Abstract: Adversarial validation and adversarial training approaches are used in order to face the concept drift problem and the the adversarial attacks to your model. These two concepts are mainly used in computer vision and NLP models. We will see a specialization in the financial stock market for predictive models, in particular in this session we will see these two approaches applied in the field of prediction of stock price market.
To face these scenario we will use a forecasting model trained to predict when to BUY or SELL a stock. We will see a complete pipeline implementation from stock market data aquisition to real time prediction in PRODUCTION. We will see the performace of this model in a first phase when subject to concept drift between training data and test data and to adversarial attacks which make the model powerless in the task of predicting. Then we will see how adopting adversarial validation and adversarial training techniques clearly improves the quality of predictions increasing the robustness to variations in the features of dataset that you will face in production. The implemented model will be evaluated on what are the optimal predictions that can be made with the test data. In order to create these benchmarks, the price and volumes of the stocks will be analyzed in advance so as to be able to calculate in advance when to BUY and when to SELL. After calculating the earnings based on these perfect transactions, the training of the model can begin. The implemented model uses as training features calculated on optimal transactions performed on the dataset. These features will be used as the foundation of the predictive model in order to be able to read possible optimal transitions even in production, in real time. A significant improvement of the model will be dictated by the adoption of the concepts of adversarial validation and adversarial training.
Bio: Federico is a Computer Science Engineer at Machine Learning Reply. He has Master's Degree in robotics and AI at UniGe and has more than five years of experience in machine learning and related fields, with a focus in computer vision and NLP, among the others. Lately, he specialized in large language models enabling the visual-language-action (VLA) paradigm and the concerns that may arise in society (the so-called AI safety).