Explaining contrastive models for financial time series forecasting

This thesis work is carried out within the context of financial time series forecasting, a subject that has undergone innovations over the years from the point of view of tools. The task consists in leveraging historical and current stock prices to predict future values over a period of time or a specific point in the future. For years the most common method used by traders and analysts have been technical indicators calculated from time series, but lately machine learning is increasingly being exploited. In this thesis work, the objective is to generate deep representations of financial time series by exploiting Contrastive Learning frameworks, explain them by means of technical indicators (i.e., a new dataset) and compare performances of the two approaches (contrastive vs technical indicators) on the same downstream task. In order to obtain the aforementioned explainability, we used an ensemble model that receives as input the time series having as features the technical indicators and the results of a clustering algorithm, applied to contrastive representations, as labels. Next, the most important features for label (cluster) prediction are filtered by a threshold. The downstream task is the forecasting of future values of the time series with 3-, 5-, 7-days horizons and is applied both to the latent representations and to the dataset consisting only of the filtered technical indicators. The goal is to verify the capability of light-weighted models (i.e. simpler models) to achieve comparable results with respect to the ones of contrastive models.