Financial time series forecasting is fundamentally an information fusion challenge, yet most existing models rely on static architectures that struggle to integrate heterogeneous knowledge sources or adjust to rapid regime shifts. Conventional approaches, relying exclusively on historical price sequences, often neglect the semantic drivers of volatility such as policy uncertainty and market narratives. To address these limitations, we propose the ASTIF (Adaptive Semantic-Temporal Integration for Cryptocurrency Price Forecasting), a hybrid intelligent system that adapts its forecasting strategy in real time through confidence-based meta-learning. The framework integrates three complementary components. A dual-channel Small Language Model using MirrorPrompt extracts semantic market cues alongside numerical trends. A hybrid LSTM Random Forest model captures sequential temporal dependencies. A confidence-aware meta-learner functions as an adaptive inference layer, modulating each predictor's contribution based on its real-time uncertainty.
Experimental evaluation on a diverse dataset of AI-focused cryptocurrencies and major technology stocks from 2020 to 2024 shows that ASTIF outperforms leading deep learning and Transformer baselines (e.g., Informer, TFT). The ablation studies further confirm the critical role of the adaptive meta-learning mechanism, which successfully mitigates risk by shifting reliance between semantic and temporal channels during market turbulence. The research contributes a scalable, knowledge-based solution for fusing quantitative and qualitative data in non-stationary environments.
Cryptocurrency markets have become an increasingly important component of global finance, with daily trading volumes exceeding $2 trillion as of October 2025 (Chokor and Alfieri, 2021;Statista and CoinGecko, 2025). Unlike traditional financial assets, cryptocurrencies operate within a market structure characterized by pronounced price volatility, uninterrupted trading activity, and heightened exposure to heterogeneous external drivers (Wątorek et al., 2023;Gherghina and Constantinescu, 2025). These drivers are not limited to macroeconomic conditions but also encompass geopolitical developments, regulatory interventions, and shifts in investor sentiment amplified through social media channels (De Leon et al., 2022). Taken together, such features give rise to a market environment in which price formation is highly unstable and difficult to model. As a result, conventional forecasting approaches frequently prove inadequate for capturing the non-linear and rapidly evolving dynamics observed in cryptocurrency prices (Kehinde et al., 2025). Accurate price forecasting in this setting is thus practically significant for institutional investors managing risk, for financial institutions developing digital asset strategies, and for regulators focused on preserving market stability.
Recent work on cryptocurrency forecasting research faces a fundamental paradox. Markets exhibit regime-dependent dynamics where prediction strategies must adapt to evolving conditions (Oyedele et al., 2023;Petropoulos et al., 2022), yet dominant methodological paradigms employ static architectures with fixed integration strategies. Statistical approaches prioritize interpretability through linear relationships (Khedr et al., 2021), while deep learning architectures emphasize temporal dependency modeling (Fischer and Krauss, 2018;Seabe and Moutsinga, 2023). Building on these foundations, recent hybrid frameworks pursue multimodal integration by combining neural networks with various data sources, including social sentiment, blockchain metrics, and volatility spillovers (Han et al., 2025;De Leon et al., 2022;Kim et al., 2022;Fu et al., 2025). Decomposition-aided approaches have demonstrated effectiveness in cryptocurrency forecasting by extracting temporal modes from volatile price series (Mizdrakovic et al., 2024). Despite their architectural heterogeneity, these paradigms converge on a common limitation. Whether employing LSTM-CNN architectures for multiscale feature extraction (Livieris et al., 2021), graph networks for cryptocurrency interrelations (Zhong et al.), or attention mechanisms for dynamic relationship modeling (Younas and Choi, 2025), existing frameworks maintain fixed model weights and predetermined integration rules. When the market microstructure changes around regulatory announcements or macroeconomic events, these systems have limited ability to rebalance their prediction strategy (Stempień and Ślepaczuk, 2025). This architectural rigidity manifests itself in systematic performance degradation during regime transitions (Oyedele et al., 2023). The core methodological gap thus becomes clear: current research lacks adaptive mechanisms that dynamically select between heterogeneous prediction paradigms based on real-time confidence evaluation and market context, particularly for integrating semantic understanding of policy uncertainty with temporal pattern recognition.
To address the gap, we propose the Adaptive Semantic-Temporal Integration Framework (ASTIF). A hybrid system that enables adaptive model selection for cryptocurrency price forecasting. The framework comprises three integrated components working in concert to optimize prediction accuracy. First, MirrorPrompt is a dual-channel small language model (SLM) architecture that processes numerical price sequences and semantic market indices through separate computational pathways. The design enables simultaneous reasoning across both data modalities. Second, an LSTM-based temporal predictor captures long-range dependencies in historical price movements, providing a strong baseline grounded in temporal patterns. Third, a metalearning module evaluates the confidence in prediction from both channels and performs adaptive model selection. Dynamically weights the semantic and temporal predictions based on their respective confidence scores. To validate the reliability of the framework across diverse market regimes, we evaluated ASTIF on a multi-asset dataset of AI-focused cryptocurrencies and traditional technology benchmarks covering the 2020-2024 period.
The contributions of this study are threefold. Methodologically, a confidencebased meta-learning framework enables dynamic selection among heterogeneous predictors, overcoming the limitations of static architectures dur-ing market regime shifts. Architecturally, a dual-channel architecture integrates semantic reasoning via Small Language Models and temporal patterns through LSTM predictors, processing market indices, such
This content is AI-processed based on open access ArXiv data.
