As the worst days in terms of governmental bans and elevated volatility of the crypto universe seem to be over – at least, for now – it’s time to start rebuilding crypto wealth. The total market cap of all coins is now close to $2 trillion, whereas the total capitalization of all stocks approaches $50 trillion. Assuming zero fiat money expansion going forward (something that is hard to believe, but still), theoretically, the maximum upside for an imaginary broad-based crypto index is around 25X. In theory that translates into Bitcoin at $1,500,000 apiece. But this is just a theoretical idealistic model, while in order to assess the real upside potential, we must assume that there will be a substantial stable residual capitalization in the stock market – let’s suppose, in the distant future, the crypto and stock market funds will be split equally. Not that we are predicting it – rather we make a necessary hypothesis.

Also, the main bolster behind most recent episodes of rapid Bitcoin and other cryptos appreciation is institutional money inflows. During these periods a crypto portfolio optimization (AKA seeking an efficient frontier) is fully warranted. Such an optimization enables an investor to maximize his or her exposure to growth momentum instead of sticking to less efficient discretionary holdings. When the inflow period ends, it is advisable that all altcoins investments are cashed and parked back in Bitcoin or Ethereum.

Understanding money flows is key for successful crypto investing. In addition, just as much as interval investing maximizes returns in stocks nowadays – same true for crypto investments. Periods of growth, like one we are facing now, must be used to seek efficient frontiers – whether by means of iterating well-known optimal Sharpe and/or Sortino ratios – or by applying various existing price forecasting models like one we discussed previously – namely, the stock-to-flow pricing model. This time around, let’s discuss two of the most interesting machine methods.

Pic.1. Example of an Excel-built Monte Carlo Simulation model involving a crypto portfolio


Several Google-searched essays in this area cover work done on Bitcoin (BTC) price prediction using different techniques and evaluation of recurrent neural network (RNN) and its system architecture. Among the most frequently cited applied machine learning methods are artificial neural network (ANN), support vector machine (SVM), and recurrent neural network (RNN) as well as k-means clustering in an asset price prediction. However, one limitation of these studies is only focused on the investors. Policy-makers can be considered as either a major partner of the system because cryptocurrency can change the dynamics of the world economy, or a major drag because of their increasingly enthusiastic regulation aspirations.

The proposed methodology considers two different deep learning-based prediction models to forecast the daily price of Bitcoin by identifying and evaluating relevant features by the model itself.

In order to better understand the basics of the machine learning methods, let’s consider two of the most widely known methods of technical analysis, RSI and MACD. The Relative strength index (RSI) charts the speed and scale of directional changes in values. The RSI has different values that determine trend behavior. It measures the velocity and magnitude of the changes in recent prices to determine overbought or oversold levels of the prices of the securities. If the value of RSI is lower than 30, then the algorithm labels the input signal as +1, higher than 70 is labeled as −1. For values between 30–70, if the value of RSI at time “t” is higher than the value at time “t−1”, the trend is “upward” and labeled as +1, and vice versa.

Moving average convergence/divergence (MACD) indicator, in its turn, is related to movements of prices. The MACD shows the relationships between two moving averages of the securities’ prices. It is calculated by using the differences of short and long Exponential Moving Averages (EMA). If the MACD increases, then prices increase and if the MACD decreases, then prices decrease. If the value of MACD at time “t” is greater than the value at time “t−1”, the trend is “upward” and labeled as +1, and vice-a-versa.

Machine Learning methods mean the creation of automatic constant checking of the calculated vs. actual data and inputting a correction signal to narrow down divergences over the testing period. A recurrent neural network (RNN) is a deep neural network characterized as a recurrent connection between the input and output of its neurons or layers and capable of learning sequences designed to capture temporal contextual information along time series data. RNN is a class of artificial neural networks where connections between nodes form a directed graph along a temporal sequence. This allows it to exhibit temporal dynamic behavior. Also, RNNs recognize data’s sequential characteristics and use patterns to predict the next likely scenario.

Pic.2. Schematics of RNN

These networks are also distinct from other types of artificial neural networks because they use feedback loops to process a sequence of data that informs the final output. These feedback loops allow information to persist. This effect often is described as memory. A compelling experiment involves an RNN trained with the works of Shakespeare to produce Shakespeare-like prose successfully. Two common RNN networks are LSTM and GRU and are presented in the subsequent sections.

While traditional deep neural networks assume that inputs and outputs are independent of each other, the output of recurrent neural networks depends on the prior elements within the sequence. While future events would also be helpful in determining the output of a given sequence, unidirectional recurrent neural networks cannot account for these events in their predictions.

So these models are more like statistical extrapolation studies, whereas in the case of the previously described stock-to-flow model we dealt with causative problem-solving. These approaches are fundamentally different, but being combined, they produce outstanding results!

Pic.3. Machine-Learning Output Example vs. Real Market Data


One solution to the problem is called long short-term memory (LSTM) networks, which computer scientists Sepp Hochreiter and Jurgen Schmidhuber invented in 1997. RNNs built with LSTM units categorize data into short-term and long-term memory cells, which is very important while separating Bitcoin’s short, incidental or seasonal, patterns from their long-term trends.

LSTMs are explicitly designed to avoid the long-term dependency problem. Remembering information for long periods of time is practically their default behavior, not something they struggle to learn.

By using an LSTM and a GRU together, networks can take advantage of the strengths of both units – the ability to learn long-term associations for the LSTM and the ability to learn from short-term patterns for the GRU.

Final Remark

Currently all the discussed Bitcoin price predicting models are pointing at BTC price reaching proximities of $100K apiece before entering a meaningful correction.


All the information contained on our website is published in good faith and for general information purposes only. Any action the reader takes upon the information found on our website is strictly at their own risk.