1. Introduction
Artificial intelligence (AI) is the simulation of human intelligence by computers. Misheva et al. [
1] underline that AI offers great opportunities for enhancing the customer experience, democratizing financial services, ensure consumer protection and significantly improve risk management. In this field, Bussmann et al. [
2] argue that Artificial Intelligence models can be used in credit risk management and, in particular, in measuring the risks that arise when credit is borrowed employing peer to peer lending platforms. Islam et al. [
3] recall that a fundamental challenge for A.I.-based prediction models is the extent to which the internal working mechanisms of an AI system can be explained in human terms.
Machine learning (ML) is a branch of AI where algorithms are used to learn from data to make future decisions or predictions. Naturally, forecasting research stands at the forefront of this blooming literature. Cohen [
4] uses particle swarm optimization and identifies that both Darvas Box and Linear Regression techniques can help traders predict the bitcoin’s price trends. Besides, Li et al. [
5] demonstrate that the Attentive LSTM network and an Embedding Network achieve superior state-of-the-art performance among all baselines for the Bitcoin price fluctuation prediction problem. Last but not least, Livieris et al. [
6] utilizes as inputs different cryptocurrency data and handles them independently to exploit helpful information from each cryptocurrency separately, which leads to better results than the traditional fully-connected deep neural networks.
At the crossroads between econometrics and machine learning, we find a paper by Chen et al. [
7], who assesses that “
the machine learning approach could be a more suitable methodology than traditional statistics for predicting the Bitcoin price”. In this spirit, we attempt to infer whether such computational methods designed to “learn” information directly from data (and adaptively improve their performance as the number of samples increases) will lead to successful Bitcoin price forecasts. While econometrics assesses potential future scenarios using advanced statistical methods (such as time series), machine learning utilizes artificial intelligence to predict behavior in new ways. Of particular interest to us is that there is still a debate on why a given algorithm can outperform conventional methods in predictive analytics.
Regarding the issue of the use of “black-boxes”, there may be fundamental reasons for using them with the suspicion that goes beyond the warnings of Zhao and Hastie [
8]. For products that would lend themselves to the use of IA/ML, in practice, these tools are rarely applied for two reasons:
Readability: For important investment choices or setting up an investment process, we cannot simply rely on a model. In most cases, it is necessary to have a specific thesis, which must be explained based on simple econometric relations, whether for investors or the CEO of a fund, rarely a specialist. From this point of view, an AR(1) model which would link the BTC to two or three indices could be, from the point of view of investment strategy, more important than a better, more complex model, because it would allow us to better explain and then justify overtime to investors what their money is used for.
Far out-of-sample robustness: In setting up complex strategies comparable to derivatives, valuations and measurements of risks are generally based on the simulation of the dynamics of an underlying process (an economic driver of the value of various products). In this case, the input simulations (in addition, calibrated in neutral risk) will generally go well beyond what has been observed in the past, and we have no idea of the relevance of the “black-box” model that will come out. It is then preferable to have a precisely specified model for which the behavior in these out-of-sample areas has been consciously established.
In economics, recent developments of machine learning can be found, for instance, in Farrell et al. [
9] who developed semiparametric inference for deep neural networks. In finance, algorithms for quants are already thoroughly documented and accessible, for instance, in the book by de Prado [
10].
This paper assesses the risks of machine learning processes as “black-box” (already built) models by detailing precisely the parameters’ choices at each step, see also Zhao and Hastie [
8], Abadie and Kasy [
11]. Adopting a prudent approach towards building sparse models (see the survey on over-fitting and regularization methods by Athey and Imbens [
12]), we select six classes of machine learning algorithms: regularization (Ridge regression), neural networks (Multilayer Perceptron with Back-Propagation), ensemble decision trees (Random Forest, AdaBoost), instance-based (
k-Nearest Neighbour), and classification (Support Vector Machine).
Using these six machine learning processes, we analyze their fitting and predictive power through an empirical application based on Bitcoin spot and futures prices. Along with the paper, we discuss some risks associated with this approach. The paper discusses the potential prediction of the machine learning processes trying to answer the following points:
Do drivers exist for Bitcoin forecasts (inside the variables we retain)?
Can we accurately produce forecasts (models in question, and how to compare them)?
Is it possible to propose robust trading strategies?
