Jump detection tests in financial time series ? a deep learning approach

In most financial market models, the asset price is driven by continuous Brownian motion. An additional complexity to such a model is the inclusion of a discontinuous jump process. Jumps are theorised to be rare, sudden, and thought to be the result of the market reacting to new information. Jump tests are identified as crucial to understand market incompleteness arising from this discontinuity. Across studies, the Lee and Mykland (2007) method emerges as one of the strongest performers in jump detection. This serves as the benchmark to the jump tests created in this dissertation. The first uses a Long Short-Term Memory (LSTM) neural network based supervised learning approach. The second uses unsupervised learning in the form of a Convolutional Neural Network (CNN) autoencoder. Bates, Merton and Stochastic Volatility double Jump (SVJJ) models provide the data used for comparison. For supervised learning, synthetic data is essential as jump labels are needed for training. The autoencoder jump test is an improvement as it does not need labelled jumps to train. This was found to be the best jump test overall when compared out of sample. Both methods were found to beat the benchmark set by Lee and Mykland (2007). The performance metrics used are suited to the imbalanced data sets arising from the assumption of jumps being rare.

Grantor

Department of Finance and Tax

Year dc:date.issued

2023

Author dc:creator

• Wagener, Justin

Advisor dc:contributor.advisor

• Ouwehand, Peter

• Finance and Tax

Handle dc:identifier.uri

http://hdl.handle.net/11427/40201

OAI identifier oai:identifier

oai:open.uct.ac.za:11427/40201