Jun 2026· International Conference on Blockchain· pp. 1-9· 0 citations· 34 references
Computer Science
TL;DR
MoFE conceptualizes cryptocurrency volatility as a superposition of multi-frequency components, which includes user network based fundamental growth, mining costs and halving mechanism caused seasonal volatility, and market sentiment-induced chaos, and delivers superior Directional Accuracy (DA) and Information Coefficient (IC).
Abstract
Forecasting cryptocurrency prices remains a formidable challenge due to inherent non-stationarity, abrupt regime shifts, and multi-scale stochastic dependencies. Conventional deep learning models often struggle to capture complex underlying dynamics, frequently resulting in persistent phaselagged predictions. To address these limitations, we propose MoFE, a novel deep learning framework that integrates Fourier Neural Operators (FNOs) within a Mixture-of-Experts (MoE) architecture. Rooted in the theoretical framework of stochastic differential equations, MoFE conceptualizes cryptocurrency volatility as a superposition of multi-frequency components, which includes user network based fundamental growth, mining costs and halving mechanism caused seasonal volatility, and market sentiment-induced chaos. Specifically, specialized adaptive FNO (AFNO) and Convolution dual-domain experts learn continuous function-to-function mappings to encapsulate global spectral trends, cyclical adjustments and microstructures, while a dynamic gating based MoE mechanism enables adaptive strategy switching across diverse market regimes. Extensive experiments on Bitcoin datasets spanning January 2020 to December 2025 demonstrate that MoFE achieves state-of-the-art (SOTA) performance in both $\mathbf{T}+\mathbf{1}$ and $\mathbf{T}+\mathbf{5}$ forecasting horizons. Notably, the model effectively mitigates the phase-lag effect, delivering superior Directional Accuracy (DA) and Information Coefficient (IC). In high-fidelity simulated trading environments, these predictive gains transfer into significant excess returns and robust risk-adjusted performance, characterized by a high Sharpe ratio.
Forecasting cryptocurrency prices remains challenging due to extreme volatility, regime‐dependent dynamics, and unstable cross‐asset correlations. Statistical methods such as ARIMA and GARCH assume stationarity and linear dependence structures, making them inadequate for capturing non‐linear temporal patterns in high volatile cryptocurrency data. Conventional machine learning methods often require hand‐crafted features and fail to capture the sequential temporal dependencies inherent in price series. Recurrent deep learning approaches such as recurrent neural networks (RNNs) and LSTMs address the issues but suffer from limited parallelization, vanishing gradients, and difficulty in learning multi‐scale temporal patterns. The advances in Transformer have demonstrated strong capability in capturing long‐range temporal dependencies through self‐attention while enabling parallel computation. However, canonical Transformer still struggle with noisy, volatile financial time series due to computational complexity and sensitivity to irrelevant temporal patterns. These limitations motivate the exploration of Transformer‐based architectures. Our study makes three key contributions. First, we propose MSPCIFormer, a novel Transformer‐based architecture that integrates multi‐scale patching with channel‐independent (CI) modelling to capture heterogeneous temporal dynamics while mitigating noise from time‐varying inter‐asset correlations. Second, we conduct comprehensive experiments comparing MSPCIFormer with state‐of‐the‐art Transformer models and strong time‐series forecasting baselines, evaluated using both statistical and economic metrics across multiple forecasting horizons. Third, we establish a unified evaluation framework incorporating both Hold‐out and Walk‐forward evaluation framework with economical metrics for regime‐robustness testing. Empirical results demonstrate that MSPCIFormer achieves the best or tied‐best predictive accuracy among all Transformer‐based baselines across three cryptocurrency assets, while maintaining competitive and stable performance across diverse market regimes.
Huali Zhao, Martin Crane, Marija Bezbradica· Expert systems· 0 citations
This survey reviews the application of Deep Temporal Convolutional Networks (TCNs) in high-frequency cryptocurrency price forecasting, a field challenged by extreme volatility and non-stationary dynamics. Recent studies demonstrate that TCNs achieve superior performance over traditional machine learning models and recurrent architectures by efficiently capturing long-range temporal dependencies through parallelizable structures. We synthesize findings across different market regimes, highlighting TCNs' advantages in computational efficiency, robustness, and adaptability to rapidly shifting trading environments. Moreover, the review examines emerging efforts to enhance interpretability, addressing a key barrier to real-world adoption in financial systems. By consolidating current progress and open challenges, this paper underscores the significance of TCNs as a promising direction for building more reliable forecasting frameworks, while outlining future opportunities to strengthen their practical impact in algorithmic trading and financial decision-making.
