Deep Learning for Robust Intraday Timing and Parameter Selection

Summary

This report describes combining individual-stock timing with multi-factor stock selection in a T+0 strategy. It uses deep-learning ideas to reduce a multidimensional space of parameter settings and win rates, then examines how performance changes across parameter variations. The stated goal is to find stable frequencies associated with trading patterns rather than optimize a single unusually profitable parameter set. The report treats consistent behavior as evidence of possible generalization and contrasts it with one-off parameter overfitting.

For the period from May 2015 to May 2019, the report gives results relative to the Shanghai 50 and CSI 300 benchmarks, including annualized returns, maximum drawdowns, information ratios, and an average position level of about 10%. These are historical reported results, not evidence of future or live performance. The document says signal coverage could rise as the strategy set expands, but provides no further detail in the supplied text. Its focus on parameter stability offers a validation idea, while the brief summary leaves implementation and cost assumptions unspecified.

Key ideas

• The proposed T+0 approach combines individual-stock timing with multi-factor selection.
• Deep learning is used to examine parameter and win-rate distributions for evidence of stability.
• The method seeks parameter ranges with consistent behavior rather than a single best-performing setting.
• The report provides historical benchmark-relative results for May 2015 through May 2019.
• The supplied text does not detail implementation, trading costs, or live validation.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.