Filtering TRIX and Williams %R Signals with a Cosine-Kernel Conv1D

Summary

The article combines TRIX trend signals and Williams %R oscillator signals in a convolutional neural network intended to filter trades. It explains Python calculations for both indicators, then describes a one-dimensional convolutional architecture whose kernel sizes and channel progression are shaped using a cosine function. It discusses design choices such as input sequence length, layer count, kernel size, frequency, dropout, and binary classification output.

The reported experiment uses CHF/JPY on a four-hour chart, with 2023 for training and optimization and 2024 for forward evaluation. Of ten indicator patterns considered previously, three had forward-walked; the model was applied as a filter to those patterns. The author reports that patterns 1, 4, and 5 walked, with pattern 4 appearing more convincing, while improvement was marginal. The evidence is limited by the short evaluation window, entries via limit orders, and take-profit exits without stop losses; the proposed architecture may also not suit all datasets.

Key ideas

• TRIX measures the rate of change of a triple-smoothed moving average, while Williams %R locates closing price within a recent high-low range.
• The proposed Conv1D model uses cosine-based variation in kernel sizes and channel counts to extract sequence features.
• The design discussion highlights tuning layer count, kernel size, frequency, dropout, and input length.
• The experiment applies the model as a trade filter to indicator patterns that had already forward-walked.
• Reported results are preliminary and use a short period with limit entries and no stop-loss exits.

Tags

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