Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
             Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018




    USING TENSORFLOW TO SOLVE THE PROBLEMS OF
    FINANCIAL FORECASTING FOR HIGH-FREQUENCY
                    TRADING
                                A.V. Bogdanov, A.S. Stankus a
     St. Petersburg State University, University Embankment 7–9, St. Petersburg, 199034, Russia

                                       E-mail: a alexey@stankus.ru


The use of neural networks significantly expands the possibilities of analyzing financial data and
improves the quality indicators of the financial market. In article we examine various aspects of
working with neural networks and Frame work TensorFlow, such as choosing the type of neural
networks, preparing data and analyzing the results. The work was carried out on the real data of the
financial instrument Si-6.16 (futures contract on the US dollar rate).

Keywords: Artificial Intelligence, recurrent neural network (RNN), financial market forecasting,
TensorFlow.


                                                          © 2018 Alexander V. Bogdanov, Alexey S. Stankus




                                                                                                        513
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
             Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018




1. Neural network selection
         With an increase in the power of computing resources, it became possible to predict the price
movement of stock markets using artificial neural networks (ANN). The most common form of ANN
used to predict the stock market is a direct transfer network, using the back-propagation error
algorithm to update network weights. These networks are commonly referred to as reverse error
propagation networks. Another form of ANN, which is more suitable for price prediction, is a
recurrent neural network (RNN) [1] or a time delay neural network (TDNN) [2]. Examples of RNN
and TDNN are the networks of Elman, Jordan, and Elman-Jordan.
         RNN was created with the ability to process long serial data and solve problems with the
distribution of context in time. The model processes one element in a sequence in one-time step. After
the calculation, the updated state is transmitted to the next step in time to facilitate the calculation of
the next element.




Figure 1. A recurrent neural network with one hidden element (left) and its unfolding version in time (right). The
 expanded version illustrates what happens in time: st - 1, st, and st + 1 are the same unit with different states at
                                      different time steps t - 1, t, and t + 1
        However, simple networks that linearly combine the current input element and the last output
element can easily lose long-term dependencies. To solve this problem, researchers created a special
neuron with a much more complex internal structure to remember the long-term context, called the
Long-Short Term Memory (LSTM) cell. He is smart enough to find out how long he has to memorize
old information, when to use new data and how to combine old memory with new input [3].




                                       Figure 2. Structure of LSTM neuron
        Stock prices are time series of length N, defined as p0, p1, ..., pN-1, in which pi is the closing
price in the period i, 0≤i <N. We have a sliding window of fixed size w), and we move it to the right
by w so that there is no overlap between the data in all sliding windows.


                                                                                                                 514
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
             Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018



        The RNN model we are building has LSTM cells as major hidden elements. We use the values
from the very beginning of the training in the first sliding window W0 to the window Wt at time t:




to forecast prices in the next Wt + 1 window:




Essentially, we are trying to find an approximation function, f (W0, W1, ..., Wt) ≈Wt + 1.




                                 Figure 3. The detailed version of the RNN
        Considering how Backpropagation through time, BPTT [4] works, we usually train RNN in a
“detailed” version, so we don’t need to back propagate too far in time and increase the complexity of
learning.
        Before conducting the experiments, a comparison of the popular framework was carried out in
order to select the optimal tool for predicting price movements (data are tabulated).
                                                             Table 1. Comparison of various frameworks
     Framework             Distributed          Architecture         Visualization          Community
                            execution           optimization                                   support
TensorFlow           High                  High                 High                  High
PyTorch              High                  High                 High                  High
CNTK                 High                  High                 Regular               -
MXNet                Regular               High                 Regular               -
Torch                -                     High                 Regular               Regular
Caffe                High                  High                 -                     -
Theano               -                     High                 Regular               Regular

To solve the problem, was chosen TensorFlow as a framework for RNN neural networks.




                                                                                                        515
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
             Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018




2. Data preparation
         The cost of the financial instrument Si-6.16 increases with time, which leads to the fact that
most of the values in the test set are out of allowable limits, and thus the model must predict values
that it has never seen before. In this situation, the network ceases to behave adequately (Figure 4).




         Figure 4. An example where the RNN tries to predict values that lie outside the training data


         To solve the problem with the scale, it is necessary to normalize prices in each sliding
window. Now the task is to predict the relative levels of change instead of absolute values. In the
normalized sliding window W't at the moment of time t all values are divided by the last unknown
price - the last price in Wt-1:



As a result of the data normalization, we get the following price movement:




                                      Figure 5. Normalized data Si-6.16




                                                                                                         516
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
             Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018




3. Experimental results
When conducting experiments, the following values were used in the TensorFlow settings:
                                                                            Table 2. Variable environment
Variable                                                                    Value
num_layers                                                                      1
keep_prob                                                                      0.8
batch_size                                                                     64
init_learning_rate                                                           0.001
earning_rate_decay                                                            0.99
init_epoch                                                                      5
max_epoch                                                                     100
num_steps                                                                      30

As a result of experiments on forecasting financial instrument Si-6.16, we obtained the following
results:




                                       Figure 6. Experimental results

        As you can see, the result of forecasting coincides quite well with the change in the price of a
financial instrument.

4. Conclusion
        Experiments have shown the possibility of predicting the price movements of stock markets
using artificial neural networks.
        However, at the moment, we can only predict the direction of price movement, rather than a
specific value, which increases the probability forecasting.
The accuracy of prediction of the price movement ≈ 62%.

References
[1] Recurrent neural network - https://en.wikipedia.org/wiki/Recurrent_neural_network;
[2] Network with a time delay - https://en.wikipedia.org/wiki/Time_delay_neural_network;
[3] Understanding LSTM Networks. August 27, 2015 - http: //colah.github.io/posts/2015-08-
Understanding-LSTMs/
[4] Backpropagation through time - https://en.wikipedia.org/wiki/Backpropagation_through_time

                                                                                                        517