Comparing CNN Depth and Parameter Choices in a Trading Model
Summary
This document summarizes experiments with a convolutional neural network (CNN) trading model called Deep Alpha. It reports that a seven-layer version performed better than a two-layer version in a same-period comparison, which the authors attribute to learning more market features. Against a Deep Alpha DNN baseline, the CNN was slightly weaker, with only a small reported difference. The text does not specify the assets, target, evaluation metrics, data construction, or test design, so the reported comparisons cannot be independently assessed from this summary.
The model’s results varied substantially after parameter changes, making them difficult to reproduce. Across six parameter groups, smaller parameter values were associated with better performance; the suggested explanation is that they capture finer-grained information. Several configurations lost profitability after 2020, raising concern that learned relationships did not persist far beyond the training period. These observations point to sensitivity and possible decay in predictive information, but the document offers no detailed results, statistical tests, or evidence that the findings generalize to other datasets or markets.
Key ideas
- A seven-layer CNN reportedly outperformed a two-layer version in the same-period comparison.
- The CNN performed slightly below the DNN baseline, with a small reported difference.
- Results changed sharply with parameter adjustments, limiting reproducibility.
- Smaller parameter values performed better across six tested groups, according to the document.
- Several configurations lost profitability after 2020, suggesting limited durability of learned signals.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.