Cryptocurrency forecasting suffers from extreme cross‑asset scale heterogeneity, non‑stationary dynamics, and structural dependencies among OHLC (Open, High, Low, Close) variables. CryptoL introduces a unified framework that first evaluates forecasting error in context‑normalized coordinates within the RevIN pipeline, preventing inverse normalization from adding a squared‑scale weighting to the mean‑squared error objective. This effect is formally characterized through empirical risk and parameter‑gradient geometry, establishing conditions under which large‑scale assets can dominate shared‑model optimization.
At the normalization level, the framework compares channel‑independent and channel‑dependent approaches. Experiments show that a shared channel‑dependent affine transformation preserves candle‑order relations, whereas independent channel transformations may break them. To further mitigate numerical distortion across orders of magnitude, CryptoL adds scale‑adaptive numerical stabilization, dynamically adjusting the normalization constant according to each asset’s magnitude.
A soft feasibility loss is also incorporated to penalize violations of OHLC inequalities, expressed as $$L_{feas}=\sum_t \max(0,\,Low_t-High_t)^2+\max(0,\,Open_t-Low_t)^2+\max(0,\,High_t-Close_t)^2.$$
Controlled ablations on heterogeneous cryptocurrency assets demonstrate that each component improves forecasting accuracy, training stability, and the frequency of financially valid OHLC predictions relative to baselines. Overall, CryptoL delivers scale‑balanced optimization, structure‑preserving normalization, numerical stabilization, and constraint‑aware forecasting.
Review: The paper offers a comprehensive solution to scale bias and physical‑constraint issues in multivariate crypto time‑series forecasting, providing a solid foundation for more reliable trading models.