This article is devoted to the issue of formalization and description of technical aspects of development of an information system for generating recommendations for risk-oriented trading strategies on stock exchanges based on the use of deep learning models. The study used a data set representing information on exchange trading in Apple assets obtained from the Yahoo Finance system. The concept of a software system including three functionally independent modules was developed, their formal schematization was carried out. A project of the system with the construction of a diagram of the main components displaying the relationships between the elements was created. In the PyCharm development environment, a structure of directories and files was developed to organize the system software. A graphical user interface with interactive widgets was implemented, providing opportunities for entering, processing and visualizing data. An analysis of the work of the developed modules was carried out, including a description of the strategic recommendations they generate for making trading decisions. The obtained results were interpreted, their key features were identified. Promising areas of further research were determined and possible ways to improve the system were outlined.
In the modern financial market, including the sphere of trading assets, there is a growing interest in
participating in exchange trading, which is becoming a current trend among investors [
In order to efectively perform analytical procedures and compete with institutional investors to
achieve target results in the formation of financial decisions, there is an increasing need for specialized
tools that can automate the process of analyzing alternatives, identify trends and evaluate trading
strategies [
Traditional methods of technical analysis often do not provide the necessary level of adaptability and
accuracy in the conditions of the modern financial environment [
Thus, the development of our own analytical system in the field of issuing recommendations on
risk-oriented strategies for conducting exchange trading is due to the need to improve the accuracy,
speed and adaptability of decisions made. A promising direction is the integration of ML and DL
methods to create comprehensive decision support tools [
Modern models of support for making trading decisions based on the use of ML and DL algorithms have been actively integrated in recent years to solve applied problems in the field of financial management and accounting, including on stock markets in order to automate data analysis processes, predictive analytics and improve the quality of choosing trading strategies. A feature of the data on trading systems is their temporal nature, i.e. the variability of parameters (price, demand, volatility, etc.) over time. In this regard, it should be noted that time series analysis methods are appropriate for use in short-term forecasting, which is due to the nature of the change in the forecast horizon, an increase in which may lead to the need to calculate new values based on the obtained indicators.
This is acceptable at some time intervals, but the accuracy can be significantly reduced, especially
in the case of an increase in the forecast horizon [
A feature that must be taken into account when choosing ML and DL models and their
parameterization for solving repressed models is the fact that time series in real economic, financial, marketing and
trade problems are most often non-stationary and homogeneous. In fact, this means that the residual
formed by subtracting a non-random component from a series is a stationary time series [
In addition to classical ML methods used for such problems, the ARIMA, GARCH approaches and
ANN models with eficient memory (for example, LSTM architecture) should be especially noted. The
ARIMA model is the result of a kind of generalization of the autoregressive moving average model, its
purpose is to build forecasts of time series of diferent types, in practice it is actively used to perform
procedures for analyzing the value of shares and various financial indicators [
This model is a type of autoregressive models, the essence of which is to organize the process of
calculating forecast values at fixed points in time based on previously obtained values. An important
aspect of the operation of such models is the study of not the values of the process directly, but focusing
on modifying its indicators relative to each other. Smoothing outliers of the series is carried out on the
basis of the moving average approach by replacing the original calculated value with the arithmetic
mean of the members closest to it [
The aim of the work is to develop a project and software implementation of an information system for issuing recommendations on risk-oriented target trading strategies.
In this paper, given the specifics of the topic under consideration, the choice of the Yahoo Finance (YF)
platform as the initial data source seems reasonable and rational [
To efectively work with YF data, it is advisable to use the yfinance library, which is implemented in the Python programming language. This tool provides a convenient interface for loading and processing ifnancial information. The library allows you to extract a wide range of data, including: • historical data on stock prices for a certain period (Open, High, Low, Close, Adjusted Close,
Volume); • real-time trading information (if available); • company financial statements; • information on dividends and stock splits; • metadata on sectors, industries and option chains; • currency rates and ETF data.
Thus, using the yfinance library in combination with Yahoo Finance data provides the necessary functionality for conducting complex financial market analysis, including developing predictive models and evaluating asset management strategies.
