Deep Learning for Financial Time Series: A Large-Scale Benchmark of Risk-Adjusted Performance

Battle Count: 51

This paper is directly and highly relevant to quantitative trading. It provides a comprehensive, large-scale benchmark of deep learning architectures specifically designed for financial time-series prediction and position sizing under Sharpe ratio optimization. The evaluation framework mirrors real-world trading considerations including transaction costs, downside risk, regime robustness, and implementation feasibility. Key findings (VLSTM highest Sharpe, xLSTM best cost buffer, hybrid models superior downside characteristics) provide actionable guidance for practitioners selecting architectures for systematic trading strategies. The cross-asset futures universe (bonds, commodities, energy, FX, equity indices) and 15-year span covering multiple market regimes make findings broadly applicable to institutional quantitative trading.

The benchmark framework involves multiple architectural families (linear, RNN, transformer, SSM, hybrid) with distinct hyperparameter spaces and training protocols. The end-to-end Sharpe ratio optimization with pooled returns, volatility targeting, EWMA estimation, and rolling-window retraining adds complexity. Hybrid architectures (VSN+LSTM, VSN+xLSTM, LSTM+PatchTST) require careful integration of components. However, the modular two-stage pipeline design (sequence architecture + projection head) provides a unified interface. Hyperparameter search spaces are well-defined but extensive. The proprietary data source and lack of public code repository increase practical implementation barriers.

Code is available upon request (not publicly hosted). Detailed hyperparameter search spaces, model configurations, training protocols (ADAM optimizer, early stopping with patience 20, gradient clipping), and evaluation metrics are thoroughly documented. However, the dataset is proprietary (Pinnacle Data Corp CLC database), limiting full reproducibility. The rolling-window scheme, seed ensembling protocol (top 10 of 50 seeds), and volatility targeting methodology are clearly specified. Architecture details including mathematical formulations are provided in appendices.