RQ-Kernel CNN for MACD and OBV Signals

Article MQL5 articles

Summary

This article revisits MACD and On-Balance Volume signal patterns and describes using supervised learning to improve their interpretation. It outlines a one-dimensional convolutional network whose kernel sizes and channel counts are designed in relation to a rational quadratic kernel. The kernel’s length scale and shape parameter control how similarity changes with distance, while its heavier tails can represent variation across multiple scales. The article also discusses possible kernel uses such as attention masks, pooling, channel weighting, and pruning, alongside trade-offs between fixed and learned kernels.

The proposed setup standardizes inputs to a 0–1 range and uses global average pooling. The article reports that forward-walk tests following a prior year of training and optimization showed improvement for two of three re-examined signal patterns, while pattern five continued to struggle, possibly because entries lacked momentum confirmation. These are limited tests of selected patterns; the article does not establish broad or durable profitability. Fixed kernels may reduce parameters and overfitting but can adapt less effectively than learned or hybrid kernels.

Key ideas

A rational quadratic kernel can represent similarity across multiple distance scales, with its shape parameter controlling tail behavior.
The described CNN uses the kernel to guide convolution sizes and channel counts for one-dimensional financial inputs.
The article identifies attention masking, distance-aware pooling, channel weighting, and weight pruning as other possible kernel applications.
Forward-walk results improved for two of the three tested MACD and OBV patterns, while one pattern remained weak.
Fixed kernels may simplify a model but can be less adaptable than learned or hybrid kernels.

Using an RQ-Kernel CNN to Filter MACD and OBV Trading Signals

Summary

Key ideas

Tags