The system project was developed using the Python programming language in the PyCharm integrated development environment. To work with the data, libraries were used that provide processing, analysis and visualization (NumPy, Pandas, Seaborn, and Sklearn), as well as tools for implementing machine and deep learning methods (TensorFlow, Keras, Sklearn). The keras and pickle formats, as well as the joblib library, are used to save the created models. The recommendation generation system consists of three autonomous data analysis modules, the results of which are integrated to form risk-oriented trading strategies: • Trading indicator analysis module for generation signals for opening and closing trades based on the intersection of two moving averages (SMA, EMA), also RSI, MACD and Bollinger Bands indicators are taken into account in proposed system; • ARIMA-based time series analysis module (ARIMA module), which supports the creation of basic models, automatic parameter selection and seasonal models development; • DL module, which usefull for building stock price forecasting models for a given time horizon.
The interaction between the modules is presented as a diagram (figure 1). Data preparation is carried out via API, which includes selecting information from Yahoo Finance based on specified parameters (time interval, ticker type), saving in a separate collection, trend analysis, calculating and displaying closing price values for the period, and taking into account the capitalization of splits in a specified time period. Based on the generated data, a preliminary study is carried out using a number of approaches: • Moving average methods for smoothing out price changes and identify trends in the data. It allow us to identify optimal moments to enter or exit transactions. • ARIMA models used for data analyzing with diferent trends (seasonality and pronounced) and future trades prices and volumes prediction. YF
API
Data collection Customization
Models serialization
Trading decisions
Calculations Moving Averages
ARIMA
Data researching
Deep Neural
Networks Visualization
Profiling
Metrics evaluation
Forecasting
Data preparation
Model formation Results analysis
• Deep learning models. Designed to predict possible market behavior scenarios. After creation, the models are adapted to the required time ranges and data parameters, which afects the accuracy of their operation. The final models are serialized and saved in pickle or keras format for subsequent use.
The models can then be loaded and used to perform analytical tasks separately from the working environment in which they were created. The next step is to evaluate the results of the generated models, including by visualizing two-dimensional graphical diagrams to highlight the necessary context of the obtained data estimates, profiling the model testing processes, assessing the quality metrics of the models, and performing the forecast procedure, which together are used to support the adoption of trading decisions on the principle of “buy/sell/wait”.
The user of the first module has the ability to analyze the loaded data, perform calculations and evaluate the values of indicators (SMA, EMA, RSI, MACD), interpret the results in graphical form, and save text results of data analysis and generated visualizations. For the ARIMA model generation module, the general list of use cases is similar to the previous module, but a number of additional aspects have been implemented.
In particular, the user enters a ticker, time range and input parameters for building ARIMA models (simple, with automatic parameter selection and seasonal), he also has the ability to initialize the process of creating models, estimating error values based on the use of the RMSE metric, performing forecast operations, displaying analysis metrics and recommendations in the interface, including their output based on forecasts, as well as functions for serializing models of analysis results into object and text ifles, visualizations in graphic files. The diagram of the main component of the system is shown in ifgure 2.
GUI +Tkinter +Input Fields +Buttons +ProgressBar +TreeView +Text Display uses
DataProcessing +prepare_data(data, train_split) +load_data() +evaluate_model(model, X_test, y_test, scaler) +compare_models() +clear_table()
outputs metrics MetricsEvaluation generates data
Visualization +plot_metrics() +plot_all_losses() +compare_models() +build_lstm_model(units, optimizer) +build_cnn_model(filters, kernel_size, optimizer) +build_gru_model(units, optimizer) triggers training
Model provides predictions
Prediction +make_prediction() +save_results() stores results saves/loads models
FileOperations +save_model() +load_model_file()
As part of the use of the DL module, the user is provided with the following options: entering input data; dividing the sample into proportions for the training and test subsets; creating DL models (LSTM, CNN, GRU); training and testing DL models; outputting values of model accuracy evaluation metrics; building visualizations of loss values; evaluating the speed of creating DL models; performing forecasts and issuing recommendations; saving visualizations; saving DL model objects to keras objects. The key components, which are blocks of individual Python functions, are: • GUI, provides various functionality for displaying widgets and special controls and user interaction with the application interface (including processing button clicks, adding data to tables, entering values and text fields); • DataProcessing, performs data processing procedures, loading them and preparing them for the training and testing of models for their evaluation; • Model, directly builds DL models (LSTM, CNN, GRU); • MetricsEvaluation, calculates model quality metrics and compares them with each other; • Prediction, implements the process of starting to build forecasts and providing the results of model operation; • Visualization, performs procedures for constructing visual diagrams and graphs for analyzing the results of model operation, including evaluating their metrics; • FileOperations, implements the functionality of serializing data and objects by saving and loading the created DL models and the results of their forecasts.
