TSHIANGOMBA, R. Financial time series analysis based on innovative Machine Learning Signal Processing approaches [online]. Brno: Vysoké učení technické v Brně. Fakulta strojního inženýrství. 2024.

Posudek vedoucího

Cicone, Antonio

In this thesis work the student tackles the forecasting of financial time series which is a challenging and open problem in Finance. Several techniques have been developed and are commonly used, like ARIMA and ARMA. In his work, Reagan, starting from a recent work published by Prof. Feng Zhou and collaborators, developed a brand new approach for financial time series forecasting. The idea was to develop a new kind of machine learning approach that mimics signal processing decomposing approaches and couple it with a machine learning feature extraction approach to produce the extension of a time series frequency by frequency. The newly developed method has been tested against artificial and real-life financial data sets. The results are really promising and the work done by the student is more than satisfactory for me.

Posudek oponenta

Sehnalová, Pavla

The thesis deals with the development of a completely new approach to forecasting financial time series. The student met the assignment requirements and followed the assignment process for developing a new kind of machine learning approach.  The diploma thesis is divided into 6 chapters. In the first chapter, the student somewhat unconventionally describes the current state of knowledge as a literature review and lists a number of sources, some of which are related to the given issue, others less so. It is rather a list of possible techniques of the last two decades. The second chapter deals in detail with two techniques for analyzing signals (IF and IMFs). In the third chapter, the student introduces the general issue of neural networks and explains all concepts in depth and mathematically. However, in the subchapter concerning convolutional neural networks, the explanations of subchapters 3.2.2 to 3.2.6 lack this technical perspective. The fourth chapter continues to explain the concepts related to Artificial Neural Network (ANN), even though the chapter is called Proposed Approach. Below in this chapter, a detailed description of the proposed framework follows, i.e. combining the two outputs of Convolution Neural Network and ANN into Fusion Neural network. The fifth chapter presents the results of 4 experiments and a comparison of the results with the determination of the most suitable variant. The sixth chapter contains the conclusion and possible future work. Although the organization of the chapters may require some rethinking, the clarity with which the methodologies are described and presented is very good. The text is mostly correct, but some statements are a bit vague, e. g. Deep learning models have usually more than 3 layers. Neurons don't just "process" the data; they apply mathematical functions to it. or While EMD allows for calculating the instantaneous frequency of IMFs using the Hilbert Transform, it's not necessarily guaranteed that the IMFs will be perfectly suited for this transformation (depending on the data). The results are impressive. The student worked with a large number of sources that are properly cited, and I rate the level of the written text as high.

eVSKP id 153993