Προβλέψεις αποδόσεων μετοχών του Ελληνικού Χρηματιστηρίου με τη χρήση μηχανικής μάθησης
Forecasting stock returns of the Greek Stock Exchange using machine learing techniques
View/ Open
Keywords
Τεχνητή νοημοσύνη ; Θεωρία χρονοσειρών ; Αποδόσεις μετοχών ; Νευρωνικά δίκτυαAbstract
In a world dominated by uncertainty, where every aspect is in constant flux, the human
imperative to formulate accurate forecasts takes precedence in risk reduction. The advent of
new technologies, coupled with the surge in computing power, has revolutionized the creation
of more precise predictions. This thesis delves into the theoretical groundwork surrounding the
analysis and forecasting of time series using classical methods. It explores synthetic elements,
model types, various statistical measures, the identification of stagnation, and diverse
forecasting methods tailored to specific problems. The second chapter addresses machine
learning, its types and distinctions from deep learning. Simultaneously, it discusses the
biological neuron, theoretical frameworks and the operational modes of four architectural
artificial neural networks, LSTM, GRU, RNN and MLP, in depth. In addition, this section will
present the most well-known optimization functions and loss functions. In the third chapter, the
thesis will provide general features about stock returns and ways to calculate them.
Subsequently, the logarithmic returns of three companies in the Greek stock market,
specifically, the group of GEK TERNA S.A., AEGEAN AIRLINES S.A., and OPAP S.A. will
be studied. This analysis will involve the use of four architectural neural networks (LSTM,
GRU, RNN and MLP) along with six different optimization algorithms. The efficiency of these
models will be compared based on specific informative error criteria, leading to the
identification of the most effective models for each network. In conclusion, the conversion of
predicted logarithmic returns to predicted closing values will be executed and the results will
be presented through charts.