NGUINABE JOSUE

Modeling and forecasting multivariate time series data using bootstrap resampling, recurrent neural networks, and transformers

Statistics and Data Science play a fundamental role in several areas of knowledge. One of its biggest sub-areas in terms of methodology and applicability is time series analysis and signal processing. Traditional models for time series analysis and forecasting, such as the Autoregressive Integrated Moving Average (ARIMA) architecture, which are linear models, have been successfully applied in time series analysis and forecasting. However, over recent years, the use of Artificial Neural Networks to tackle similar tasks has gradually captured the attention of the research community. This PhD project intends to develop and apply methodologies for time series forecasting, both point forecasts and prediction intervals, for multivariate time series. For that, we will consider adaptations and combinations of statistical and machine learning concepts and methodologies, such as ARIMA, exponential smoothing, recurrent neural networks, transformers, and singular spectrum analysis.