A Controlled Comparison of Deep Learning Architectures for Multi-Horizon Financial Forecasting: Evidence from 918 Experiments

Battle Count: 67

Highly relevant for quantitative trading infrastructure decisions. Provides evidence-based model selection guidance: ModernTCN as default recommendation, PatchTST as Transformer alternative, N-HiTS for altcoin specialist, DLinear for resource-constrained environments. However, the finding that directional accuracy is indistinguishable from 50% across all architectures means these models cannot directly generate trading signals without explicit directional loss redesign. The 2-2.5x error amplification at longer horizons and architecture-specific degradation patterns inform horizon-dependent strategy design. The variance decomposition showing architecture explains >99% of forecast variance (vs. seed <0.04%) provides actionable guidance for model selection effort allocation.

Moderate to high complexity. The five-stage pipeline requires: Optuna HPO infrastructure, multi-seed training orchestration (648 runs), chronological data splitting, z-score normalisation with inverse scaling, rank-based evaluation, and comprehensive statistical testing (Friedman, Wilcoxon, DM, Spearman, ICC, JT). However, the authors provide a complete open-source codebase with unified CLI, YAML configurations, checkpointing/resume support, and modular organisation separating preprocessing, models, training, evaluation, and benchmarking. Individual model implementations follow standard PyTorch patterns with unified interface (input: B,w,5; output: B,h).

Exceptional reproducibility: full codebase, data, trained models (648 checkpoints), evaluation outputs, HPO trial logs, frozen configurations, and statistical test results are released under open licence. Deterministic seed management across Python, NumPy, PyTorch CPU/CUDA, and PYTHONHASHSEED. Unified CLI for pipeline execution. Environment specified via environment.yml and requirements.txt. Checkpointing and resume support included. Only minor limitation: torch.use_deterministic_algorithms not enabled, allowing small floating-point variation across GPU architectures.