Filtering Ichimoku and ADX-Wilder Signals with a Spectral Mixture Network

Summary

The article applies supervised learning as an additional filter for three Ichimoku and ADX-Wilder signal patterns that had previously performed poorly in forward-walk tests. Its model uses a spectral-mixture-inspired feature layer to represent oscillatory behavior, followed by fully connected neural network layers that map the learned features to an output. Separate subnetworks estimate frequency, variance, and weight parameters; the design combines periodic feature extraction with a regression head. The article also discusses tuning component count, numerical stability, overfitting, and deployment through ONNX.

The author reports that the network changes the earlier signal-only results and describes one tested pattern as profitable, but with a choppy equity path. The article suggests combining the learned filter with conventional indicator signals, while acknowledging that the tests are limited and require further diligence. It does not establish robust performance across markets or periods. There are also conflicting descriptions of whether the output uses a sigmoid, so the exact model configuration is not fully clear from the text.

Key ideas

• A spectral mixture-inspired layer is used to extract periodic features before a neural network regression head.
• Separate subnetworks learn frequency, variance, and weight parameters for the feature transformation.
• The model is applied as an extra filter to selected Ichimoku and ADX-Wilder signal patterns.
• The author reports a profitable but choppy equity trajectory for one tested case and calls for further testing.
• Component count, overfitting, compute demands, uncertainty estimates, and output-activation descriptions are limitations to consider.

Tags

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