Time Series and Sequence Learning2023HT
|
|
Course plan
No of lectures
11 lectures, 3 pen and paper exercise session and 7 computer lab sessions
Recommended for
Given as MSc level course but where PhD students are invited as well and will get PhD study subject points upon finished course, in agreement with their supervisor. For PhD students: Registration through this portal ONLY, not through Ladok.
The course was last given
Goals
After completion of the course, the student should on an advanced level be able
to:
- apply state-of-the-art methods for the analysis of sequential (e.g., time
series) data,
- account for major principles for the selection, estimation and validation of
sequential models,
- use statistical and numerical software to fit appropriate time series models
to given data sets, make inference about time series components, and compute
forecasts and their statistical uncertainty,
- demonstrate insightful assessment of the generalization capacity of the
statistical relationships on which forecasts can be based.
Prerequisites
Organization
Content
The course provides basic skills and knowledge about state-of-the-art methods
needed for professional work in which sequential data are explored, modified,
modelled and assessed. The course focus is on:
Linear autoregressive models (AR and ARMA)
Nonlinear autoregressive model, including temporal convolutional networks
State space models, Kalman filtering and smoothing
Nonlinear state space models and Sequential Monte Carlo filtering
Recurrent neural networks
Model estimation, validation, and forecasting
Literature
A side-effect of this uniqueness, however, is that we have not been able to
find a single (good) text book that covers all the topics of the course. We
have tried to make the lectures as self-contained as possible, but of course,
there is still a point in being able to read about the material on your own. To
help you find relevant literature, a list of suggested reading is given below.
The reading advice mainly refer to three different text books:
[SS] R. H. Shumway and D. S. Stoffer, Time Series Analysis and Its
Applications: With R Examples, Springer, 4th edition, 2016. There are a few
copies available at the library. You can find the pdf online.
[DK] J. Durbin and S. J. Koopman, Time Series Analysis by State Space Methods,
Oxford University Press, 2nd edition, 2012. Available through the university
library.
[GBC] I. Goodfellow, Y. Bengio and A. Courville, Deep Learning, MIT Press,
2016. Freely available here http://www.deeplearningbook.org/
Lectures
Introduction. Classical regression. [SS 1.1-1.2 and 2.1]
Stationarity, ACF, AR models. [SS 1.3-1.5] and some parts of [SS 3.1, 3.5].
Alternatively [K 2, 4.1, 7.3]
NAR, TCN. The papers https://arxiv.org/pdf/1803.01271.pdf and
https://arxiv.org/pdf/1609.03499.pdf for the TCN model. Also, [GBC 6, 9] for an
introduction to MLP/CNN.
Local level model and the Kalman filter. [DK Ch 2]
Structural time series models. [DK 3.1-3.2, 3.4, 3.6, 3.10, 4.1-4.3]
alternatively [K 9-12]
Smoothing and the EM algorithm. [DK 4.4-4.5, 7.2.1, 7.3.1-7.3.4]
Nonlinear SSM and particle filtering. [DK 9.1, 11.1-11.4.2, 11.5, 11.7,
12.1-12.4]
Parameter Estimation in Nonlinear SSM. [DK 11.6]
From State-Space Models to Recurrent Neural Networks. [GBC 10.1-10.2], [DK
4.3.5]
Recurrent Neural Networks cont.'d. [GBC 10.5-10.11]
Summary.
Examination
Computer. based exam and computer labs
Examiner
Johan. Alenlöv
Credits
6
Comments
The course follows the Master course 732A80
Page responsible: Anne Moe