250
Blockchain Technology
was an absolute challenge to forecast with high accuracy. And moreover the bitcoin
price prediction is a linear time-bound property.. Hence, a better model has been
proposed with times series analysis and artificial neural network methodology for
bitcoin prediction.
The ARIMA model is one of the well-recognized time series models, and it has
been applied in many domains to identify future price predictions. For example, it
has been applied in domains such as stock price prediction, electricity prices, supply
chain, traffic interval prediction (Lv et al., 2021), chiller system performance (Ho &
Yu, 2021), COVID-19 – probable evolution of the pandemic (Benvenuto et al., 2020;
Alaraj et al., 2021), wind speed forecasting (Liu et al., 2021; Kavasseri & Seetharaman,
2009), forecasting primary fuel demand (Ediger and Akar, 2007) and so on.
The hybrid model of ANN (Artificial Neural Network) and ARIMA has outper
formed all the other forecasting time series models (Guo et al., 2021), (Patel et al.,
2020) for bitcoin price prediction. This model has been designed from the advantages
of linear and non-linear modelling. Moreover, the different kinds of ARIMA models
produce results with minimal errors; however, the error did not make much differ
ence with actual data (Ismail et al., 2021). The linear model ARIMA, combined with
backpropagation neural network (BPNN) has shown better prediction accuracy com
pared with other combinations of linear and non-linear models (Yang et al., 2021).
The performance of short-term models are compared with neural networks for fore
casting in the supply chain process.. In this analysis, the neural networks performed
well compared with short-term models (Bousqaoui et al., 2021). Moreover, research
has been conducted on stock price prediction using neural networks with ARIMA.
In this research, the performance of ARIMA is improved when it is combined with
the ANN models (Kinasz, 2021).
The method used in this research is a hybrid model, which is introduced by sand
wiching the ARIMA model with RNN, which, it is hoped, will improve the accu
racy over the current prediction, since this method is very similar to the RNN and
LSTM method. The results obtained by the ARIMA model with RNN and RNN
with LSTM are evaluated to find the performance and efficiency. In this research, the
input stream is first forwarded into the ARIMA model, which generates a relatively
good and accurate result. This result is again passed into the gates of the RNN model
for its neural convolutions under 300 epochs, which further subdivides and classifies
to produce a better-valued prediction. The proposed model is expected to create a
better result with low resource consumption as well.
15.4 EXPERIMENTATION
For the experiment, a data set was collected from Kaggle for the years 2017 and 2018
(Kaggle, 2021). The details of the data set are as follows:
• coincheckJPY_1-min_data_2014-10-31_to_2018-01-08.csv
• bitflyerJPY_1-min_data_2017-07-04_to_2018-01-08.csv
• coinbaseUSD_1-min_data_2014-12-01_to_2018-01-08.csv
• bitstampUSD_1-min_data_2012-01-01_to_2017-01-08.csv