Time Series Classification with Convolutional Neural Networks: Automated Trading by Pattern Recognition
Specialeforsvar ved Laura Perge
Titel:Time Series Classification with Convolutional Neural Networks: Automated Trading by Pattern Recognition
Abstract: Deep learning models, and within the field, even more, convolutional neural networks have become widely popular for large-scale regression and classification purposes. These models can be utilized by traders in the financial markets. In the paper, we outline an automatic trading solution that is based on a convolutional neural network model. The inputs of the model are 30 x 30 x 3 images obtained by transforming the daily return series of financial assets in a 30-day rolling window fashion. Three image transformation strategies are applied on the return series to ensure that we capture both the dynamic and static properties of the times series: Recurrence Plots, Gramian Angular Fields and Markov Transition Fields. Each image is labelled as a trading order, ”Buy”, ”Sell”, or ”Hold”, based on whether the last price making up the return series which is used to create the image is a local minimum, maximum, or neither. The convolutional neural network is then trained on such images so that given a new price appears, a new image is created, and the model returns the most probable trading order. The model is built both in an asset-specific and in a universal setting. The asset-specific is trained on the returns of the one asset that the predictions are made for, while the universal model is trained on the returns of many different assets from various asset classes. The results show that, during the 2008 financial crisis, both model builds provided significantly higher annual returns than other more traditional algorithmic trading strategies, and, in periods of increasing prices, they do not provide significantly different annual returns either.
Furthermore, the universal model setting is more robust, and achieves better classification performance on the test set than the asset-specific, however, on average, there is no significant difference between the annual returns obtained by the two different approaches.
Vejleder: Rolf Poulsen
Medvejledere: Kenneth H. M. Nielsen, Lasse Bøhling, Ernst & Young
Censor: Elisa Nicolato, Aarhus Universitet