Deep Learning Based Adaptive Multimodal System for Bitcoin Price Prediction

Abstract

Cryptocurrencies are the best example of Blockchain. And the Blockchain is a decentralized network where they are multiple nodes that control the network. Bitcoin established itself as the first decentralised cryptocurrency in 2009. Bitcoin revolutionized the crypto world and all other cryptocurrencies other than bitcoin such as Ethereum, Ripple, etc. are called altcoins. In the crypto market, the price of bitcoin reflects the other cryptocurrencies. Therefore, our intention is to create prediction models which uses deep learning concepts for bitcoin and to forecast its future price.

Introduction

I. INTRODUCTION

Investors and scholars find market price forecasting fascinating and difficult due to its complexity. There are a lot of unknowns and things to think about. The market may be affected by economic conditions and other reasons.  [1] As a result of recent political developments, the market has grown not only more competitive but also more diversified. When it comes to the stock market, foreign exchange (FX), and cryptocurrencies, ODO defines cryptocurrency as a digital currency. A cryptocurrency is a form of money that is governed by encryption algorithms. The establishment of money units and the verification of currency transfers, that is not subject to central bank supervision. The prediction of future prices is one of the key objectives of bitcoin analytics. The price dynamics are influenced by a variety of things. The supply-demand relationship, investor appeal, financial and macroeconomic information, technical indicators like difficulty, the number of freshly generated blocks, and so on are the most crucial elements. Social media and search engine trends have a big impact on bitcoin values. In order to forecast the future price of Bitcoin, we are constructing the ARIMA and FBProhpet models in this work. The goal of this research is to create a model of a neural network that can be used to forecast Bitcoin price movement over the long and short terms. This research is inspired by work done in the stock market price prediction field, which achieves higher accuracy than the cryptocurrency price prediction field by predicting the long-term price.

II. LITERATURE REVIEW MATERIALS AND METHODOLOGY

As traders and investors are anxious to know the price movement of bitcoin in the future, bitcoin price detection is one of the most important aspects today. So, there are several models that have been designed to forecast the price of bitcoin in the future.

This study uses data from daily time series, 10-minute intervals, and 10-second intervals to forecast the price of bitcoin. They created three sets of time series data with intervals of 30, 60, and 120 minutes, and then generated three linear models from the datasets using GLM/Random Forest.

The price of Bitcoin is determined by linearly combining these three models. [2] states that the author is investigating efforts made to forecast the American stock market. At last the result he got that the excess return standard deviation was far more than the mean square error of the predicted network. However, the author demonstrates how a number of fundamental financial and economic factors can forecast the result.

With a 55 percent accuracy rate, Reference [4] used SVM and ANN to forecast the price of Bitcoin using data from the Bitcoin blockchain. In order to predict the short-term Bitcoin price, Random Forest, SVM, and Binomial Logistic algorithms are utilised. This research has one flaw, according to [6]: the outcome was not cross checked, therefore it may have overfit the data and one cannot be sure if the model would generalise.

As a result, [6] uses an LSTM network and accomplishes an accuracy rate of 52%. The majority of prior attempts at predicting the price of cryptocurrencies have estimated short-term Bitcoin values with low accuracy and without cross-validation, endangering traders' confidence in the model.

With Multilayer Perceptron having the most accuracy when projecting the next 60-day price change and Recurrent Neural Networks having the best accuracy when forecasting the following 56-day price change, long-term forecasting triumphs over short-term forecasting with accuracy of 81.3 percent, precision of 81 percent, and recall of 94.7 percent, Multilayer Perceptron surpasses Recurrent Neural Networks. The variations in bitcoin prices based on execution orders like purchase or sell were investigated by Tian et al. [1]. Moving average values and regression approaches were discussed. They developed a time series model that projected bitcoin values using a Gaussian time model. They did, however, show that their model works well with time series data. We tested our proposed model based on the current price of bitcoin using a dataset spanning several years.

A. Arima

The autoregressive and moving average models are working together as a single unit to generate the ARIMA model. This model has been widely used and tested for different kinds of time series. Since seasonal time series make up the majority of data from the actual world, modelling of both seasonal and non-seasonal series must be investigated. In ARIMA, we find the difference between time series from one time stamp to another rather than predicting the time series.

Multiplicative model is used to describe the seasonal time series, which is defined as follows: where x is a periodic term,, and B is the difference operator specified as

Since B(Zt) = Zt-1, (1-Bx)D

D stands for the x seasonal difference in Dth.