Bitcoin is an electronic currency based on a vast peer-to-peer network, totally decentralized. New bitcoins are introduced to the market via a process called mining. The miners receive rewards as soon as they validate recent transactions after solving an optimization problem using a Proof of Work, which needs intensive computation. The first Bitcoin was created in 2009 (Nakamoto [
13]). A cryptocurrency can be defined as a digital asset designed to work as a medium of exchange using cryptography to secure the transactions and to control the creation of additional units of the currency. Since the origin of this cryptocurrency, we have observed high volatility of its price and specific features, which could be an interesting feature to understand, in the context of investment objectives. In February 2021, Bitcoin hit a market capitalization of
$1 trillion (all digital coins combined have a market cap of around
$1.7 trillion, according to Reuters [
14]). By hitting the
$1 trillion market cap, the Bitcoin market is gaining acceptance among mainstream investors and companies, from Tesla and Mastercard to the bank BNY Mellon. (See Reuters (2021) at
https://www.cnbc.com/2021/02/19/bitcoin-hits-1-trillion-in-market-value-as-cryptocurrency-surge-continues.html, accessed 19 February 2021).
Investment managers are doubtful with respect to the forecastability of Bitcoin, much like currency forecasters. For instance, the Chief Investment Officer of Citi Private Bank, David Bailin, reminds that “
to get an exposure to Bitcoin, if you do not own the actual Bitcoin, any such fund or structure can be a very, very inefficient way to do that”. (See Yahoo Finance (2021) at
https://autos.yahoo.com/unstoppable-trends-better-bitcoin-long-142808181.html, accessed on 17 March 2021). Bitcoin’s private key custody problem has essentially three practical solutions: (i) kept on an exchange, it constitutes a “honey pot” for hackers (recall the Magic The Gathering Online eXchange (Mt.Gox) where 650,000 BTC were lost); (ii) kept in banks, it undermines the 21 million Bitcoins scarcity by paving the way for securitization all over again (recall the 2008 sub-primes crisis); (iii) kept on a hardware wallet (such as Trezor or Ledger), there is the risk of theft and physical harm (Ledger’s marketing database was famously hacked, containing the clients’ private addresses). Therefore, no solution appears satisfactory. That is why investment in cryptocurrencies will be a (small) part of the “opportunistic side” of the client’s portfolio. According to Amundi’s asset managers Vincent Mortier and Didier Borowski, Bitcoin and other cryptocurrencies do not possess the intrinsic qualities of money, i.e., to be a metric unit, a store of value, and a medium of exchange. They do not have any real economic underlying, and there exists no pricing model. Both asset managers are wary of the speculative nature of cryptocurrencies. (See Amundi (2021) at
https://research-center.amundi.com/article/crypto-currencies-bubble-or-emergence-new-paradigm-decentralised-finance, accessed on 24 March 2021).
If we investigate the literature on Bitcoin, we observe a considerable amount of papers on this cryptocurrency recently to predict its price or the associated return or to determine the trend of these two quantities. Much literature focuses on the prediction of the volatility for this cryptocurrency. Nearly all the models existing in the linear and non-linear time series have been applied to predict prices or volatility. We give, in the next section, a summary of the more recent papers. Nevertheless, an interesting question remains: is it possible to predict the price of this cryptocurrency, whatever the model used and the period considered? This opens the question of the validity of the conclusions of all these papers. If everything ‘works’ even if the methods are antinomic, what is the robustness of these predictions?
