Improving Time Series Forecasts Under Volatility

Summary & Methodology Analysis

The paper identifies a common failure mode in Graph Neural Networks (GNNs), which are models that process data structured as a network of nodes, when the underlying graph topology is not static. To quantify this, the authors propose the Temporal Correlation Volatility metric, a model-agnostic approach for measuring how intensity and distribution of variable correlations evolve. This enables developers to detect when standard graph representations, which treat dependencies as invariant, become unreliable for time series forecasting. The methodology utilizes a Temporal Convolution Layer for capturing piecewise-smooth temporal dynamics alongside the new GLIDE layer designed specifically for dynamic topologies. GLIDE introduces two primary design components: Path-based Message Passing, which aggregates features across higher-order neighborhoods to filter out transient noise, and Static and Dynamic Propagation Separation, which decouples long-term topological structures from rapid shocks captured by first-order temporal difference kernels. The system maps these processed representations to future sequences using a 1D convolutional projection. While the architecture shows significant performance gains, the current implementation has limitations. The Temporal Correlation Volatility metric is restricted to linear correlations, and the theoretical derivation provided in Theorem 3.1 relies on the assumption of mutually independent temporal perturbations, a condition that may not hold in all real world domains.