(1-B) The seasonal moving average model's order is P, the non-seasonal moving average model's order is Q.The corresponding positions of the AR and MA parameters in the ARIMA model are detected using the Auto Correlation Function (ACF) and Partial Auto Correlation Function (PACF) curves.

B. FBProphet

Prophet is a method for forecasting time series data that matches non-linear patterns with seasonality on a yearly, monthly, and daily scale as well as the effects of holidays. It works well with historical information from different seasons and time series with sizable seasonal effects. Prophet typically does a good job of handling outliers and is tolerant of missing data and trend shifts. Prophet is a popular Facebook app that generates precise estimates for planning and goal-setting. We've discovered that it performs better than any alternative method in the vast majority of situations. We use Stan to fit models so that you may get forecasts in a matter of seconds.
An open-source time-series model creation method called Facebook Prophet combines some tried-and-true ideas with some fresh twists. It avoids some of the drawbacks of other techniques and excels at modelling time series with a variety of seasonalities. The total of the three-time functions plus an error term is growth g(t), seasonality s(t), holidays h(t), and error e(t).

V. ACKNOWLEDGEMENTS

 We would especially want to thank our guide S. G. Nagaraju Valluri and co-ordinator Dr. T. Rama Swamy for providing us with the chance to complete a fantastic project on this subject. It forces us to conduct extensive research and pick up new knowledge. We are very appreciative of that. Additionally, we would like to thank my friends who greatly contributed to the timely completion of this project.

Conclusion

In this study, we looked at how well the ARIMA Model and FBProphet models predicted fluctuations in the price of bitcoin. Additionally, these models are used to forecast the price of bitcoin in the future. By combining many features into a single feature and removing extraneous features, feature engineering can be used to increase the learning speed and accuracy of neural networks. Among the hyperparameters that can be changed are batch size, dropout rate, and weight initialization. We are also interested in investigating more complex deep learning models, including architectural variants of temporal convolutional neural networks and sequence to sequence models.

References

[1] L.Khaidem, S. Saha and S. R. Dey, \"Predicting the direction of stock market prices using Random Forest\", arXiv preprint arXiv:1605.00003, 2016 [2] Mohammad Ali, Swakkhar Shatabda, \"A Data Selection Methodology to Train Linear Regression Model to Predict Bitcoin Price\", Advanced Information and Communication Technology (ICAICT) 2020 2nd International Conference on, pp. 330-335, 2020. [3] Busra KUTLU KARABIYIK, Zeliha CAN ERGÜN, \"FORECASTING BITCOIN PRICES WITH THE ANFIS MODEL\", Dicle Üniversitesi ?ktisadi ve ?dari Bilimler Fakültesi Dergisi, vol. 11, pp. 295, 2021. [4] K.L. Varma, K.S. Reddy, S. Jancy and M.P. Selvan, \"Classification and Prediction of Text Data by Using a Natural Language Processing Algorithm\", Lecture Notes in Electrical Engineering, vol. 691, pp. 635-641, 2021. [5] S. Prince Mary, B. Bharathi, S. Vigneshwari and R. Sathyabama, \"Neural computation based general disease prediction model\", International Journal of Recent Technology and Engineering, vol. 8, no. 2, pp. 5646-5649, 2019. [6] Sankar Phani, P. Anil Kumar, P. Bharathi and B., \"Blockchain-Based Incentive Announcement In Vanet Using CreditCoin\", Lecture Notes in Electrical Engineering, vol. 709, pp. 567-574, 2021. [7] Sankar Phani, P. Anil Kumar, P. Bharathi and B., \"Blockchain-Based Incentive Announcement In Vanet Using CreditCoin\", Lecture Notes in Electrical Engineering, vol. 709, pp. 567-574, 2021. [8] M.P. Selvan, N. Navadurga and N.L. Prasanna, \"An efficient model for predicting student dropout using data mining and machine learning techniques\", International Journal of Innovative Technology and Exploring Engineering, vol. 8, no. 9, pp. 750-752, 2019. [9] H. Chebrolu, S. Buragadda, Viji Amutha Mary and A. Jancy, \"Blockchain Integration in House Voting System\", Lecture Notes in Electrical Engineering, vol. 691, pp. 529-537, 2021. [10] Chowdary Ajay, M. Kundan, M., Viji Amutha and A. Mary, \"Effective Credit Card Forgery Prevention Using Multilevel Authentication\", IOP Conference Series: Materials Science and Engineering, vol. 590, no. 1, pp. 012021, 2019.

Copyright

Copyright © 2023 A Rajesh Kumar, S Madhu, A Haripriya. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.