Forecasting bitcoin: Decomposition aided long short-term memory based time series modeling and its explanation with Shapley values

• DOI: 10.1016/j.knosys.2024.112026
• Publication Date: 2024-06-06T15:54:41Z
• Authors: Vule Mizdrakovic; Maja Kljajic; Miodrag Zivkovic; Nebojsa Bacanin; Luka Jovanovic; Muhammet Deveci; Witold Pedrycz

Bitcoin price volatility fascinates both researchers and investors, studying features that influence its movement. This paper expends on previous research examines time series data of various exogenous endogenous factors: Bitcoin, Ethereum, S&P 500, VIX closing prices; exchange rates the Euro GPB to USD; number Bitcoin-related tweets per day. A period three years (from September 2019 2022) is covered by dataset. two-layer framework introduced tasked with accurately forecasting price. In first layer, account for complexities in analyzed data, variational mode decomposition (VMD) extracts trends from series. second Long short-term memory hybrid Bidirectional long networks were used forecast prices several steps ahead. work also an enhanced variant sine cosine algorithm tune control parameters VMD neural attaining best possible performance. The main focus combining modified metaheuristics improve cryptocurrency value forecast. Two sets experiments conducted, without VMD. results have been contrasted models tuned seven other cutting-edge optimizers. Extensive experimental outcomes indicate can be forecasted great accuracy using selected decomposition. Additionally, model was analyzed, Shapley values indicated such as EUR/USD rates, Ethereum prices, GBP/USD a significant impact forecasts.