CATCH: Frequency Patching for Multivariate Time-Series Anomaly Detection

Summary

The article explains CATCH, a neural network framework for detecting anomalies in multivariate market time series. It transforms normalized data with an FFT, divides the real and imaginary spectra into patches, and uses a channel-masked Transformer to learn cross-instrument relationships within frequency bands. An inverse transform reconstructs the series, with reconstruction error used to identify anomalies.

The article also describes an MQL5 implementation in progress, including complex-valued convolution and GPU processing with OpenCL. Its examples motivate how high-frequency deviations, broader trend changes, and unusual cross-asset behavior might appear in spectral data. The article provides no evaluation on historical markets; it says performance testing will come in a later installment. Its account is therefore an architectural overview and implementation discussion, not evidence that CATCH produces profitable signals or reliably detects market manipulation.

Key ideas

• CATCH transforms multivariate time series into frequency-domain patches using the FFT.
• A channel-masked attention module is designed to focus on learned dependencies between instruments.
• The model reconstructs the series through an inverse transform and flags anomalies using reconstruction error.
• The article outlines complex-valued convolution in MQL5 and OpenCL but does not report market performance tests.

Tags

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