AI-based forecasting of complex timeseries: A tutorial with hands-on artificial and real-world data

The activity is a mini-course with tutorial/practical cross-disciplinary focus on the design, training, and testing of Artificial Intelligence (AI) models for the forecasting of complex oscillating time series. Participants will be introduced to basic concepts of model identification and forecasting of temporal data series, specifically adopting linear auto-regressive (AR) models to fix a benchmark and Feed-Forward (FF) and recursive (R) Neural Networks (NN) to advance the forecasting accuracy and robustness with AI-powered nonlinearities.

When predicting one-step ahead, AR models are the first “cheap” choice, grounded on the efficient least squares method. FFNNs add nonlinearities, though still offering a “static” approach that considers past samples as independent inputs, and losing efficiency and accuracy over long regressive horizons. RNNs are the edge-cutting technology. They fully exploit the data temporal dimension, thanks to an internal state that is updated using the same efficient recursive rule over the input series.

When predicting multiple steps ahead, the recursive use of one-step predictors is not the best practice, especially on series that show a sensitivity to perturbations typical of random-like (i.e., chaotic) oscillating data. While a multi-model approach (one model for each step ahead to be predicted) is possible for both the AR and NN frameworks, NNs offer the easy and efficient multi-output alternative, where the same model is trained to optimize the forecasts over an horizon of future steps.

Participants will be guided step-by-step through these methodologies (AR, FNN, RNN, single-step, multi-step recursive, multi-model, multi-output) and their pipeline of implementation, from the configuration of the relevant tools (only freeware tools will be used, like PyTorch and TensorFlow), to the end-product: the one- or multi-step predictor. Didactical case studies, based on artificial data generated by simulating known dynamical models, will be used to present and compare the methodologies, and proposed as exercises alike. Real-world case studies, based on data series in the fields of natural and environmental sciences, civil and environmental engineering, bio-medical engineering, and finance will be used to showcase the descriptive power of the presented methodologies.

The activity is co-designed with:

• One of the EU-leading group on AI-based analysis of natural and environmental data series (the Chair of Smart Water Management at the Technische Universität Berlin – Einstein Center Digital Future), with long-lasting experience on both short-term high-frequency (hourly, daily) and historical data;
• A group of computer scientists in the Faculty of Electronics and Information Technology at the Warsaw University of Technology, EU-leading experts in genomic data.

Biophysical and financial data series will be selected with the help of biomedical and management colleagues in the PoliMi campus. Mostly publicly available data will be used.

Lectures and technical/practical notes will be provided to participants.

The reference for further reading is "Deep Learning in Multi-step Prediction of Chaotic Dynamics: From Deterministic Models to Real-world Systems", by M. Sangiorgio, F. Dercole, G. Guariso, Springer 2022.

By the end of the course, participants will be able to:

• Understand and compare linear (AR) and AI-based (FFNN, RNN) models for forecasting complex oscillating and chaotic time series;
• Design, train, and implement one-step and multi-step forecasting models using open-source tools (e.g., PyTorch, TensorFlow);
• Evaluate and select appropriate forecasting strategies (recursive, multi-model, multi-output) based on data characteristics and prediction horizon;
• Apply advanced forecasting methodologies to synthetic and real-world datasets in environmental, engineering, biomedical, and financial domains.

Essentially no specific knowledge or skill is required, because of the tutorial style of the mini-course. Basic knowledge of dynamical systems, linear and nonlinear, in both continuous and discrete time, and coding skills in matlab/python are definitely useful for a proficient participation.

Participation is limited. In case requests exceed the available spots, selection will be based on GPA and credits earned from completed exams.

Total of 8 sessions of 45-minute hours each (+ 15-minute break in between):

• April 20^th - 4:30–7:45 PM CET;
• April 21^st - 4:30–7:45 PM CET;
• April 22^nd- 4:30–7:45 PM CET;
• April 23^rd- 4:30–7:45 PM CET;
• April 27^th- 4:30–7:45 PM CET;
• April 28^th- 4:30–7:45 PM CET;
• April 29^th- 4:30–7:45 PM CET;
• April 30^th - 4:30–7:45 PM CET.

Total hours: 28.

The sessions will be held in a room at Polimi, campus Città-Studi, equipped with Wi-Fi access and power outlets. Participants are requested to use their own laptop (Windows, Mac, Linux). The meetings will be recorded and streamed through the Webex platform for students of the ENHANCE universites who will participate remotely. Recordings will be made temporarily available to all participants.

The two-level mutual trust-based quality assurance scheme has been adopted:

• at the university level: Politecnico di Milano has applied its internal quality assurance procedures and structures to the proposal of AI-based forecasting of complex timeseries: A tutorial with hands-on artificial and real-world data it submitted to ENHANCE and to its implementation - the related learning activities,
• at the Alliance level: the body composed of Education Officers has made decisions regarding the inclusion of AI-based forecasting of complex timeseries: A tutorial with hands-on artificial and real-world data proposed by Politecnico di Milano to the Innovative Learning Campus part of the joint ENHANCE educational offer, based on the compliance with the formal requirements and ENHANCE goals.

For the final assessment, participants will be divided into small groups and assigned case studies. Each group will work independently on the project in the weeks following the course. The results of the group work will be presented by the participants in online meetings.

Applications are open from March 3 to 24 at 12:00 PM following the procedures listed below. Please, take the time to read them carefully in order to choose the right path.

ENHANCE students (a part from Polimi):

Applications will be submitted via the following form.

Please, be aware that for submitting your application you should have to be logged in with a Google account.

OR

Polimi students (only):

Applications will be submitted via the following link: Passion in Action.

Please, be aware that only the applications submitted through the Passion in Action website will be considered valid.

Either online or on-site at Politecnico di Milano.

Available seats:

• 20 Polimi ;
• 20 for ENHANCE partner universities.