Development of Computational Intelligence on Cryptocurrency Forecasting

Cryptocurrency is a digital or virtual money that is difficult to fabricate or replicate because it is safeguarded by encryption. Blockchain technology is used by several cryptocurrencies to establish decentralized networks. Cryptocurrencies vary from traditional currencies in that they are not issued by a centralized entity, possibly making them resistant to government interference. The term “crypto” refers to the encryption and encryption techniques used to protect these entries, such as elliptic curve encryption, public-private key pairs, and hashing algorithms.

Forecasting is widely used in a wide range of sectors and enterprises as a tool for effective and efficient planning. Forecasting is a difficult task, and those that succeed at it will have a substantial advantage. Finding acceptable methodologies and models for estimating the values of cryptocurrencies is an important and difficult challenge due to their great volatility and the lack of legal control over their activities, which has resulted in major dangers connected with investing in crypto assets.

In 2022, a substantial amount of research has been conducted in trying to anticipate cryptocurrency prices. Many various factors have been studied in order to predict cryptocurrency prices using multiple methodologies and procedures in Machine Learning, Deep Learning, and Optimization algorithms.

Backpropagation neural network prediction for cryptocurrency bitcoin prices (Sovia, et al.)

The background of this paper is the difficulty in predicting the price of a crypto coin, especially the price of Bitcoin. The changes in the price of bitcoin are influenced by many things such as the closing of the bitcoin market in a country, the occurrence of hacker attacks on the bitcoin blockchain and the emergence of new coins that use technology like Bitcoin. This paper tries to implement an artificial neural network using backpropagation method to predict the price of Bitcoin by giving a form of predictive results that are strengthened with a fairly good value of accuracy. This paper uses 5 parameters, such as open price, higher price, lower price, volume of Bitcoin, and request Bitcoin. This paper uses Bitcoin data at 26 April 2018. After calculating thevartificial neural network with backpropagation method, the process will be at testing stage of the prediction process using MATLAB software. The result is the prediction price of Bitcoin is very close to the actual price that will be happen in the next day.

Bitcoin price forecasting with neuro-fuzzy techniques (Atsalakis, G. S., et al.)

This paper tries to forecast the price of Bitcoin using a computational intelligence technique that use a hybrid Neuro-Fuzzy controller, namely PATSOS, to forecast the direction in the change of the daily price of Bitcoin. This paper uses daily historical time series data of Bitcoin closing prices as raw inputs to the model, to forecast the direction in the change of the price. The data are from September 13, 2011 to October 12, 2017, a total of 2201 observations. The data were divided into two subsets: a traning set and a validation set, which the proportion of training set is 97% (2141 observations) and the validation set is 3% (60 observations). The result is using PATSOS trading strategies will have a Rate of Return (ROR) for 37.34%, compared to B&H Strategy only have 21.81% of ROR. The performance difference between PATSOS and B&H is 71.21%, which means PATSOS trading strategy is more useful for trading a Bitcoin.

Forecasting cryptocurrency price using convolutional neural networks with weighted and attentive memory channels (Zhang et. al, 2021)

This paper tries to forecast six cryptocurrencies’ values by following three steps: data pre- processing, weighted & attentive memory channel model, and training process using CNN. The data is obtained from CoinMarketCap from July 2017 until July 2020 that contains 1,089 samples. The model was trained by using Keras and Adam with a learning rate of 0.001. The dataset ratio for training, validating, and test is 60:20:20. It uses 50 – 100 epochs. The result was tested by using RMSE, MAE, and MAPE to evaluate error and R-Squared to evaluate accuracy. The proposed method only requires a short time window and a few layers to achieve high prediction accuracy. This model also reduces forecasting errors with various training ratios. It also shows a promising accuracy for predicting whether the price will increase and reduce.

Using Genetic Algorithm and NARX Neural Network to Forecast Daily Bitcoin Price (Han et. al, 2019)

This paper tries to suggest a daily bitcoin return model using a genetic algorithm and NARX neural network. It generates 10,000 numbers to produce a simulated data series, after that, the data were divided into ratio of 70:15:15 for training, validating, and testing. For GA, the length of chromosome was 10 bits for input, 2 bits for the number of hidden layers, and 4 bits for number of the neurons per layer then use sigmoid function as a transfer function. The initial population is 100 and generate 100 binary numbers with 12 bits at random, the crossover rate is 0.9 and mutation rate of 0.1. At the end, the mean squared error for predicting by using this model is equal to 0.00.142. The model also has some limitations, such as it can predict jump of the data effectively.

An Optimized Support Vector Machine (SVM) based on Particle Swarm Optimization (PSO) for Cryptocurrency Forecasting (Hitam et. al, 2019)

This paper introduced an improved Support Vector Machine (SVM) based on Particle Swarm Optimization (PSO) for forecasting the future price of cryptocurrencies. PSO is used to optimize the SVM algorithms in cryptocurrency predictions. Particle Swarm Optimization (PSO) is known as a superior algorithm for a static and basic optimization issue, and it is stated to be easier to adapt than other methods such as GA with few parameter adjustments. The dataset utilized was compiled over five (5) years of daily price for all six types of cryptocurrencies from 2013 to 2018. The experimental results show that an optimized SVM-PSO algorithm can effectively forecast the future price of cryptocurrency, outperforming single SVM algorithms with the best performance accuracy of 97 percent on Ethereum accurate forecasting, but the outcome is also dependent on the population and quality of training dataset.


Ghani Rizky Naufal – 2201799404 Gregoryus Imannuel Perdana – 2201755283 Muhammad Donny Devanda – 2201829621

Antoni Wibowo