Motivated by the growth of the Bitcoin market and the recent interest of market participants (for instance, in February 2021 alone, corporate adoption of BTC involved ARK Invest, Blackrock, BNY Mellon, Mastercard, Microstrategy, Square. BlackRock, the world’s largest asset manager “started to dabble” in BTC. Tesla invested
$1.5B in BTC and announced plans to accept crypto payments. North America’s first Bitcoin Exchange-Traded Fund (ETF), the Canadian-based Purpose Bitcoin, amassed
$421M in Asset Under Management in its first two days of trading) and academics, this study focuses on machine learning modeling. We illustrate some features that could explain the dynamic behavior of Bitcoin’s price by taking into account the non-stationary behavior of the data in place of classical parametric modelings (ARMA, related-GARCH, VAR modelings). For recent extensions in econometrics, see for instance Abedifar et al. [
15], Ahelegbey et al. [
16] regarding correlation networks, Billio et al. [
17] for multivariate models such as Granger Causality, or Baumöhl [
18] for connectedness
à la Diebold and Yilmaz approaches for crypto-assets and exchanges (Dahir et al. [
19], Le et al. [
20], Mensi et al. [
21] and further papers). For recent literature on forecasting non-stationary time series based on machine learning, see, e.g., Cao and Gu [
22], Kurbatsky et al. [
23], Wang and Han [
24]. The interest of machine learning is that the notion of non-stationarity is not crucial as in econometric models for which we need to have stationarity to be sure to have a solution (because this corresponds to the assumptions imposed by econometric models), which is irrelevant in ML. In that sense, machine learning could be a promising technology. We know that it has been used and advanced for asset price/return prediction in recent years since the financial time series are non-stationary and volatile. The development of machine learning and its interest in finance is not new. A seminal paper on the introduction of this methodology was given by Rosenblatt [
25], and more recent developments can be found in Russell and Norvig [
26] with a lot of references therein. For applications using financial assets, the current paper of Iworiso and Vrontos [
27] provides evidence that machine learning techniques permit us to get exciting results concerning the forecasts of the direction of the U.S. equity premium.
In the spirit of the previously-cited papers, in the present paper, we analyze the behavior of the Bitcoin cryptocurrency and its futures with a class of machine learning techniques. We investigate its behavior in the future. As soon as Bitcoin is used for diversification in portfolios, we complete our analysis by looking at Bitcoin’s interaction with stocks, commodities, bonds, and other cryptocurrencies. Our research based on these data analytics techniques focuses on (i) their capability to fit a data set. We observe that the Adaboost method and the random forest processes are the winners inside a competition based on six competitors. (ii) Regarding their predictability power, we observe a high variability of the results depending on the period on which we work and the input data used for the training. Thus, the question that emerges from this work is the possibility of predicting the spot or the future for this cryptocurrency against luck or uncertainty, without obviously calling into question the methodology used.
Central banks convey this idea of Bitcoin being an extremely inefficient way to process transactions, highly speculative, and used mainly for the financing of illicit activities. In the view of the Treasury Secretary Ms. Janet Yellen, Central Banks Digital Currencies (CBDC) should be the only solution for printing digital money (through its proprietary core ledger). This view is largely echoed in finance journals. To cite a few, Foley et al. [
28] estimate that around
$76 billion worth of illegal activity per year involve Bitcoin (46% of bitcoin transactions), which is close to the scale of the U.S. and European markets for illegal drugs. Among other “Silk Roads” dismantled by the FBI regarding drug trafficking, the risks of “black e-commerce” are heightened by the anonymous file server Tor (The Onion Router), and by secret cryptocurrencies’ operational design such as Zcash or Monero.
Spanning daily data from 13 January 2015, to 31 December 2020, our analysis is based on several steps to analyze the main drivers of the Bitcoin currency. First, we look at the realm of seventeen cryptocurrencies. Second, as representative of traditional financial markets, we investigate the relationships of Bitcoin with eleven stocks, four bonds, and four foreign exchange markets. Third, we examine the interactions with four energy, seven metals, three grain commodities, five softs, and two cattles as an alternative investment class. The analysis is robust to Bitcoin spot or futures prices as the underlying asset. The novelty lies in (i) considering six machine learning models and one parametric model (an AR) in a horse race to forecast the price of Bitcoin, (ii) developing trading strategies issues to investigate the potential use of crypto assets in portfolio management. As robustness checks, we identify several sub-sample forecasts for results sensitivity purposes.
Regarding the central methodological and empirical contributions, our paper stresses the key ingredients to make a ‘good’ machine learning model in quantitative economics, a.k.a: (i) proceeding to an excellent data collection (our ‘financial markets’ approach as opposed to a ‘blockchain approach’ feeding the models with technical and non-stationary data) while controlling for low multi-collinearities; (ii) assessing using a wide array of visualization tools (clusters, maps, diagrams) the main finding of segmentation of Bitcoin concerning traditional financial and commodity markets (e.g., Bitcoin reacts mainly to the information content of other cryptocurrencies); and (iii) favoring either the AdaBoost or Random forest algorithms as predictors of the Bitcoin spot and futures prices, which allows us to implement trading strategies; and to open the debate on the forecasting accuracy of Bitcoin.
