SparseTSF: Periodic Decomposition for Lightweight Time-Series Forecasting
Summary
The article explains SparseTSF, a compact approach to long-horizon time-series forecasting that assumes a known period in the input data. It splits a sequence into period-aligned subsequences, forecasts those subsequences with shared linear parameters, and then reassembles the outputs. A sliding aggregation step, implemented with convolution, combines nearby observations before sparse forecasting to reduce information loss and lessen the effect of outliers. The method also centers inputs by subtracting their mean and restores that mean to the forecast.
The article describes an MQL5 neural-network implementation and says it was trained and tested on historical data, reporting profitable results on both datasets. It provides no detailed performance statistics or comparison in the supplied text, so that claim alone does not establish robustness or a tradable edge. The authors caution that the example programs demonstrate an implementation and are not ready for live-market use. The approach also depends on useful periodicity being known and may not capture irregular dynamics.
Key ideas
- SparseTSF divides input data into subsequences aligned to an assumed period.
- Shared linear parameters forecast each subsequence and keep the model compact.
- Sliding aggregation combines nearby observations to address information loss and outlier sensitivity.
- Mean normalization is applied before forecasting, with the mean added back afterward.
- The article reports historical training and test results but does not provide enough detail to establish live trading performance.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.