Can neural networks understand Fedspeak? — Analyzing minute level effects of FOMC press conferences on US stock market with various multimodal transformers
Loading...
URL
Journal Title
Journal ISSN
Volume Title
School of Business |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Department
Major/Subject
Mcode
Degree programme
Language
en
Pages
45 + 4
Series
Abstract
This paper examines the usage of various neural network models to model the minute-level price changes in the S&P 500 tracking SPY ETF during Federal Open Market Committee press conferences. The models use combinations of past price changes, speech segments of the FOMC Chairman, and various linguistic variables extracted from said segments to predict the one-minute price changes. 38 FOMC press conferences between January 2019 and September 2023 are studied through the following model families: linear regression of linguistic features to replicate a previous study, unimodal transformer-based BERT models (BERT, FinBERT, ELECTRA), multimodal versions of the same BERT models that incorporate previous price changes, and a Temporal Fusion Transformer time series model. This study shows that BERT models are not able to predict the direction or size of the movements to any capacity, no matter which model, inputs, or target variables are used. Temporal Fusion Transformer only predicts an average line but is able to make quantile predictions that mostly cover the actual changes. Where the original regression study found highly future-oriented speech, moments discussing the statement, and the interaction of the terms to be statistically significant, this study finds only the interaction and statement-related moments to be significant. Additionally, after accounting for autocorrelation, only the interaction term remains significant and negative, very loosely implying that the Chairman speaking in a future tense about the statement may calm the markets. Overall, the tested neural network models are not able to accurately model the price changes during the conferences. There is some support that the moments discussing the statement are more important, but the evidence is not conclusive. Likely reasons are that the models are not able to take in enough context, as the markets do not necessarily react to what the Chairmans says, but rather to how it differs from market expectations. Despite significant autocorrelation in the price changes, the changes behave very erratically in a way that cannot be reasonably modeled from the past changes alone.Description
Thesis advisor
Malo, PekkaSihvonen, Jukka