What is the accurate information set to predict Bitcoin prices? Our reply is largely that Bitcoin appears segmented to crypto-assets only, and not much connected to financial markets. Therefore, we do not deploy standard econometrics tests (with Granger causality or reverse causality). We follow a purely data-driven machine learning approach. In a nutshell, this paper contains the results of a set of prediction exercises. The critical emphasis is placed on the proper use of machine learning techniques (Artificial Neural Network, Support Vector Machine, Random Forest, k-Nearest Neighbours, AdaBoost, Ridge regression) to forecast daily movements of the price of Bitcoin. We demonstrate that the performance of such machine learning methods is highly dependent on several design choices (hyperparameters, optimizers, network topology). The forecast statistics retained are the Root Mean Square Error (RMSE), the Mean Absolute Error (MAE), and the Mean Absolute Percent, Error (MAPE). The paper concludes that, in this particular exercise, AdaBoost stands out as the best machine learning. The Random Forest algorithm also performs well among the six considered.
The remainder of the paper is organized as follows.
Section 2 summarizes some of the papers interested in predicting Bitcoin with different modelings, classifying these modelings concerning the models.
Section 3 describes, in a uniform way, all the machine learning models used.
Section 4 introduces the data.
Section 5 contains the results for the whole sample, distinguishing the results obtained with all the risk factors we have listed to explain the spot’s behavior and Bitcoin futures. This section proposes an in-depth analysis of the inter-relationships between Bitcoin, other cryptocurrencies, and the stable coin Tether, traditional asset markets, and alternative commodities.
Section 6 provides robustness checks along four sub-samples corresponding to different periods characterizing Bitcoin’s price behavior: a restricted sample to the newest cryptocurrencies, Tether’s introduction in 2017, the 2016–2018 Bitcoin economic cycle, the recent 2019 trend, and the 2020 “bull run”.
Section 7 provides some discussions and conclusions.
2. Background
The literature on Bitcoin pricing is developing in finance. Among various topics tackled, Easley et al. [
29] document the level of transaction fees on this particular market and assess that a high volume of transactions is required. Bitcoin mining is computationally intensive on the network, and a model calibration was achieved by Prat and Walter [
30] (including the electricity cost). Mining rewards are the main incentives for miners to invest in expensive mining pieces of equipment (e.g., dedicated GPU cards or ASIC miners). Hence, the motto “Get Rich or Die Mining” is often found on crypto forums). Another concern on this market is the ability of traders to benefit from price deviations that occur due to multiple trading places: this is called arbitrage between exchanges (Makarov and Schoar [
31]). In management, the focus is more on the “cryptocurrency mania” that risks leading to speculative bubbles, as in Cheng et al. [
32], Wei and Dukes [
33]. Financial practitioners are also concerned about the security of the blockchain (Pagnotta [
34]). Quantum computers are posing a serious challenge to the security of the Bitcoin blockchain indeed. (See Deloitte (2021) at
https://www2.deloitte.com/nl/nl/pages/innovatie/artikelen/quantum-computers-and-the-bitcoin-blockchain.html, accessed on 14 March 2021).
Several studies have been conducted in the literature concerning predicting Bitcoin spot price or the evolution of its volatility trend. We provide some references without being exhaustive.
Some are based on classical econometric modeling, including: (i) time-series techniques (e.g., vector autoregressive (VAR), vector error correction (VEC), quantile regression), for instance, see Fantazzini et al. [
35] and references therein. (ii) GARCH and DCC modeling: for example, Briere et al. [
36] investigate the volatility behavior of Bitcoin. Using the same models, Aslanidis et al. [
37] compare the volatility of different cryptocurrencies, including Monero. Caporale and Zekokh [
38] use Markov switching modelings to investigate the volatility of Bitcoin and other cryptocurrencies. (iii) Long memory and jump modeling: in several papers, authors try to detect a possible long memory behavior using different techniques, see, for instance, Bariviera et al. [
39], Alvarez-Ramirez et al. [
40], Begušić et al. [
41]. Some authors try to use this long memory behavior for trading strategies; see, for instance, Khuntia and Pattanayak [
42], Al-Yahyaee et al. [
43]. With the possible long memory behavior, models with jumps have been used to investigate both the returns’ behavior and the volatility. Some references are Phillip et al. [
44], Mensi et al. [
45]. There exists a large literature on the bubble behavior of Bitcoin, which has been observed since 2014. We can cite, among others, Su et al. [
46], Guegan and Frunza [
47], Geuder et al. [
48]. At the same time, looking at the evolution of the price on specific periods, authors try to show that Bitcoin can be considered a commodity (the idea is that this cryptocurrency corresponds to a limited resource), e.g., Guesmi et al. [
49], or as gold (Dyhrberg [
50]). In many cases, the authors are interested in discussing the potential (or not) of Bitcoin for diversification, (e.g., Polasik et al. [
51], Bouri et al. [
52], Selmi et al. [
53]).
