QUEEN’S BUSINESS SCHOOL FINANCE SEMINAR SERIES

Wednesday 10^th December

1pm

“Forecasting Volatility with Sentiment: Do we need to catch the hype of AI and LLM?”

Maggie Chen

Cardiff University

Abstract

Volatility forecasting for Bitcoin is of great interest but challenging due to its complex features. Although the GARCH  models are typically used, they are limited to accommodate properties exacerbated by extreme movements and behavioural factors. In the era of AI, we investigate whether the old (GARCH) can work with the new (machines or agents) that are equipped with great capacity to actually produce better forecasting for the highly liquid and risk Bitcoin. We start with a simple GARCH-LSTM (Long Short-Term Memory)  structure and find strong evidence of better forecasting performance. Then, we introduce deep learning algorithms or LLM agents to news sentiment signals - that includes three architectures benchmarked under strict leakage controls: Convolutional Neural Network (CNN)-(LSTM), Attention-LSTM and Transformer. We find that the Attention-LSTM, which employs an attention mechanism to dynamically weight time steps and sentiment features beats others, achieving the lowest errors for 1- and 5-day forecasting in- and out-of-sample. We conclude that sentiment signals are important for accurate Bitcoin volatility forecasting; machines and AI agents are stronger than traditional models but AI is not necessarily more intelligent than the well defined machines.

QBS Conference Hub, Seminar Room 01.012