Adaptive Periodic Time-Series Models with Rotary Positional Encoding

Summary

This concluding article in a LightGTS implementation series explains how adaptive periodic segmentation feeds a Transformer forecasting model. Time series are divided into cycles identified with an FFT, converted into tokens through flexible projection, and processed with attention before parallel decoding generates forecasts. Its main technical focus is Rotary Positional Encoding (RoPE): coordinate pairs in query and key embeddings are rotated according to token position so attention can represent relative temporal offsets without adding trainable positional parameters. The article also describes OpenCL kernels for the forward rotation and its reverse operation for gradient propagation.

The discussion places this component within a larger neural trading system and mentions multi-horizon forecasts and an observed tendency toward long-only behavior, with missed short opportunities during reversals. That observation motivates further optimization, especially two-way trading and risk controls. The supplied material emphasizes architecture and implementation; it does not provide enough quantitative results to establish predictive or live-trading performance, and it calls for further testing on real data.

Key ideas

• FFT-based segmentation identifies periodic chunks that are projected into tokens for Transformer processing.
• RoPE encodes relative temporal position by rotating pairs of embedding coordinates according to token position.
• The OpenCL forward and backward kernels apply the rotation and its inverse efficiently across token dimensions.
• The larger model combines adaptive patching, attention, and parallel decoding for time-series forecasts.
• The reported long-only tendency and lack of presented performance validation leave strategy effectiveness unresolved.

Tags

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