Institutional Repository of Vilnius University
Sintetinių duomenų naudojimas finansinių laiko eilučių analizėje /
Abstract
dc:descriptionForecasting financial time series, particularly in the cryptocurrency market, faces challenges from limited historical data and high price volatility. This thesis investigates whether the use of synthetic data during the training of deep learning models improves cryptocurrency price forecasting accuracy. Three synthetic data generation methods: the stationary bootstrap, "QuantGAN", and "Time-Causal VAE", are compared as training set augmentation sources for three forecasting models (LSTM, BD-LSTM, TCN), used to predict daily prices of three cryptocurrencies (BTC, ETH, XRP) over the 2020–2026 period. A total of 90 experiments were conducted and evaluated using the Friedman and Wilcoxon signed-rank tests. The results show that synthetic data augmentation reduces forecasting error on average, although the magnitude of the effect depends strongly on the specific configuration. "Time-Causal VAE" proved to be the most reliable method, it improved results in 88.9\% of configurations and was the only method to statistically significantly improve the BD-LSTM model across all three cryptocurrencies. The largest benefit was obtained by the TCN model, which had the weakest baseline performance, and by XRP, the cryptocurrency with the heaviest distribution tails. The best results across all three cryptocurrencies were achieved by the TCN and "Time-Causal VAE" configuration, although the gap to BD-LSTM and "Time-Causal VAE" is very small, making both configurations suitable for practical use.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Adamonis, Ąžuolas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308333857&prefLang=en_US