Repository of colleges and higher education institutions

Show document
A+ | A- | Help | SLO | ENG

Title:Napovedovanje časovnih vrst z modeli globokega učenja : primerjalna analiza TFT, LSTM IN GRU
Authors:ID Bučar, Janez (Author)
ID Boškoski, Pavle (Mentor) More about this mentor... New window
Files:.pdf MAG_2025_Janez_Bucar.pdf (3,93 MB)
MD5: B1BDEF4F3F8A7485E042D8B0B95BBBA2
 
Language:Slovenian
Work type:Master's thesis/paper
Typology:2.09 - Master's Thesis
Organization:FIŠ - Faculty of Information Studies in Novo mesto
Abstract:V magistrskem delu smo raziskovali učinkovitost treh modelov globokega učenja za napovedovanje prihodnjih cen Bitcoina. Modeli, ki smo jih uporabili, so Long-Short Term Memory (LSTM), Gated Reccurent Unit (GRU) in Temporal Fusion Transformer (TFT), ki velja za najnovejšo arhitekturo med izbranimi modeli. Analizo uspešnosti modelov smo opravili na več evalvacijskih metrikah, kot so RMSE, MAE, MAPE, SMAPE in MASE. Poleg natančnosti napovedi smo analizirali tudi interpretabilnost modelov ter vpliv makroekonomskih kazalnikov in kategoričnih spremenljivk. Optimizacijo hiperparametrov smo izvedli s pomočjo knjižnice Optuna, da smo izkoristili njihov potencial. Rezultati kažejo, da je TFT model dosegel najboljše napovedne rezultate ter omogočil interpretacijo dejavnikov, ki imajo vpliv na prihodnjo ceno Bitcoina. Hipoteze, ki so se nanašale na primerjavo natančnosti, časa učenja in pomen vhodnih značilk, smo ovrednotili na podlagi eksperimentalnih rezultatov.
Keywords:napovedovanje časovnih vrst, Bitcoin, LSTM, GRU, Temporal Fusion Transformer, interpretabilnost, Optuna
Place of publishing:Novo mesto
Place of performance:Novo mesto
Publisher:J. Bučar
Year of publishing:2025
Year of performance:2025
Number of pages:XV, 76 str.
PID:20.500.12556/ReVIS-12397 New window
COBISS.SI-ID:252972291 New window
UDC:004.8(043.2)
Note:Na ov.: Magistrska naloga : študijskega programa druge stopnje;
Publication date in ReVIS:13.10.2025
Views:162
Downloads:5
Metadata:XML DC-XML DC-RDF
:
Copy citation
  
Share:Bookmark and Share


Hover the mouse pointer over a document title to show the abstract or click on the title to get all document metadata.

Licences

License:CC BY-NC-SA 4.0, Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
Link:http://creativecommons.org/licenses/by-nc-sa/4.0/
Description:A Creative Commons license that bans commercial use and requires the user to release any modified works under this license.

Secondary language

Language:English
Abstract:In this master’s thesis, we explored the performance of three deep learning models for forecasting future Bitcoin prices: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Temporal Fusion Transformer (TFT), the latter being the most recent among them. We evaluated the models using several metrics, including RMSE, MAE, MAPE, SMAPE, and MASE. In addition to forecast accuracy, we examined model interpretability and the influence of macroeconomic indicators and categorical variables. Hyperparameter optimization was performed using Optuna. Results show that the TFT model achieved the best predictive performance and provided valuable insights into the key factors affecting Bitcoin’s future price. Hypotheses related to accuracy, training time, and feature importance were evaluated based on experimental results.
Keywords:time series forecasting, Bitcoin, LSTM, GRU, Temporal Fusion Transformer, interpretability, Optuna


Back