COMPARATIVE ANALYSIS OF MODIFIED LSTM, TCN, AND TRANSFORMER ARCHITECTURES FOR MULTI-FREQUENCY EU FINANCIAL TIME SERIES

This article addresses the fragmented evidence on deep learning for EU financial forecasting by developing a theoretically grounded comparative framework for modified LSTM, TCN, and Transformer architectures under multifrequency conditions.It formulates the scientific problem of frequency-volatilitygovernance mismatch and sets the objective of structuring architecture suitability to support subsequent EU-focused empirical benchmarking and cluster analysis.Architectural inductive biases are mapped to frequency profiles (daily/weekly/ monthly), horizon profiles, volatility regimes, and EU instrument characteristics.The methodology operationalizes this mapping through two structured typologies: a multi-frequency comparative model table and a cluster-based series taxonomy linking data regimes to the architectures with strongest theoretical fit.The results indicate systematic, non-universal dominance: modified LSTM variants align most strongly with persistent weekly-monthly series shaped by delayed transmission and institutional inertia; TCNs are theoretically advantaged in dailyweekly settings where localized motifs and moderate volatility prevail; Transformers show the strongest fit in high-volatility daily environments with non-local dependencies and regime shifts.A cluster-based perspective further suggests that regime-shifting mixed-frequency series favor hybrid or constrained-attention designs over monolithic models.The framework is theory-driven and does not report unified, large-scale EU benchmarking results; this limits immediate quantification of effect sizes.The typologies abstract from instrument-specific microstructure and from cross-country data harmonization issues that can materially affect model training.These limitations № 13(54) 2025 1716 can be addressed by future studies implementing standardized EU datasets, multihorizon protocols, and controlled architecture sweeps.The framework supports context-sensitive model selection for EU financial supervision, stress testing, and macroprudential monitoring by aligning architectures with frequency-volatility regimes while recognizing governance constraints on opacity.It also guides experimental design for benchmarking by indicating which model families should be prioritized within specific series clusters.Unlike prior studies that evaluate models in isolated settings or singlefrequency datasets, this study formalizes the interaction between architectural design, temporal resolution, volatility regime, and EU institutional requirements, providing a coherent basis for comparative empirical testing and cluster-informed interpretation.