Διπλωματική εργασία--Πανεπιστήμιο Μακεδονίας, Θεσσαλονίκη, 2023.
(EL)
Approved for entry into archive by Κυριακή Μπαλτά (
[email protected]) on 2023-10-02T07:08:30Z (GMT) No. of bitstreams: 2
PavlidisPavlosMsc2023.pdf: 1115325 bytes, checksum: face55b74e2d687d1dbf93ad5fce1e32 (MD5)
license_rdf: 805 bytes, checksum: 4460e5956bc1d1639be9ae6146a50347 (MD5)
(EN)
Submitted by ΠΑΥΛΟΣ ΠΑΥΛΙΔΗΣ (
[email protected]) on 2023-09-30T11:33:14Z
No. of bitstreams: 2
license_rdf: 805 bytes, checksum: 4460e5956bc1d1639be9ae6146a50347 (MD5)
PavlidisPavlosMsc2023.pdf: 1115325 bytes, checksum: face55b74e2d687d1dbf93ad5fce1e32 (MD5)
(EN)
Made available in DSpace on 2023-10-02T07:08:30Z (GMT). No. of bitstreams: 2
PavlidisPavlosMsc2023.pdf: 1115325 bytes, checksum: face55b74e2d687d1dbf93ad5fce1e32 (MD5)
license_rdf: 805 bytes, checksum: 4460e5956bc1d1639be9ae6146a50347 (MD5)
Previous issue date: 2023-07-03
(EN)
This paper attempts to provide a comprehensive study on stock price prediction using LSTM neural networks and compare their performance. Using 10 years of data from the US stock market index S\&P 500, several simple LSTM and LSTM with Attention models were trained. A novel rolling window approach was utilized for the training procedure, where each model was trained on subsequent, non overlapping subsets so that the weights of the model are updated regularly to capture the ongoing trends. The experimental results revealed that models with smaller architecture outperformed larger models and that dropout, loss function, and model type all have little impact on performance.
(EN)
University of Macedonia
