Liu Jin, Yahya M.H., Saidatunur Fauzi Saidin, Li Lu
Abstract Multivariate cryptocurrency forecasting is challenging because market series exhibit non-stationarity, cross-variable dependence, heterogeneous temporal scales, and abrupt short-term fluctuations. Although Transformer-based forecasting models can capture long-range temporal relationships, directly modeling raw high-frequency sequences may obscure dominant periodic structures and increase computational cost. This study proposes a frequency-guided multi-scale decomposition and patch Transformer, termed FMDP-Transformer, for multivariate cryptocurrency time-series forecasting. First, a frequency-guided multi-scale representation module estimates dominant temporal periods from the Fourier amplitude spectrum and constructs scale-specific representations through period-dependent average pooling. This module is designed to extract multi-scale periodic information and attenuate short-term disturbances rather than to perform explicit anomaly detection. Second, the resulting representation is decomposed into trend and residual components. A lightweight linear projection is used for parsimonious trend extrapolation, while the residual component is divided into overlapping patches and processed by a Transformer encoder to model local and long-range temporal dependencies. The forecasts produced by the two branches are subsequently combined. Experiments on Bitcoin, Dogecoin, and Binance Coin data derived from the G-Research Crypto Forecasting dataset evaluate the model under multiple forecasting horizons. Comparisons with recurrent, decomposition-based, patch-based, inverted-Transformer, and multi-scale forecasting models, together with component ablations and computational-complexity analysis, are used to assess its effectiveness. The results indicate that frequency-guided multi-scale representation, decomposition, and patch tokenization provide complementary benefits for multivariate cryptocurrency forecasting. Nevertheless, the proposed frequency-guided smoothing operation does not explicitly identify statistical anomalies, and abrupt market movements may contain predictive information rather than noise.