International Conference on Advanced Communications Technology(ICACT) 144

Bitcoin Price Prediction using Machine Learning

Siddhi Velankar*, Sakshi Valecha*, Shreya Maji*
*Department of Electronics & Telecommunication, Pune Institute of Computer Technology, Pune, Maharashtra, India
[email protected], [email protected], [email protected]

Abstract— In this paper, we attempt to predict the Bitcoin price

accurately taking into consideration various parameters that II. LITERAURE SURVEY
affect the Bitcoin value. For the first phase of our investigation, Bitcoin is a new technology hence currently there are few
we aim to understand and identify daily trends in the Bitcoin
price prediction models available. [1] deals with daily time
market while gaining insight into optimal features surrounding
Bitcoin price. Our data set consists of various features relating to series data, 10-minute and 10-second time-interval data. They
the Bitcoin price and payment network over the course of five have created three time series data sets for 30, 60 and 120
years, recorded daily. For the second phase of our investigation, minutes followed by performing GLM/Random Forest on the
using the available information, we will predict the sign of the datasets which produces three linear models. These three
daily price change with highest possible accuracy. models are linearly combined to predict the price of Bitcoin.
According to [2] the author is analysing what has been done to
Keywords— Bayesian regression, Bitcoin, Bitcoin prediction, predict the U.S. stock market. The conclusion of his work is
Blockchain, crypto currency, generalized linear model (GLM), the mean square error of the prediction network was as large
machine learning. as the standard deviation of the excess return. However, the
author is providing evidence that several basic financial and
I. INTRODUCTION economic factors have predictive power for the market excess
return.
A. Bitcoin: In [3], instead of directly forecasting the future price of the
Bitcoin is a crypto currency which is used worldwide for stock, the authors predict trend of the stock. The trend can be
digital payment or simply for investment purposes. Bitcoin is considered as a pattern. They perform both short term
decentralized i.e. it is not owned by anyone. Transactions predictions (day or week predictions) and also long-term
made by Bitcoins are easy as they are not tied to any country. predictions (months). They found that the latter produced
Investment can be done through various marketplaces known better results with 79% accuracy. Another interesting
as “bitcoin exchanges”. These allow people to sell/buy approach the paper reflects is the performance evaluation
Bitcoins using different currencies. The largest Bitcoin criteria of the network. Based on the predicted output the
exchange is Mt Gox. Bitcoins are stored in a digital wallet performance evaluation algorithm decides to either buy, sell or
which is basically like a virtual bank account. The record of hold the stock.
all the transactions, the timestamp data is stored in a place From [4], a comparison between Multi-Layer Perceptron
called Blockchain. Each record in a blockchain is called a (MLP) and Non-linear autoregressive exogenous (NARX)
block. Each block contains a pointer to a previous block of model is made. They conclude that MLP can also be used for
data. The data on blockchain is encrypted. During transactions stock market prediction even though it does not outperform
the user’s name is not revealed, but only their wallet ID is NARX model in price prediction. The authors made use of
made public. MATLAB’s neural network toolbox to build and evaluate the
performance of the network.
B. Prediction:
The Bitcoin’s value varies just like a stock albeit III. FLOW OF PAPER
differently. There are a number of algorithms used on stock
market data for price prediction. However, the parameters The first part of the paper is database collection. We have
affecting Bitcoin are different. Therefore it is necessary to acquired bitcoin values from two different databases namely:
predict the value of Bitcoin so that correct investment Quandl and CoinmarketCap. After acquiring this time-series
decisions can be made. The price of Bitcoin does not depend data recorded daily for five years at different time instances, it
on the business events or intervening government unlike stock must be normalized and smoothened. For this, we have
market. Thus, to predict the value we feel it is necessary to implemented different normalization techniques like log
leverage machine learning technology to predict the price of transformation, z-score normalization, boxcox normalization,
Bitcoin. and so on. After this, data is smoothened over the complete
time period.

ISBN 979-11-88428-01-4 ICACT2018 February 11 ~ 14, 2018

 International Conference on Advanced Communications Technology(ICACT) 145

The next step is to select parameters that will be fed to the 1) Log Normalization: In this method, the range is compressed
predictive network. From an array of available features, some and we get the values that were close to zero before
are mentioned below: normalization. The function is: A'= log(A)/log(max)

TABLE 1. BITCOIN FEATURES AND THEIR EQUATIONS 2) In built MATLAB method: The function used is 'normc' to
normalize database columns. It compresses the range to the best
S.no. Features Equations/Definitions possible extent as compared to other methods.
1. Block Size Average block size in MB
2. Total bitcoins Total number of bitcoins mined
3) Standard deviation normalization: Here, we take into
3. Day high, day Highest and lowest values of
consideration the difference of every value with respect to the
low different days
mean value. The advantage of this technique is that we get the
4. Number of Total number of unique Bitcoin
negative values as well due to proper compression of the Y axis.
transactions transactions per day
The formula is z = (x – μ) / σ.
5. Trade Volume USD trade volumes from the top
exchanges
4) Z score normalization: This method uses technique similar to
After feature selection, the sample inputs will be fed to the standard deviation method by considering the mean value.
model. The variation in the bitcoin values can be considered
as a pattern. The pattern can be either going up, down or 5) Boxcox normalization: The function used is:-
staying within a certain margin of the previous day's price.
data(λ)=(data^λ−1)/λ ….λ is not = 0
The next choice that is available is the number of layers and data(λ)=log(data) …λ is = 0
the number of neurons per layer. Hence, the model will
perform a pattern aided regression algorithm and artificial The sudden changes in data are observed significantly in this type
neural networks to correctly predict the bitcoin value. The of normalization, so that the data can be processed more accurately.
accuracy can be compared with different models after the final
prediction. Here are the results obtained after implementing various
normalization techniques:
IV.ONGOING WORK AND ACHIEVED RESULTS
The first step towards Bitcoin prediction is database
collection. For our paper we have collected database from the
following sources:

A. Quandl
Quandl holds databases related to financial, economic, and
social background from over 500 publishers. Data available on
Quandl can be used on different platforms such as Python,
MATLAB, Maple and Strata. We were able to procure
datasets for Bitcoin for up to 5 years of timestamp data with
specifications such as –Data high, Data low, Open, Close,
volume of transaction, weighted price.
Figure 1. Graph of data before normalisation

B. CoinMarketCap
CoinMarketCap keeps a track of all the cryptocurrencies
available in the market. They keep a record of all the
transactions by recording the amount of coins in circulation
and the volume of coins traded in the last 24-hours. They
continuously update their records as they receive feeds from
various cryptocurrency exchanges. CoinMarketCap provides
with historical data for Bitcoin price changes.

The next step is database normalization. We basically

perform this step to achieve consistency i.e. reduce or
eliminate duplicate data, insignificant points and other
redundancies. For normalizing our data, we have used five
different techniques.
Figure 2. Graph of data using Log Transform normalisation technique

ISBN 979-11-88428-01-4 ICACT2018 February 11 ~ 14, 2018

International Conference on Advanced Communications Technology(ICACT) 146

Figure 3. Graph of data after using matlab in-built function

Figure 6. Graph of data after boxcox normalisation technique

V. PROPOSED WORK
A .Bayesian Regression
1. Break the first third of the data into all possible
consecutive intervals of sizes 180s, 360s and 720s.
Apply k-means clustering to retrieve 100 cluster
centers for each interval size, and then use sample
Entropy to narrow these down to the 20 best/most
varied and hopefully most effective clusters.
2. Use the second set of prices to calculate the
corresponding weights of features found using the
Bayesian regression method. The regression works as
follows –
x At time t, evaluate three vectors of past prices of
Figure 4. Graph of data after standard deviation normalization technique
different time intervals (180s, 360s and 720s).
x For each time interval, calculate the similarity
between these vectors and our 20 best kmeans
patterns with their known price jump, to find the
probabilistic price change dp_i.
x Calculate the weights, w_i for each feature using a
Differential Evolution optimization function.
3. The third set of prices is used to evaluate the
algorithm, by running the same Bayesian regression
to evaluate features, and combining those with the
weights calculated in step 2.

B. GLM/Random forest:

1) Construct three-time series data sets for 30, 60, and 120
Figure 5. Graph of data after z-score normalization technique
minutes (180, 360, 720 data points respectively) preceding
the current data point at all points in time respectively.
2) Run GLM/Random Forest on each of the two time series
data sets separately.
3) We get two separate linear models: M1, M2 corresponding
to each of the data sets. From M1, we can predict the price
change at t, denoted ∆P1. Similarly, we have ∆P2 for M2.
4) Combine these values to predict the macro price change
defined as ∆P=W0+∑Wj∆Pj, where W0 is initial market
value at t=0, and Wj denotes the weight at the given
interval.

ISBN 979-11-88428-01-4 ICACT2018 February 11 ~ 14, 2018

 International Conference on Advanced Communications Technology(ICACT) 147

5) In addition to using 10-second interval data, we can also Shreya Maji, age 21, was born in Pune, India.
use 10-minute interval data to gain a longer-term picture of She has been extremely devoted towards the
technical education and is currently studying in
the price trends. final year of the undergraduate engineering
course from SPPU, Pune, India. Her area of
specialization is electronics and
VI.CONCLUSIONS telecommunications. With the current project
of Bitcoin Prediction, Ms. Shreya Maji is
After establishing the learning framework and completing aiming for a place in IEEE publications for the
the normalization, we intend to use the two methods first time.
mentioned above and choose the best method to solve the
Bitcoin prediction problem.

ACKNOWLEDGMENT
We would like to thank our project guide Mr. Kaustubh
Sakhare for encouraging us to work on this project and our
parents for the constant support.

REFERENCES

[1] D. Shah and K. Zhang, “Bayesian regression and

Bitcoin,” in 52nd Annual Allerton Conference on
Communication, Control, and Computing (Allerton),
2015, pp. 409-415.
[2] Huisu Jang and Jaewook Lee, “An Empirical Study on
Modelling and Prediction of Bitcoin Prices with
Bayesian Neural Networks based on Blockchain
Information,” in IEEE Early Access Articles, 2017, vol.
99, pp. 1-1.
[3] F. Andrade de Oliveira, L. Enrique Zárate and M. de
Azevedo Reis; C. Neri Nobre, “The use of artificial
neural networks in the analysis and prediction of stock
prices,” in IEEE International Conference on Systems,
Man, and Cybernetics, 2011, pp. 2151-2155.
[4] M. Daniela and A. BUTOI, “Data mining on Romanian
stock market using neural networks for price prediction”.
informatica Economica, 17,2013.

AUTHOR BIOGRAPHIES

Siddhi S. Velankar was born in Pune, India on

23rd December 1995. She is currently studying
Electronics and Telecommunication engineering
in Pune Institute of Computer Technology. She
has immense interest in fields like signal
processing and machine learning and wishes to do
further research in the same. This will be her first
IEEE published paper.

Sakshi Valecha was born in Patiala, Punjab,

India, in 1996. She is currently pursuing her
B.E. from Savitribai Phule Pune University in
the field of Electronics and
Telecommunication from Pune institute of
computer Technology in Pune, Maharashta,
India. With your conference Ms. Valecha is
attempting to get a place in IEEE publication
for the first time.

ISBN 979-11-88428-01-4 ICACT2018 February 11 ~ 14, 2018