Applying Computer Vision techniques to improve stock market prediction

Abstract

Linear models have a long standing in algorithmic trading, they are easy to implement, inference is nearly instantaneous and they rarely over-fit. Nevertheless, these models have one main drawback: they are incapable of finding non-linear relationship without tedious and suboptimal feature engineering. The expressivity of neural networks makes them prime candidates to solve this problem but they often generalise very poorly to new samples (i.e that are not in the training data).

During my internship at BNP Paribas, I implemented techniques that were firstly introduced in Computer Vision to solve this problem. On one hand, we show that learning meaningful representations with Variational Auto-Encoders increases the compression rate of information while preserving a highly structured and continuous space (compared to the PCA). On the other hand, using this representation of the market instead of the raw features improves the prediction score of our indicators.

More concretely, we present a method to reduce by the error rate compared to the PCA algorithm while maintaining a latent space of high quality. Moreover, using the deep neural network architecture used here increases the explained variance of the stock market by compared to linear methods.

How to cite

@mastersthesis{lagesse2022,
  title  = {Applying Computer Vision techniques 
            to improve stock market prediction},
  author = {Adrien Lagesse},
  year   = {2022},
  month  = {September},
  url    = {https://adrien-lagesse.io/publications/cv-for-stock-market-prediction},
  school = {École Polytechnique, Paris},
  type   = {Master's thesis}
}