Adaptive Periodic Patching for Time-Series Forecasting Models
Summary
This article details a practical token-generation stage for a LightGTS-style forecasting model. It first estimates each input series’ dominant frequency with a Fourier transform, then uses that estimate to set patch length and overlap while keeping the number of patches fixed. The fixed token count is intended to fit the memory and allocation constraints of MQL5 and OpenCL execution.
The implementation uses a maximum-sized weight matrix and zero padding so inactive parts of a patch contribute nothing, avoiding repeated weight reconstruction or pseudoinverse calculations as frequencies change. GPU work-items process segment-filter combinations, while separate kernels and a coordinating class handle the model’s computation and training operations. The article explains the architecture and implementation choices but provides no forecasting accuracy results or trading evaluation, so it does not establish that the method improves predictions or trading outcomes.
Key ideas
- A Fourier transform supplies a dominant-frequency estimate for each univariate input series.
- Patch lengths and overlap adapt to frequency while the output token count remains fixed.
- Zero padding allows a maximum-sized projection matrix to serve patches of different active lengths.
- OpenCL work-items parallelize token generation across segments, filters, and variables.
- The implementation discussion does not report predictive or trading performance.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.