the remarkable abilities of these models in understanding and generating human-like text. An exciting new development is to repurpose LLMs for general time series forecasting. If we encode time series, such as past price action, as a string of numerical digits, we can frame the time series forecasting problem as the problem of next-token prediction in text. This opens the possibility that we can zero-shot extrapolate time series, that is, generate future price values without additional training.

In this talk, we explain how generative AI, that is, AI capable of generating text, images, but also time series, can be utilized in forecasting tasks. The discussion includes (1) time series foundation models, (2) LLMs, and also (3) large multimodal models (LMMs), where the model can be prompted not only with text, but also with images, such as candlestick charts. We present our studies in forecasting Bitcoin and stocks price action based on generative AI. Based on these (often surprising) findings, we are confident that our work can make a significant contribution to automated swing trading.

About the Speaker

Helmut Prendinger is a full professor at the National Institute of Informatics, where he works in the Digital Content and Media Sciences Research Division. He carries out research in the area of Artificial Intelligence, specifically Deep Learning and recently large language models (LLMs).