Xue Cheng· Applied and Computational En...· 0 citations
Exchange rate forecasting remains a fundamental challenge in financial time series analysis due to the inherent non-linearity, non-stationarity, and complex temporal dynamics of currency markets. This paper introduces AFD-Net (Attention-Enhanced FAN with DLinear), a novel deep learning framework designed to synergistically integrate frequency-domain decomposition, attention mechanisms, and linear time series modeling. The key architectural innovation lies in the sequential pipeline design: the FAN module first transforms non-stationary inputs into stationary residuals, enabling the subsequent attention module to operate in a low-noise environment where temporal dependencies are more reliably captured, and the DLinear backbone then decomposes the attention-enhanced representation into trend and seasonal components for final prediction. This causally motivated design ensures that each component addresses a distinct and well-defined challenge in financial time series modeling. Specifically, the model employs Feature-wise Adaptive Normalization (FAN) to isolate dominant frequency components, thereby mitigating non-stationary patterns effectively. Concurrently, an attention module is incorporated to adaptively emphasize salient temporal features, while a DLinear backbone serves as the core predictive structure, ensuring efficient sequence modeling with minimal computational overhead. We evaluate the effectiveness of AFD-Net through comprehensive experiments on the Exchange-Rate and IMF Exchange Rate datasets, benchmarking against seven state-of-the-art models, including Informer, Autoformer, PatchTST, and iTransformer. The results demonstrate that AFD-Net consistently outperforms baseline models across multiple forecasting horizons, exhibiting particular superiority in long-term trend modeling. Furthermore, ablation studies quantitatively confirm the distinct and complementary contributions of the FAN and attention components. A comparison against ReVIN-based normalization further justifies the FFT-based design of FAN, demonstrating that the performance gain substantially outweighs its modest computational overhead. This work highlights the potential of frequency-aware architectures in enhancing the accuracy and robustness of financial time series predictions.
The cryptocurrency evolution series has piqued the interest of many researchers seeking to understand the underlying factors that drive evolution and predict future values, yielding useful insights across social and economic fields. The paper introduces a new methodology implementing a cluster-based universal framework to forecast daily cryptocurrency realized volatility from exogenous inputs. The mathematical models include internal and external determinants, and, in the proposed universal model, the input set is further augmented with time-series data on realized cryptocurrency volatility based on the hierarchical Dynamic Time Warping (DTW) clustering dendrogram. The forecasting stage is performed by the regressors in the Deep Learning (DL) class. The performance of the models is assessed using Mean Accuracy Percentage Error (MAPE), Root Mean Squared Error (RMSE), and Mean Directional Accuracy (MDA). The final stage of the proposed methodology is the model interpretation carried out by the Gradient Explainer technique. The comprehensive experimental analysis in our work shows that the universal model trained on a time series from the same class reliably outperforms the other models. In addition, the results obtained are better than those reported for other universal approaches in the literature.
C. Cocianu, C. Uscatu, Alexandru Stan et al.· IEEE Access· 0 citations
Although Bitcoin is acknowledged as the largest cryptocurrency by market capitalization and trading volume in the world's financial market, investors face a great deal of risk and uncertainty due to its exceptionally high volatility and non-linear price changes. To provide a data-driven foundation for risk reduction and forecasting support, accurate modeling techniques are crucial. This work attempts to provide a thorough comparative analysis mapping the precise accuracy–efficiency trade-off between Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models under a standardized Grid Search hyperparameter optimization pipeline using a recent Bitcoin closing-price dataset spanning from January 1, 2020, to January 1, 2026. The research methodology follows a structured data science pipeline, beginning with data acquisition from Yahoo Finance, followed by preprocessing using Min-Max Scaling fitted strictly on the training partition to eliminate data leakage. Model development involves an experimental approach where both LSTM and GRU neural controllers are tuned to extract optimal structural weights. The predictive precision of these models is rigorously evaluated using three standard metrics: MAE, RMSE, and MAPE, while processing throughput is measured via hardware execution times. The research findings indicate that the optimized LSTM model achieved superior one-step-ahead predictive precision with a MAPE of 2.32%, whereas the GRU model recorded a higher error rate of 3.94%. However, the GRU model demonstrated a significant advantage in computational efficiency, completing the training process 8.45 times faster than LSTM. In conclusion, while LSTM is recommended as a forecasting support tool for high-precision financial analysis, GRU remains a viable, parameter-efficient alternative for real-time monitoring on resource-constrained systems before real-world financial deployment.
In the highly volatile cryptocurrency market, trading decision support based on price prediction remains a challenging task. Although machine learning and deep learning techniques have been widely applied to cryptocurrency price prediction, many existing approaches rely on correlation-based black-box models, which limits interpretability and robustness. In this study, we employed a NOTEARS-Linear-based Prediction Model (NLBPM) that directly incorporated causal structures inferred through a causal discovery method as structural constraints within the prediction model. Unlike conventional approaches that focus primarily on minimizing prediction error, the NLBPM emphasized return maximization as its objective function, thereby prioritizing practical economic value. Using Bitcoin as a case study, we constructed a model to predict the direction of price movement four hours ahead and evaluated its performance using a rolling-window scheme with a one-month sliding window. Analysis of the inferred causal structures showed that the returns improved when trades were executed only during rolling-window trials in which specific directed edges to the target variable were detected. Based on this finding, we proposed a causal filter strategy that restricts trading to periods in which specific directed edges to the target variable are detected. In the data period analyzed in this study, the selected edge was the one from the opening price (Open) to the target variable. Backtesting experiments incorporating a transaction fee of 0.1% demonstrated that, while the benchmark LSTM model achieved a negative monthly average return of −3.20% and the NLBPM without filtering yielded −0.72%, the NLBPM with the Open filter attained a higher monthly average return of 10.35%. This study supports the usefulness of using inferred causal structure for cryptocurrency trading decision support.
