Σχολή Επιστημών και Τεχνολογίας της Πληροφορίας
Μόνιμο URI για αυτήν την κοινότηταhttps://pyxida.aueb.gr/handle/123456789/2
Η Σχολή Επιστημών και Τεχνολογίας της Πληροφορίας περιλαμβάνει τα Τμήματα: - Τμήμα Πληροφορικής - Τμήμα Στατιστικής
Περιήγηση
Πλοήγηση Σχολή Επιστημών και Τεχνολογίας της Πληροφορίας ανά Συγγραφέα "Aloupis, Konstantinos A."
A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω
Τώρα δείχνει 1 - 1 από 1
- Αποτελέσματα ανά σελίδα
- Επιλογές ταξινόμησης
Τεκμήριο Towards an artificially intelligent trading system. Using conventional time series analysis, along with Long Short Term Memory Recurrent Neural Networks for stock price forecasting: a comparative study, on the seven highest capitalization stocks of the FTSEASE25 index(12/21/2018) Aloupis, Konstantinos A.; Αλούπης, Κωνσταντίνος Α.; Athens University of Economics and Business, Department of Informatics; Karlis, Dimitrios; Vassalos, Vasilios; Koutroumbas, KonstantinosMachines baring the gift of intelligence have been dreamt by inventors for eons. This long lasting aspiration dates back to Daedalus, Hephaestus, Talus and Pandora. In the last few years a critical mass of interconnecting counterparts such as increase in computational power and cognitive advances in algorithms, has empowered and democratized artificial intelligence, while broadening its applications. Numerous fields have been benefited by this conjuncture; visual pattern recognition, natural language processing, data mining to name a few. A spill-of effect from these scientific regions to the financial field is silently increasing its pace. Long tested and widely accepted traditional techniques in financial time series forecasting, are gradually grafted with Machine learning and Artificial Intelligence so as to expand the forecasting arsenal. The aim of this thesis is to carry out a preliminary investigation on the ability of Long-Short term memory recurrent neural networks (LSTM-RNNs) to perform reliable one step ahead predictions on the closing price of Greeks stocks. Their performance will be evaluated against a variety of well-established time series models in traditional statistical and technical analysis, such as moving averages (MAVG) and Auto Regressive Integrated Moving Average (ARIMA). Additionally they will be tested against the basis of feed forward neural networks (FFNN), the Multilayer Perceptron (MLP). All the above mentioned models will also be used for performing one step ahead prediction on the Euro-Dollar spot exchange rate, due to its extreme liquidity and long history of prices.Each of the available time series in this study are split into two smaller ones; the first constitutes the so called training set, which is used to estimate the parameters of each model -were necessary-, while the second constitutes the test set, which is used for model evaluation. The latter is performed in terms of the Root mean square error (RMSE) applied on the test set. The model with the lowest average RMSE, will be the one that outperforms the others. For this endeavor, as an implementation tool for constructing the LSTM-RNN, Google’s deep learning framework –known as TensorFlow- is used. The main reason for adopting this extremely complicated and with an inherently quite flat learning curve tool, is its immense scalability and in-depth parameterization and configuration abilities; compared of course to other similar higher level solutions such as Keras and Torch.From our preliminary results it becomes obvious, as it is self-evident, that none of the models achieves high accuracy in short term stock price forecasting. Noticeably better –reduced RMSE approximately by 20% on average when compared to benchmarks-, is found to be the LSTM-RNN under several restrictions only for the large capitalization commercial stocks of TITAN, OPAP, OTE, MOH, EEE and MPELA; representing almost 50% of the FTSEASE25 index capitalization. These findings are completely in line with previous studies on foreign stock exchanges, signaling a new terra incognita for the analysis of the Greek stock market which is in a shocking shortage of such inquires; that is the main contribution of this thesis. One of the many peculiarities with this task at hand, and commonly admitted by the literature to be the hardest part of every research, is the clarification of the data input feed pipeline for the TensorFlow model building stage. This process is demonstrated here methodically, not lacking the proper attention.