Some papers use high-frequency data and are interested in shock transmission: using realized volatility of the cryptocurrencies, some authors detect asymmetries in shock transmissions between the cryptocurrencies and traditional assets, see, for instance, Kurka [
54] and references therein. The informational efficiency of Bitcoin has also been investigated using high-frequency in Zargar and Kumar [
55], extending some previous works on different papers whose references can be found in this last paper.
Some papers have investigated the cross-correlation between cryptocurrencies and different stocks and bonds using related GARCH and DCC modelings to use Bitcoin for diversification. For instance, in a recent paper, Aslanidis et al. [
37] detect that the correlation of traditional assets against Monero is even closer to zero than against other cryptocurrencies. Other papers investigate the correlation with different stocks, such as Fang et al. [
56], Gillaizeau et al. [
57], among others.
Sentiment analysis using Twitter and Google Trends forms another new tool to forecast Bitcoin prices. For instance, Wołk [
58] recently mobilized this computational tool to predict the prices of Bitcoin and other cryptocurrencies for different time intervals. The author highlights that people’s psychological and behavioral attitudes significantly impact the highly speculative cryptocurrency prices. Further, on informative signals derived from Twitter and Google Trends, Shen et al. [
59] find that the number of tweets is a significant driver of next-day trading Bitcoin volume. Philippas et al. [
60] identify that Bitcoin prices are partially driven by momentum on media attention in social networks, justifying a sentimental appetite for information demand. Guégan and Renault [
61] explore the relationship between investor sentiment on social media and intraday Bitcoin returns. The authors document a statistically significant relationship between investor sentiment and Bitcoin returns for frequencies of up to 15 min. The impact of news is further documented by Dey et al. [
62] regarding the use of chainlets to evaluate the role of the local topological structure of the blockchain on the joint Bitcoin and Litecoin price formation and dynamics, or by Nicola et al. [
63] regarding information theory measures extracted from a Gaussian Graphical Model constructed from daily stock time series of listed US banks.
Finally, machine learning modeling has recently been used to understand the behavior of cryptocurrencies. Atsalakis et al. [
64], Jang and Lee [
65], Mallqui and Fernandes [
66] investigate the direction prices for daily cryptocurrencies. Atsalakis et al. [
64] uses a hybrid Neuro-Fuzzy controller based on artificial neural networks for Bitcoin prices. Jang and Lee [
65], for the same data set, use a Bayesian neural network. Mallqui and Fernandes [
66] focus on Artificial Neural Networks (ANN), Support Vector Machines (SVM), and k-Means clustering method for Bitcoin predictions introducing other stocks in their study. On another side, Nakano et al. [
67] explore Bitcoin intraday technical trading strategies based on deep learning for the price direction return prediction (up and down) on the period of December 2017 January 2018. They provide interesting results on the role of the layers, outputs, and inputs for their trading strategies. Sun et al. [
68] adopt a novel Gradient Boosting Decision Tree (GBDT) algorithm, Light Gradient Boosting Machine (LightGBM), to forecast the price trend. Further on this, [
69] hierarchically cluster Bitcoin prices from different exchanges and classic assets by enriching the correlation-based minimum spanning tree method with a primary filtering method based on the random matrix approach. Using a stochastic neural network model, Jay et al. [
70] trained the Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) models for Bitcoin, Ethereum, and Litecoin. The results show that the proposed model is superior in comparison to the deterministic models.
Our paper is close to this last class of articles, with new and different findings.
7. Conclusions
Machine learning algorithms attempt to find natural patterns in data to enhance decision-making. Machine learning typically prescribes a vast collection of high-dimensional models attempting to predict quantities of interest to solve problems in computational finance while imposing regularization methods.
