Beyond the Hype: Forecasting with Graphs and Foundation Models

Author: Denis Avetisyan

New research reveals that while foundation models excel at short-term forecasting, success isn’t guaranteed and depends heavily on the data itself.

Time Series Forecasting Models (TSFM) operate on a five-minute sampling interval, establishing a granular temporal resolution for predictive analysis.

A comparative analysis demonstrates that model performance in spatio-temporal forecasting is significantly impacted by data characteristics, requiring careful alignment between model architecture and spatiotemporal conditions.

Despite advances in machine learning for environmental sensing, a universally optimal forecasting model for spatiotemporal data remains elusive. This work, ‘No One-Model-Fits-All: Uncovering Spatio-Temporal Forecasting Trade-offs with Graph Neural Networks and Foundation Models’, systematically evaluates the performance of classical, neural, graph, and time series foundation models on real-world temperature data from a wireless sensor network. Our results demonstrate that while graph neural networks excel in sparse deployments, multivariate time series foundation models generally outperform them, contingent on sufficient sampling frequency and spatial coverage. How can we best align forecasting model architecture with the inherent characteristics of spatiotemporal data to build truly efficient and robust systems?

The Limits of Conventional Forecasting

Traditional time series models struggle with spatiotemporal data due to their inability to capture the complex, nonlinear dependencies inherent in these datasets. They often assume temporal independence or rely on simplistic spatial smoothing, inadequate for representing intricate relationships across time and space. Accurate short-term temperature forecasting is vital for diverse applications, yet achieving high-fidelity forecasts remains challenging, particularly at fine spatial resolutions. Existing methods generalize poorly, lacking robustness due to overreliance on localized data and insufficient accounting for regional and global climate drivers.

Spatial Relationships as Mathematical Structures

Spatio-Temporal Graph Neural Networks (STGNNs) offer a powerful framework for modeling interconnected time series by explicitly incorporating spatial dependencies. A core component is the representation of these relationships using an Adjacency Matrix, defining connections between nodes and facilitating information propagation. Connection strength is often determined by statistical measures like Pearson Correlation, quantifying interdependence between time series.

The adjacency matrix heatmaps reveal that STGNNs exhibit distinct connectivity patterns at 20% and 60% sparsity levels.

Graph Convolution and Temporal Dynamics

Graph Convolutional Networks (GCNs) enable convolution operations directly on graph-structured data, capturing spatial dependencies inherent in the graph’s topology. Combining GCNs with Recurrent Neural Networks (RNNs), particularly Gated Recurrent Units (GRUs), extends modeling to incorporate temporal dynamics. Architectural variations like GRUGCN and TGCN optimize performance; GRUGCN achieved an MAE of 2.088 with 30-minute intervals and 8 nodes, while T-GCN reported an MAE of 1.976 using a 45-minute interval and 25 nodes.

Foundation Models and the Pursuit of Generalization

Time Series Foundation Models (TSFMs) represent a significant advancement in spatiotemporal forecasting, establishing robust baselines for complex prediction tasks. Models like Chronos, Moirai, and TimesFM are pre-trained on extensive data, enabling transfer learning and improved generalization. Moirai demonstrates state-of-the-art performance, achieving an MAE of 0.93 with eight nodes, substantially surpassing GRU models (MAE of 2.98) and Transformers. Comparative analyses consistently demonstrate Moirai’s superiority over Graph Neural Networks, suggesting its architecture effectively captures the underlying dynamics of spatiotemporal data.

Sensitivity and the Limits of Observation

Forecast accuracy is demonstrably sensitive to sampling rate and spatial density. Higher rates and densities generally yield improved accuracy, but at the cost of increased computational demands and storage. This trade-off presents a challenge for large-scale applications. Future research should prioritize optimizing these parameters in conjunction with developing efficient STGNN architectures. Investigating adaptive sampling strategies and model compression techniques will be crucial for scaling STGNNs to handle increasingly large datasets and complex dynamics.

The study’s findings regarding the interplay between model architecture and data characteristics echo a fundamental tenet of robust system design. It demonstrates that simply applying a powerful model – such as a Time Series Foundation Model or Graph Neural Network – does not guarantee optimal performance. As Tim Berners-Lee once stated, “The Web is more a social creation than a technical one.” This applies to forecasting as well; the ‘technical’ models must be carefully aligned with the ‘social’ context of the data – its inherent spatiotemporal conditions and heterogeneity. The research highlights that without such alignment, optimization efforts become, as Berners-Lee would observe, a form of self-deception, masking underlying issues rather than addressing them.

What’s Next?

The demonstrated sensitivity of forecasting accuracy to data characteristics suggests a fundamental limitation within the current paradigm. The observed outperformance of Time Series Foundation Models over Graph Neural Networks is not a victory for one architecture, but a symptom of a deeper issue: the tacit assumption that a universally optimal model exists for spatiotemporal prediction. Such a notion feels increasingly improbable, given the inherent complexity and heterogeneity of real-world systems. The field must move beyond merely achieving good results and focus on understanding when and why certain approaches fail – or, more precisely, when their approximations become unacceptable.

Future work should prioritize rigorous analysis of the alignment between model inductive biases and the underlying data-generating processes. The reliance on empirical benchmarks, while convenient, obscures the crucial question of whether a model is genuinely capturing the system’s dynamics or simply memorizing patterns. A more mathematically grounded approach, perhaps leveraging techniques from dynamical systems theory or information geometry, is required to establish verifiable criteria for model selection and hyperparameter optimization.

Ultimately, the pursuit of increasingly complex models feels like a detour if it doesn’t yield a corresponding increase in our understanding. The goal is not simply to predict temperature, but to articulate the principles that govern its evolution. The observed trade-offs are not merely engineering challenges; they are invitations to re-examine the very foundations of spatiotemporal modeling.

Original article: https://arxiv.org/pdf/2511.05179.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/