SAGDFN: Sparse Adaptive Graphs for Multivariate Forecasting

Summary

The document introduces SAGDFN, a graph neural network framework for forecasting multivariate time series when relationships among assets change over time. It describes learning connections from data and using Significant Neighbors Sampling to focus on a compact set of candidate neighbors, reducing the graph from a full N×N structure to a smaller N×M representation. The account also discusses sparse attention and α-Entmax as ways to filter weaker connections, with model components trained together using an L1 loss.

The proposed design aims to lower computation and memory costs while retaining useful cross-series information and limiting noisy links. Financial examples include currencies, commodities, and equity indices, but the framework is presented as applicable to other networked time series too. The document explains the method conceptually and begins discussing its implementation, but the supplied text is truncated. It provides no completed empirical comparison or trading results, so its claims about speed, accuracy, and robustness should be treated as proposed benefits rather than demonstrated performance.

Key ideas

• SAGDFN learns changing relationships among time series instead of relying on a fixed graph supplied in advance.
• Significant Neighbors Sampling selects a compact set of candidate connections to reduce graph computation.
• Sparse attention and α-Entmax help emphasize stronger links and suppress less relevant ones.
• The model jointly trains node representations, neighbor selection, attention, and forecasting components.
• The document describes expected efficiency benefits but gives no complete empirical trading evaluation.

Tags

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