Partial multivariate transformer as a tool for cryptocurrencies time series prediction

Battle Count: 83

The paper is highly relevant to quantitative trading as it directly addresses cryptocurrency price forecasting, benchmarks multiple deep learning and statistical models, and critically examines the gap between statistical prediction accuracy and actual trading profitability. The finding that lower prediction error does not consistently translate to higher financial returns is a crucial insight for quantitative traders. The trading simulation framework (sign-based strategy with ROI, Sharpe Ratio, drawdown, and directional accuracy metrics) is directly applicable to strategy development. However, the simple trading rule and lack of transaction cost modeling limit direct practical applicability.

The PMformer architecture involves dual-attention mechanisms (temporal and feature-wise self-attention), learnable embeddings for time steps and feature identities, and a partial multivariate training strategy with random feature subset sampling. While the architecture is described clearly, implementing the dual-attention encoder and the feature subset sampling training loop requires moderate-to-advanced deep learning engineering. Bayesian hyperparameter tuning across 12 models with extensive search spaces adds computational complexity. The model is based on existing Transformer variants (Informer-based), which reduces some implementation burden.

The paper provides detailed hyperparameter configurations (Tables I and II), explicit data splits (70/20/10 chronological), feature engineering descriptions (OHLC, volume, trade count, technical indicators), training procedure (100 Bayesian trials, 3 seeds: 42, 1337, 2025), and hardware specification (NVIDIA T4 GPU). However, no code repository is mentioned, and the exact feature subset sampling strategy during training is not fully detailed. The PMformer architecture is referenced from prior work [9, 21].