Dual Demand and Supply Forecasting with Spatial-Temporal Fusion

Summary

The article adapts Spatial Temporal Fusion (STF), a framework for jointly forecasting related demand and supply series, to trading signals. It treats bullish and bearish price movement as proxies for demand and supply, using high-minus-open and open-minus-low values. A spatial input is built from autocorrelation among those series, while a time input represents the weekday. Separate multilayer perceptrons forecast the two sides, with the supply model also receiving prior demand; subtracting the supply forecast from the demand forecast yields a directional signal.

The author explains the framework’s motivation through two-sided markets and remote-sensing examples, then describes an MQL5 implementation using a hand-coded neural network instead of the source paper’s transformer architecture. The proposed inputs are acknowledged as subjective proxies, and alternatives such as volume or macroeconomic measures are mentioned. The article presents the approach as a prototype rather than validated trading evidence: it provides no clear performance results in the excerpt and warns that neural-network training requires substantial data and compute, with cross-validation needed before relying on forecasts.

Key ideas

• STF forecasts related demand and supply series jointly, incorporating spatial and temporal inputs.
• High-minus-open and open-minus-low price movements serve as bullish and bearish proxies.
• Separate neural networks forecast demand and supply, and their difference forms the directional signal.
• The implementation uses autocorrelation values for spatial input and weekday as its time index.
• The proxy choices are subjective, and the approach needs extensive data, compute, and validation.

Tags

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