The FF4EUROHPC cascade funding initiative aims at fostering joint research between companies and academic partners that make massime use of computattional resources.
UNIMORE is the domain expert partner of the experiment "LEVERAGING HPC FOR AI AND DL-POWERED SOLUTIONS FOR ASSET MANAGEMENT" with AxyonAI
The objectives of the Experiment are:
- To build a new set of product features and predictive models based on highly innovative deep learning technologies and requiring high computation power,
- To enhance the current offering of Axyon AI extending its target market especially to small/medium-size customers that need a complete, end-to-end solution,
- To implement solutions to better adapt the models to the rapid changes that often occur in the financial markets.
The Experiment has the overall goal of improving the service offered by Axyon to its clients through several technological advancements. Three main areas of improvement (scalability, risk management and adaptiveness) have been identified and each will be the target of a dedicated task. In particular, during the first part of the Experiment Axyon and Cineca will work on improving the computational scalability of the Axyon ML platform. Then, the focus will be on enhancing Axyon IRIS risk management features for end-to-end portfolio construction. In the last part of the Experiment, with the collaboration of AImageLab, the activities will be focusing on increasing the adaptiveness of Axyon IRIS forecasting models.
In detail Continual learning models will be applied to asset management in order to obtain systems that can effectively learn online in time and adjust to smooth dataset shifts.
Continual learning aims at preventing the well known phenomenon called catastrophic forgetting that prevents a deep learning model to retain its original knowledge when trained in an online fashion.
The research is focused on:
- Demonstrating and observing that forgetting occurs in asset management and time series analysis
- evaluating the continual learning strategies that better fit the problem and provide baseline solution to the forgetting
- design and develop new CL models that enables the exploitation of unlabeled data by discovering new trends hidden inside the data stream and exploit them for strenghtening the online training regime.