Many data sets are temporal by nature, i.e. time series. Typical tasks faced by data scientists are: analyzing time series in different domains and developing statistical models based on the data, in order to forecast future values or classify the time series into predefined categories.
This course provides a comprehensive introduction to analysis, forecasting and classification of sequential data. The course adopts a practical approach: theoretical concepts are illustrated and applied in specific case studies. Students will also learn to identify the tools best suited for a given task.
The first part of the course presents a survey of the state-space approach to exponential smoothing, which currently achieves state-of-the-art performance in time series forecasting. It starts from the simpler exponential smoothing model and makes it more realistic by adding trend components, seasonal components, and so on. The techniques are applied to real data set containing hundreds of related time series.
In the second part of the course students learn how to analyze digital signals in different domains, i.e. time and spectral domain; they learn how to extract meaningful features from digital signals suitable for classification. Finally, they learn how to set up and learn statistical models, such as HMMs or DNNs, for recognizing and classifying time series.