KNOWLEDGE MODEL FOR DETECTING CORRELATION DEPENDENCES IN CORPORATE ONTOLOGICAL GRAPHS ON HETEROGENEOUS DATA

Abstract

The article presents a knowledge model for detecting correlation dependencies in ontological systems (CLCO – Context- Linked Corporate Ontology), built on heterogeneous data, in particular, a structured corporate graph of economic relations and a time-bound text context. The main scientific novelty of the work lies in the architectural separation of the structural and contextual layers: the graph of economic relations remains topologically stable, and text data performs exclusively the role of time and event binding without modifying the graph structure. This approach allows for the formalization of corporate relations as specific economic roles and creates conditions for interpretation of the results. The knowledge model is formalized as a typed attributed directed graph, in which vertices corresponding to corporate entities and text documents are highlighted, and typed economic relations between companies (supply, customer relations, partnerships, competitive relations, control). Integration of heterogeneous data is performed by linking textual description of events to the corresponding nodes of the corporate graph. The types of economic relations preserve their role semantics and are used to describe the context, which increases the expressiveness of the situation representation in the model. Text news is aggregated in time intervals and tied to nodes without changing the graph topology. The empirical validation of the model was carried out on daily stock market data (OHLCV) of large public companies traded on the NYSE and NASDAQ, using sliding time windows from 30 to 250 trading days. The target variable was formulated as a binary class of the price direction for the next trading day. Four configurations were compared: the traditional technical analysis model (TA), aggregation of news and technical features without a graph (AGG), and a graph model taking into account types of economic relationships (GRAPH). The experimental results demonstrate the practical effectiveness of the proposed model. In the close–close mode, the GRAPH strategy provided a return of +13.42% versus +3.58% for AGG and −4.35% for TA. In the close–open mode, the graph context increased the return from +17.82% (AGG) to +19.80%. The analysis shows that the graph context acts as a factor coordinating the behavior of related companies, refining the interpretation of events in the temporal slice without changing the structure of relations. Thus, the work confirms the effectiveness of integrating a fixed corporate knowledge graph and a time context for short-term forecasting of price movement direction. The proposed approach provides a transparent and interpretable representation of economic relations, allows assessing the influence of the structural context and serves as the basis for the development of analytical and predictive systems in the financial sector.