Convolutional neural networks are currently very popular in a wide variety of applications. The aim of this article is to verify the reasonableness for using convolutional neural networks in cases when time-series data is available. The predictions of a convolutional neural network, that analyzes the graphical representation of a time-series are compared with the predictions of long shortterm memory network, that analyzes time-series data in numerical representation. We show how the accuracy of cryptocurrency predictions could be improved with time-series data inputs versus image arguments.
Market prices of cryptocurrencies are subject to great volatility, partially due to the absence of agreed-upon fundamentals to back up their price value. Thus, the variety of psychological factors and emotional perception of the market situation by the trader is very important when he decides whether to buy or sell a cryptocurrency.
Every cryptocurrency exchange or analytical service has several basic elements,
shown on their index pages. The main one – is a price chart. An example of this chart
is shown in Figure 1. We believe that the appearance of this chart has a great
influence on a trader’s decision to buy or sell a cryptocurrency. [
For this study, we will compare prediction results made by CNN with chart images used as input data to LSTM with time-series data used as input data. The output will be the change in BTC market price for the next minute.
We will use a statistical F-test to determine the quality of predictions. [
For our setup, we used a server equipped with Intel® Core™ i7-6700 Quad-Core CPU, 64 GB DDR4 RAM, 2 x 500 GB SATA SSD hard drives and GTX 1080 Ti GPU. This setup with a powerful GPU allows for rapid model compilation and training of the model with the data.
We used the Keras library and the TensorFlow library to operate with neural network models.
The data itself is stored and grouped into CSV files. Every CSV file contains two columns and 1440 rows. 1440 rows correspond to 1440 minutes in a 24-hour timeframe. The first column includes the closing market price for each minute and the second column represents a minute change of the market price.
We had 525600 files with training examples in total, matching every minute situation during the year. Thus, the critical F-test value using alpha=0.05 is 1.83.
CNN
First of all, we prepared training samples for CNN. To do it, we generated an image for every CSV file. The example of such image is shown in Figure 2. Our CNN consists of three convolutional layers and Multilayer perceptron. The experimental results, using trained CNN are shown in Figure 3.
Using trained CNN results, the calculated F-test value turned out to be -2.9. Although the value of -2.9 is above critical, this level is still low.
To train LSTM network we just fed the CSV files.
Our LSTM network consists of one LSTM layer and three direct distribution layers.
The experimental results using trained LSTM are shown in Figure 4.
Using trained LSTM results, the calculated F-test value turned out to be -1.5. The value of -1.5 is below critical. 5
4. Calculation of the coefficient of relative change in the linear component over the period T by the formula:
EL = a (1) * T * 100 / Emin, (2)
The value of EL is relative and does not depend on the absolute value of the series. If the quantity EL > 0 then the linear component increases.
5. To estimate the harmonic component of the series, we perform a Fourier transform (FT) for the series E'.
6. Determine the moduli of the oscillation amplitudes in the frequency domain A(w). Carry out the filtering of frequencies according to the amplitude values.
To analyze the efficiency of the proposed algorithm, we used MATLAB environment.
Step 1. The linear component is constant. The harmonic component is absent. The results are shown in Figure 5. The value of the coefficient EL displayed on the second chart. In this case, EL = 0.
Step 2. The linear component grows. The harmonic component is absent. The results are shown in Figure 5. The coefficient EL = 20%. The spectrum modules have values in the low-frequency range (1-2 Hz.). It should be remembered that the main frequency band for the module of the real sequence lies in the interval 0<k <= N / 2-1.
Therefore, frequencies above 15 Hz in our example should be ignored. If it is necessary to analyze higher frequencies, we will increase N - the number of sampling points, Figure 6.
Step 3. The linear component decreases. The harmonic component is absent. The results are shown in Figure 7.
Step 4. The linear component is absent. The harmonic component is present at a frequency of 3 Hz - amplitude 20. The results are shown in Figure 8.
Step 5. The linear component increases. The harmonic component is present at a frequency of 3 Hz - an amplitude of 400. The results are shown in Figure 9.
We applied the proposed algorithm to the data presented in Figure 1. The linear component predicted an uptrend with a rate of 10.2% per period. Modules of amplitudes of the harmonic components had an influence in frequencies of up to 10 Hz.
Using trained LSTM with additional inputs results we conducted the F-test resulting in the F-test value of -5.8. This value is above critical, and better than CNN.
In this paper we compared CNN and LSTM efficiency to process time-series data in numerical and image form.
In our experiment, CNN shows non-critical results in terms of F-test, while LSTM with only one input – BTC market price, shows under critical results. But the extension of inputs, based on series data analysis can improve LSTM performance to the non-critical level, prevailing results produced by CNN.