Today, electronic payments by credit card have become widespread on the planet thanks to the Internet. These payments represent an ever-increasing volume of more than US$500 billion annually, and fraud accounts for less than 0.5% of the number of these transactions. Globally, this represents an estimated US$25 billion in losses per year and this loss is increasing.
Irrespective of the amount of this fraud, all possible detection measures must of course be implemented to limit its spread. This is why a great deal of research has been carried out on this problem over the last 20 years.
As the publisher of a high-performance transactional platform for payment systems, LUSIS is therefore primarily interested in fraud countermeasures. Within the framework of the LUSIS chair, we wish to study fraud detection both from the point of view of algorithm performance, but also with a constraint of realism of implementation and this on real data.
From a technical point of view, the difficulties and locks are at different levels :
- very unbalanced data sets, there are less than 0.5% of fraudulent transactions,
- need to avoid false positives at all costs,
- emergence of new fraud strategies,
- online detection more difficult than offline,
- changes in customers’ consumption habits (concept drift),
- need to explain the refusal of a transaction
Automated trading systems seek to place orders on the financial markets so as to make capital grow while at the same time limiting the inherent risk. The prices of financial products are time series that vary according to many parameters that are not all observable. Within machine learning, time series are very specific data that require the use of specific methods
One of the characteristics of machine learning approaches to automated trading lies in the evaluation of performance. More often than not, the assessment is based solely on the accuracy of the market direction (up or down). However, on a series of predictions, a model with high accuracy can also result in a high loss since the amplitude is not taken into account.
As part of the Chair, we study:
- metrics with true exploitability beyond accuracy,
- the use of backtesting for the construction of more efficient models,
- reinforcement learning reinforcement for the construction of robust models.