Использование ИИ в образовательных целях

Аpplication of neural networks in the learning processs

Abstract. The change in Apple stock prices was studied. A model was built and a neural network was trained. The software implementation was carried out in Python. The main function of the program is to predict stock prices using a neural network.

A method for calculating the mean square error was determined. Mean Squared Error was used for software implementation. This is the average deviation in all iterations of neural network training.

Keywords: neural networks, machine learning, forecasting, LSTM neural network.

I. Introduction

In this work, the DataSet (input data) is divided into two parts: the training part and the validation part. The ratio of the training part to the validation part is 80% to 20%.

We train the neural network. The result of the network training is compared with real data. Such a comparison demonstrates the ability of the constructed model to predict, the error (deviation) of the model and the accuracy of calculating metrics.

II. Data preparation and network training

The robot has carried out three stages of creating a neural network model to solve the forecasting problem:

 the first stage is data preparation;

 the second stage is building a model and training the neural network;

 the third stage is choosing the error function.

Neurons are interconnected and grouped into layers. Our neural network has 3 layers of neurons:

 input layer;

 control layer (computation block and analysis block);

 output layer.

Information is fed to the neurons of the input layer. In our case, the input layer consists of 32 neurons.

The output layer is where the result is taken. One neuron is defined on this layer. This neuron predicts the price for one day. This price is the subject of the study. Input signals can be weakened or strengthened. This happens due to the weights of interneuronal connections.

First, it is necessary to train the neural network. To train the neural network, the data was divided into two parts: training and validation. In the training set, the future stock price is used to train the network. The data was divided in a ratio of 4/5 to 1/5. For training, we have a sample of 4/5, for validation - a sample of 1/5. All data is grouped into blocks.

III. Software implementation

The software implementation is carried out in Python. The main function of the program is to forecast stock prices using a neural network.

The application was created using the Google Collaboratory cloud service from Google. The developed application allows flexible modeling of neural networks. This will help to obtain a higher-quality model for solving the problem.

The program code is divided into several blocks:

• libraries;
• data analysis (Pandas library);
• dataframe splitting into training and validation parts;
• creating a dataset for a time series;
• training an RNN model using Tensorflow;
• eliminating overfitting. Dropout regularization method.

Fig. 1 – Graphs of changes

1. T. Zoumpekas, E. Houstis, and M. Vavalis, “Eth analysis and predictions utilizing deep learning,” Expert Systems with Applications, vol. 162, p. 113866, 2020.
2. S. Alonso-Monsalve, A. L. Suárez-Cetrulo, A. Cervantes, and D. Quintana, “Convolution on neural networks for high-frequency trend prediction of cryptocurrency exchange rates using technical indicators,” Expert Systems with Applications, vol. 149, p. 113250, 2020.
3. A. Dutta, S. Kumar, and M. Basu, “A gated recurrent unit approach to bitcoin price prediction,” Journal of Risk and Financial Management, vol. 13, p. 23, Feb 2020.