Filtering Weak Gator and Accumulation/Distribution Signals with a CNN

Summary

The article tests whether supervised learning can improve three previously weak signal patterns based on the Gator oscillator and the Accumulation/Distribution oscillator. Its proposed filter is a one-dimensional CNN that periodically applies dot-product kernel regression: feature vectors at different time positions are compared, normalized weights are formed, and each position receives a weighted blend of features across the sequence. The author also discusses varying convolution kernel sizes and channel depths, plus alternatives such as local windows, temperature scaling, downsampling, parallel attention blocks, and multiscale fusion.

In tests on GBP/JPY at a 30-minute interval, the author reports that feature pattern 4 clearly reversed its earlier poor showing, while the other patterns did not benefit equally. The stated test window is limited to two years, so the reported result is preliminary and calls for independent diligence. The method also has practical costs: its full similarity matrix grows quadratically with sequence length, uses extra memory, and may oversmooth features or destabilize training.

Key ideas

• Dot-product kernel regression lets CNN features at each time position incorporate information from distant positions.
• The tested model alternates convolutional layers with kernel-regression steps and uses a supervised output as a signal filter.
• The reported tests found a clear improvement for feature pattern 4, but not comparable gains across all revisited patterns.
• The evidence comes from a two-year test on one currency pair and timeframe, limiting the strength of conclusions.
• Quadratic computation and memory use, oversmoothing, and training instability are key concerns.

Tags

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