Yin, Zexuan; (2024) Neural Time Forecasting With Latent Dynamics. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Neural_Time_Series_Forecasting_With_Latent_Dynamics.pdf - Accepted Version Download (3MB) | Preview

Abstract

This thesis investigates the use of neural network models for time series forecasting with an emphasis on modelling latent dynamics (unobservable time series). Time series forecasting is of great research interest to both industry and academia. It describes the task of predicting the future values of one or more time series conditioned on past information. With growth in the availability of data, researchers have started developing machine learning - more specifically neural networks - to model more complex temporal dynamics. Neural networks excel at modelling non-linearity and requires fewer assumptions on the underlying process. To contribute to the area of deep learning for time series forecasting, I focus specifically on the fact that not all time series can be observed. Financial markets for example, contain unobservable market regimes which drive observed time series such as stock returns and prices. Being able to model these latent dynamics provide us with a useful tool to study what is happening beneath the surface. To achieve this, I bring topics together from other areas of machine learning such as representation learning and Bayesian inference. This thesis is broken down into four experiments. In the first experiment, I develop a stochastic variant of the recurrent neural network which can be used to perform multi-step-ahead time series forecasting, and generate confidence intervals for the predictions. In the second experiment, I study the concept of Granger causality in the presence of a potential confounder. I develop a neural network architecture to model the confounder and show that by taking this into account, one can obtain better forecasting accuracy on the target time series. The third and fourth experiments are concerned with the application of latent variable modelling in financial markets. In experiment three, I bring together deep learning and GARCH models from the field of econometrics and propose a neural architecture for volatility (variance/covariance) forecasting in a low dimensional setting (<5 assets). Finally, in the fourth experiment, I build on my work in the third experiment to propose a model capable of forecasting the volatility of an investment portfolio in higher dimensional settings.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item