Improvement of the method for forecasting the structure of financial market dependencies using machine learning methods and graph-based models

Abstract

The study is devoted to forecasting the dependency structure of the financial market considered as a dynamic system of interconnected assets. Unlike traditional approaches focused on predicting the values of time series, an improvement of the existing approach is proposed in which the object of analysis and forecasting is the dependency graph between financial instruments. The introduction substantiates the relevance of the problem associated with the necessity to account for the dynamic nature of interrelationships between assets.
The conducted review of contemporary scientific literature covers correlation-based methods, vector autoregressive models, the Graphical Lasso approach, and machine learning algorithms applied to identify dependencies among financial assets. It has been established that existing approaches have limitations in forecasting the future configuration of dependency structures, as they are primarily oriented toward static estimation of such structures or prediction of individual market indicators.
The aim of the study is to advance the method for forecasting the dependency structure of the financial market through the integration of graph-based models of market interrelations with machine learning methods. Within the framework of the study, the financial market is formalized as a graph model constructed on the basis of the inverse covariance matrix, which enables the representation of direct conditional dependencies between assets. The Graphical Lasso method is employed to estimate this structure and construct a sparse graph of market interrelations.
An improved approach to modeling the temporal evolution of dependency structures is proposed through the use of a sliding window for sequential graph estimation at different time points and analysis of parameter changes over time. Based on the current state of the graph structure and the characteristics of its dynamics, a feature set is formed and subsequently utilized by machine learning models to predict the future configuration of links between financial assets.
In addition, the architecture of an information system implementing the proposed approach has been developed. The system includes subsystems for market data collection, preprocessing, graph construction, feature engineering, and dependency structure forecasting.
The effectiveness of the proposed approach is intended to be evaluated experimentally using real financial data and graph structure recovery metrics, including Precision, Recall, F1-score, as well as graph structural similarity measures. To validate the feasibility of the proposed improvement, the results are planned to be compared with baseline dependency analysis methods, including correlation-based approaches and VAR models. The obtained results are expected to provide an assessment of the practical applicability of the proposed approach and determine prospects for its further development and application in financial analytics tasks.

Keywords: financial market, dependency structure, graph models, Graphical Lasso, time series, machine learning, forecasting.