This paper presents a large-scale benchmark of modern deep learning architectures for financial time series prediction and position sizing, with a primary focus on Sharpe ratio optimization. The study evaluates 16 models, including linear models (AR1x, DLinear, NLinear), recurrent networks (LSTM, xLSTM, PsLSTM), transformer-based architectures (iTransformer, PatchTST, TFT), state-space models (Mamba, Mamba2), and hybrid approaches (VLSTM, VxLSTM, LPatchTST, VSN+Mamba2). Using 15 years of daily futures data across commodities, equity indices, bonds, and FX, the authors assess out-of-sample performance through risk-adjusted returns, statistical significance, downside risk, breakeven transaction costs, robustness to seed selection, and computational efficiency.

The key finding is that models explicitly designed to learn rich temporal representations consistently outperform linear benchmarks and generic deep learning models. The hybrid VLSTM (VSN + LSTM) achieves the highest overall Sharpe ratio of 2.39, while LPatchTST and TFT also deliver strong performance. xLSTM demonstrates the largest breakeven transaction cost buffer, indicating improved robustness to trading frictions. In contrast, linear models and generic transformers (iTransformer, Mamba) underperform, highlighting the importance of architectural inductive bias for financial time series. The results suggest that effective financial forecasting models benefit from jointly denoising returns, learning asset-specific dynamics, and encoding temporal structure in a stable and adaptive manner.