The graphical interface of the system is implemented as a desktop software application, with various widgets from the Tkinter library used as interface elements.
The interface includes buttons for training models, loading and saving to files, performing a forecast, plotting a graph based on loss values, saving forecast results and comparing models with each other, as well as clearing tables.
A separate table with scrolling was created to specify the hyperparameter values for each created model.
Based on the implemented modules of the system, a study and analysis of the results of practical use was performed, in particular, the module for assessing trading indicators. The conclusions made based on the results of the analysis are supported by visualizations, an example of which is shown in figure 3, below is an interpretation of the key results.
The RSI analysis revealed several key features. The RSI indicator more often reaches values corresponding to the overbought area, which indicates more frequent price declines in short time periods. Such dynamics indicate short-term corrections, typical for highly volatile markets. In some parts of the time series, there is a weak divergence between the RSI values and price movements. This may indicate a weakening of the current trend and the possibility of a reversal.
In the process of using the DL module, various models of artificial neural networks (ANN) were built, after which their results were compared by the loss metric, as shown in figure 4.
Main conclusions from the analysis: • during training over 100 epochs, there is practically no overfitting efect. This indicates a good generalizing ability of the models; • the GRU model demonstrates the fastest decrease in the loss metric values, already in the first 3 training epochs; • the LSTM model stabilizes last of all, around the 20th epoch.
Comparison of the training speed of the models: • the convolutional ANN model trains the fastest, but shows the lowest accuracy; • training the LSTM model takes more than 6 times longer than the convolutional model; • the GRU model requires the longest time to train among all, but it turns out to be the leader in accuracy labels.
The GRU model demonstrated the highest values of the 2 coeficient, which indicates its best predictive ability. All models used the same optimizer, Adam, which ensures uniformity in the training process and allows for a more objective comparison of their results. From the analysis, we can conclude that the GRU model is the most accurate in terms of accuracy marks and 2 coeficient, despite the longer training time, while the LSTM model requires more time to stabilize and is less eficient in terms of training speed.
The GRU model (figure 5) demonstrates the ability to accurately account for local trend changes and efectively describe periods of market turbulence. When forecasting for the next 30 days (monthly time horizon), the GRU model predicts an optimistic increase in asset prices, which suggests a long-term purchase strategy. At the same time, the CNN model gives a pessimistic forecast, focusing on selling assets, and LSTM ofers a neutral approach, closer to the asset holding strategy.
Thus, if we compare the forecast results obtained on the basis of SMA, EMA, RSI, MACD, ARIMA, GRU, LSTM, CNN with each other in terms of accuracy and completeness, we should note higher values for deep learning models (especially GRU) and the seasonal ARIMA model, allowing us to make more balanced and confident decisions on selling or buying assets. Using only indicator values does not show high reliability of forecasts and is almost 1.5 times less accurate than LSTM models.
The conducted studies confirm that the developed system for analyzing data and supporting trading decisions on the exchange provides the user with the opportunity for a comprehensive assessment of various strategies. It is based on a combination of forecasts generated by trading indicators, ARIMA, GRU, LSTM and CNN models, which is especially important in the conditions of high uncertainty typical of exchange markets. These uncertainties are associated with the influence of many factors – social, political, environmental and economic – that are dificult to formalize.
The key novelty of the work lies in the adaptation, aggregation and hybrid software implementation of diferent approaches to the formation of recommendations for making trading decisions in a single system built on a modular architecture, as well as in the development and optimization of diferent deep learning models with an assessment of their efectiveness.
We can note that decisions should be based on the interpretation of forecasts from all models. To improve accuracy, it is important to consider both the coincidences in their results and the diferences: if all models give consistent forecasts, this increases confidence in the choice of strategy and if there are any disagreements in prediction conservative strategy is more preferable; linear conducting additional analysis of external factors. Also, we want to accentuate that the final trading decision should be made jointly with the data analyst.
The integration of hybrid neuro-fuzzy methods into the system will increase the stability of forecasts, especially in conditions of uncertainty, and minimize the influence of external factors on decisionmaking.
Thus, using the system not only helps to improve the quality of forecasts, but also serves as a tool for building more sustainable trading strategies in turbulent conditions.
Declaration on Generative AI: The authors have not employed any Generative AI tools.