FORMATION OF AN INVESTMENT PORTFOLIO BASED ON STOCK MARKET FORECASTING USING MULTI-SOURCE DATA

Abstract

The study uses data from various sources: historical stock prices, technical indicators, news reports, and social media publications. This approach makes it possible to combine both quantitative and qualitative factors that directly influence stock market dynamics and to build a more complete and comprehensive information base for forecasting.Unlike traditional models, which mainly take into account only the statistical characteristics of time series, the use of multi-source data allows for a better representation of the complex interrelationships between economic, informational, and behavioral factors that determine fluctuations in the value of securities. The experimental results confirmed that increasing the number of layers and neurons in neural networks, as well as applying regularization in the form of dropout, significantly improves forecast accuracy and helps to avoid overfitting. The best-performing model – a four-layer BiLSTM with 150 neurons and a dropout parameter of 0.2 – achieved a MAPE of 4.36, outperforming the comparable LSTM model. This demonstrates BiLSTM’s ability to better capture long-term dependencies and hidden patterns in time series.At the same time, a significant limitation was the considerably longer training time required by BiLSTM, which may complicate or even preclude its use in real-time systems where decision-making speed is critical. The aim of the study is to develop an effective model for forecasting stock market dynamics based on multi-source data using modern deep learning methods, capable of ensuring high accuracy and reliability of results even under unstable conditions. The object of the study is the process of stock price forecasting under conditions of high volatility and unpredictability of financial markets, driven by both internal economic factors and external informational influences. The subject of the study is deep learning methods and models (in particular LSTM and BiLSTM) focused on the analysis of time series and textual data from various information sources, the integration of which makes it possible to improve forecasting quality. The obtained results indicate the practical feasibility of applying deep learning in investment management, since such methods provide more accurate and reliable forecasts under conditions of significant uncertainty and instability. They can become an effective tool for optimizing investment portfolio formation, reducing risks, and increasing the profitability of financial decision-making. Further research may be directed toward improving model architectures to accelerate the training process and developing hybrid systems that combine deep learning with natural language processing methods for a deeper analysis of the influence of informational factors on the stock market.