Application des techniques d'apprentissage automatique pour la prédiction de la tendance des titres financiers

Abstract

This thesis examines the use of machine learning techniques for predicting the trend of financial intraday time series (e.g.: minutes to hours) and explain how we could design an autonomous decision-making system to automate the trading of financial securities. One of the challenges of financial time series is the asymmetric distribution of the price movements that introduce bias to the data used to train our models. Three methods are proposed in this thesis to deal with this problem. Among other things, resampling the data, modifying the default probability cutoffs used to generate the classifications and the use of cost sensitive learning to penalize the misclassifications of the majority class. We have observed the existence of time dependencies in the price series of the securities used in this project. This component needs to be considered during the process of training and evaluating our models. Indeed, the accuracy of the inference made from the training data on new unseen data will depend on the similarity of the statistical properties of the two sets of data. The existence of drifts in financial time series is a common phenomenon. By comparing data on different time windows, we can see how the distribution of the target variable changes over time. One solution we propose in this thesis is to use a dynamic rolling evaluation and forecasting technique, where each new instance is predicted from a dynamic set of data. For each new data point, the most recent collected value is compared to the model prediction and integrated into the training data. We also illustrated that the selection of features to use in prediction (the technical indicators for this project), is a task that requires a lot of rigor. Strong multicollinearities can be found for technical indicators that are derived from the same series (e.g. price series). Using indicators that reflect the same information may introduce a lot of noise into our models and affect their performance. We have explained how we can deal with this problem and how we can reduce the dimensionality of our data as well as the interaction between the technical indicators. At the end of this thesis, we demonstrated that it is possible to predict intraday financial time series trend by using machine-learning techniques. From a statistical perspective, the performance of our predictions was significantly better than a random guess, which contradicts the hypothesis of the random walk. We also back-tested our models to see if it is possible to implement winning trading strategies based on machine learning techniques. The results are significantly higher than the expected performance of a passive trading strategy.

Date	12 Nov 2019
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Edmond T. Miresco (Supervisor)