Is It Possible to Forecast the Price of Bitcoin?

Abstract

This paper focuses on forecasting the price of Bitcoin, motivated by its market growth and the recent interest of market participants and academics. We deploy six machine learning algorithms (e.g., Artificial Neural Network, Support Vector Machine, Random Forest, k-Nearest Neighbours, AdaBoost, Ridge regression), without deciding a priori which one is the ‘best’ model. The main contribution is to use these data analytics techniques with great caution in the parameterization, instead of classical parametric modelings (AR), to disentangle the non-stationary behavior of the data. As soon as Bitcoin is also used for diversification in portfolios, we need to investigate its interactions with stocks, bonds, foreign exchange, and commodities. We identify that other cryptocurrencies convey enough information to explain the daily variation of Bitcoin’s spot and futures prices. Forecasting results point to the segmentation of Bitcoin concerning alternative assets. Finally, trading strategies are implemented.

Conducting Causal Analysis by Means of Approximating Probabilistic Truths

“…While Bitcoin, as an asset class, has only recently attracted the public attention of large institutional investors, many researchers have already analyzed the time-series behavior of Bitcoin prices. An overview of recent developments and more discussion on forecasting Bitcoin prices is by [ 76 ]. They investigate a large set of covariates that cover nearly all important classes of financial assets, except bonds.…”

Ensemble and Multimodal Approach for Forecasting Cryptocurrency Price

“…Recently, Chevallier et al [6] proposed a novel approach for cryptocurrency forecasting that improved the performance significantly while keeping the approach simple, which is an AdaBoost that uses multiple decision trees weak learners. Surprisingly, the results have demonstrated that AdaBoost outperforms all ANNs, LSTMs, KNN, and SVMs by an average RMSE of $23.42 per 1 Bitcoin.…”

“…The main limitation of these previously mentioned studies [2,6,11,17,23,25] is that all tweets and posts are equally treated without considering the level of interactions (likes, comments, and shares) that can significantly influence the impact of a particular post on the overall sentiments. Another limitation is the focus on only online communities and social media posts as a fluctuation factor despite the fact that other factors are proven to be as crucial as sentiments to the fluctuation changes [16].…”

“…Inspired by the work of Chevallier et al [6] and Sun et al [29], this paper proposes to forecast cryptocurrency price in the next day using taking advantage of all indicating factors (RQ1). To tackle RQ2, we propose an ensemble learning approach that adopts adaptive boosting and LSTM architectures.…”

“…All these characteristics and others make the forecasting of the fiat-crypto currency exchange rate 1 https://gadgets.ndtv.com/cryptocurrency/news/bitcoin-price-btc-cryptocurrencymarket-crash-usd-1-trillion-coinmarketcap-2726233 2 https://www.globenewswire.com/news-release/2021/04/12/2208331/0/en/At-30-CAGR-CryptoCurrency-Market-Cap-Size-Value-Surges-to-Record-5-190-62-Millionby-2026-Says-Facts-Factors.html (or simply cryptocurrency price forecasting) extremely challenging and often inaccurate [2]. Therefore, financial and AI experts are constantly analysing the market to better understand the trends of price fluctuations [1,2,5,6,13,16,17].…”