Stock Market Movement Forecasting Using Machine Learning and Complex Network Measures

Abstract

We investigate whether complexity-inspired network descriptors extracted from visibility-graph (VG) representations of price dynamics can improve short-horizon forecasting of the Standard & Poor’s 500 movements. We consider two aligned forecasting tasks at a 14-day horizon: (i) Up/Down directional classification and (ii) regression of the 14-day-ahead standardized forward return. Using daily close prices from 1981 to 2025, we construct sliding window graphs (50/75/100 trading days) and compute a broad set of interpretable spectral, topological, and mesoscopic features, including spectral radius, algebraic/natural connectivity, graph index complexity, clustering/transitivity, efficiencies, assortativity, and path-length statistics. Beyond the standard (“natural”) VG, we introduce a volatility-adaptive quantile VG (VAQ-VG) and two additional geometric/motif descriptors—visibility angle entropy and square-motif density—to form a multi-view feature bank. Features are lagged (1–7 days), standardized, and filtered via mutual information. Six learning algorithms (regularized linear models, random forest, k-nearest neighbors, Gaussian process, and stochastic gradient descent) are evaluated under a purged expanding cross-validation protocol with a 150-day embargo to eliminate sliding window leakage, with hyperparameters tuned by randomized search. Empirically, the directional task shows modest but consistent skill: top models reach the area under the receiver operating characteristic curve ≈ 0.70–0.72 with accuracy ≈ 0.65–0.66, remaining reliably above chance across thresholds. In contrast, point prediction of 14-day return magnitudes remains challenging, with the best R 2 below 0.20 and limited gains from more complex regressors. Overall, VG-derived features provide a compact, interpretable representation that supports leakage-robust directional forecasting, while VAQ-VG yields small, model-dependent shifts that suggest complementary structure across views rather than a decisive single best construction.

Received: 31 August 2025 | Revised: 24 February 2026 | Accepted: 10 April 2026

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The data that support the findings of this study are openly available in GitHub at https://github.com/Butman2099/Complex-systems-book and https://github.com/Butman2099/Complex-Network-Paper-for-Artificial-Intelligence-and-Applications-AIA-journal.

Author Contribution Statement

Andrii Bielinskyi: Conceptualization, Methodology, Software, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization. Vladimir Soloviev: Conceptualization, Methodology, Writing – review & editing, Supervision, Project administration. Andriy Matviychuk: Conceptualization, Methodology, Validation, Investigation, Writing – original draft, Writing – review & editing. Vitalii Bezkorovainyi: Methodology, Software, Formal analysis, Investigation, Data curation, Writing – review & editing, Visualization.