Exportar Publicação

Pratas. T.E., Ramos, F.R. & Rubio, L.J. (2023). Forecasting Bitcoin’s Volatility: Exploring the Potential of Deep-Learning. 42nd Eurasia Business and Economics Society.

T. E. Pratas et al.,  "Forecasting Bitcoin’s Volatility: Exploring the Potential of Deep-Learning", in 42nd Eurasia Business and Economics Society, Lisboa, 2023

@misc{pratas2023_1732396434919,
	author = "Pratas. T.E. and Ramos, F.R. and Rubio, L.J.",
	title = "Forecasting Bitcoin’s Volatility: Exploring the Potential of Deep-Learning",
	year = "2023",
	howpublished = "Digital",
	url = "https://ebesweb.org/42nd-ebes-lisbon/42nd-ebes-conference-lisbon/"
}

TY  - CPAPER
TI  - Forecasting Bitcoin’s Volatility: Exploring the Potential of Deep-Learning
T2  - 42nd Eurasia Business and Economics Society
AU  - Pratas. T.E.
AU  - Ramos, F.R.
AU  - Rubio, L.J.
PY  - 2023
CY  - Lisboa
UR  - https://ebesweb.org/42nd-ebes-lisbon/42nd-ebes-conference-lisbon/
AB  - The importance of using appropriate statistical, mathematical and computational techniques can significantly impact decision-making. With the recent computational progress, Deep Learning methodologies based on Artificial Intelligence seem to be pointed out as a promising tool to study financial time series, characterized out-of-the-ordinary patterns. Cryptocurrencies are a new asset class with several particularly interesting characteristics that still lack deep study and differ from the traditional time series. Bitcoin is characterized by extraordinary high volatility, high number of structural breaks and other characteristics that might difficult the study and forecasting of the time series. The goal of this study is to critically compare the forecasting properties of classic methodologies (ARCH and GARCH) with Deep Learning Techniques (with MLP, RNN and LSTM architectures) when forecasting Bitcoin’s volatility. The empirical study focuses on the forecasting of Bitcoin’s volatility obtained by squaring the daily returns of Bitcoin’s Price time series resulting in 2753 observations from 08-09-2014 to 01-05-2022. Forecasting quality, using MAE and MAPE for one, three- and seven-day’s forecasting horizons are compared. The Deep Learning methodologies show advantages in terms of forecasting quality (when we take in consideration the MAPE), but also require huge computational costs. In detail, both MLP and RNN architectures seem to have a smoother prediction that fluctuate less but fails to capture the big volatility spikes. While the LSTM architecture seems react strongly to such movements and tries to adjust their forecast accordingly. This happens due to its long-term memory proprieties that allows the model to ‘remember’ such past volatility spikes might result on high volatility spikes on the following day. In addition, Diebold-Mariano tests were also performed to compare the forecasts, reinforcing the superiority of Deep Learning methodologies.
ER  -