In a seminal paper, Zhao and Hastie [
8] warns us against the pitfalls of ‘black-box’ machine learning models and urge us to make proper use of (i) a good predictive model, (ii) a sound selection of the dataset, and (iii) visualization tools to ensure the quality of the research work.
This paper addresses the following central research questions: How far the is use of machine learning techniques useful to forecast daily movements of the price of Bitcoin? How strong is the relationship between cryptocurrencies and traditional financial assets? How to perform a trading strategy based on cryptocurrencies? We overcome the risks of machine learning in predicting Bitcoin spot and futures prices by documenting: (i) either the AdaBoost or Random forest algorithms perform as the ‘best’ machine learning models among the six considered and could be implemented in a banking institution’s internal computing system for Bitcoin forecastability. (ii) A financial market approach favors price relationships between asset classes (cryptocurrencies, stocks, bonds, foreign exchange, and commodities) for the dataset’s quality. (iii) We implement a myriad of visualization tools (e.g., Louvain clustering, self-organizing map, t-distributed SNE, Sieve diagrams, and multidimensional scaling) to strengthen our findings.
During the period under study, the key takeaway is that Bitcoin appears as a problematic asset to forecast, subject to frequent price swings, as highlighted by the latest end-of-year 2020 run-up. From our empirical results, Bitcoin appears segmented from traditional financial and commodity markets. It seems to react more to the information content stemming from other cryptocurrencies, enhancing the forecast accuracy of Bitcoin. Whatever the period in which we work (full period, sub-sample periods), the best result is achieved during the full period (lowest MAPE = 0.15).
Across the trading strategies, we have documented that (i) machine learning algorithms (configured as bots following buy/sell signals) do not teach how to trade, (ii) the buy-and-hold strategy appears the best, which incites owners of Bitcoin to ‘hodl’. Because of the variability of the forecasting results, it is necessary to let cool heads prevail before investing in Bitcoin with private individuals’ money. Wealthy clients from investment banks can have access to Bitcoin funds in special entities such as Morgan Stanley, for individuals with “an aggressive risk tolerance” and a net worth of
$2 million held by the firm. (See CNBC (2021) at
https://www.cnbc.com/2021/03/17/bitcoin-morgan-stanley-is-the-first-big-us-bank-to-offer-wealthy-clients-access-to-bitcoin-funds.html, accessed on 17 March 2021). Such clients are especially waiting for the approval of ETF funds by the U.S. SEC led by the new appointment of Mr. Gary Gensler, a professor in the digital economy from MIT. The SEC’s cryptocurrency commissioner, Ms. Hester Peirce (dubbed “crypto mom”), has long advocated a Bitcoin ETF. The argument behind the use of cryptocurrencies lies in the modernization of the financial system, whereby assets worth billion dollars can be transferred securely and quasi-instantly (compared to the SWIFT/IBAN system) between agents. Transaction fees are recorded to be minimal. It is possible to follow the transactions in real-time on
blockchair.com. For example, one timestamp 6 April 2021 13:22 (UTC) records
$1,783,170,000 exchanged through Bitcoin for a transaction fee of
$45.45. At the time of writing, only Canada has approved the opening of ETFs on Bitcoin (such as Purpose Investment, Evolve Funds Group, or CI Galaxy). (See Coindesk (2021) at
https://www.coindesk.com/third-bitcoin-etf-expected-to-launch-in-canada-this-week, accessed on 8 March 2021).
Overall, we believe that Bitcoin remains a difficult beast to tame for a modeler, and that the debate on its pricing and forecasting accuracy remains open. We wish to insist here on the speculative nature of investing in Bitcoin, which remains an artificial mechanism without an underlying (unless it can be considered that the proof-of-work algorithm is a receivable underlying in financial derivatives markets?). It remains hard to gauge the fundamental value of Bitcoin from an economist’s standpoint, hence the challenge attached to any attempt at forecasting.
As avenues for future research, we would like to stress that our method can be extended to alt-coins, with a specific interest obviously for Ethereum, Litecoin, or Ripple. (i.e., the “main” alt-coins, as listed on
https://coinmarketcap.com/, accessed on 27 May 2021). Having forecast the spot and futures prices of Bitcoin in this article, it might be interesting to look at the spreads between the two next. Indeed, with the “democratization” of ML among institutional investors, we can imagine that there will be more significant volumes of trading and arbitrage between the two markets, but the differences in performance between the two were due for a certain time. reflect distinct populations of participants.