TopoPrimer, July 2026: topology as input for forecasting

TopoPrimer, published July 2026 by Zara Zetlin, Kayhan Moharreri and Maria Safi, makes the global topological structure of a series population an explicit input to forecasting models. The framework is precomputed once per domain using persistent homology and spectral sheaf coordinates, and the authors report improvements on four public benchmarks drawn from Chronos and TimesFM, including a gain of up to 7.3% MSE on ECL.

What is TopoPrimer and how does it work?

TopoPrimer is a framework that supplies explicit topological context to forecasting models: the global topology of the series population is computed once per domain via persistent homology and spectral sheaf coordinates, then provided per token to models as either a fully trained per-token input or as a lightweight adapter for pre-trained backbones. Sheaf coordinates are identified as the primary accuracy driver, and the system is designed so the topology component can be precomputed independently of model training.

The pipeline separates topology extraction from model fitting. Persistent homology captures global structure across series, spectral sheaf coordinates summarize that structure into per-series inputs, and those coordinates are consumed by forecasting backbones either during full training or as an adapter on pretrained models.

How much does TopoPrimer improve forecasts?

Across four public benchmarks on Chronos and TimesFM, TopoPrimer consistently improves forecasting accuracy, with reported gains of up to 7.3% MSE on the ECL dataset; the topology advantage persists with near-identical magnitude whether backbones are zero-shot or fine-tuned. The authors report the improvements are largest in difficult regimes: under peak seasonal demand classical and zero-shot models can degrade by up to 50% while TopoPrimer stays within 10%, and at cold start with no item history TopoPrimer reduces MAE by 27% relative to a topology-free baseline.

Those concrete numbers come directly from the paper: up to 7.3% MSE improvement on ECL, peak seasonal degradation for non-topology models up to 50% versus TopoPrimer within 10%, and a 27% MAE reduction at cold start.

Why does the topology help?

The paper shows topology supplies signals that are complementary to per-series training: the topology advantage persists with near-identical magnitude across zero-shot and fine-tuned backbones, which the authors interpret as topology and per-series training capturing complementary information. Practically, precomputing the global structure concentrates cross-series patterns that individual series histories may miss, helping both robustness under seasonal spikes and performance when item history is absent.

Sheaf coordinates emerge as the primary driver of accuracy gains, which suggests that the way the global topology is mapped back to per-series inputs matters more than merely computing persistent homology. The framework’s split between domain-level precomputation and lightweight per-token deployment also makes it usable as an adapter for pretrained backbones, according to the authors.

What to watch

See whether the topology advantage generalizes beyond the four public benchmarks on Chronos and TimesFM and whether practitioners reproduce the reported numbers in production. Two concrete signals will matter: independent replication of the paper’s reported up-to-7.3% MSE gain on ECL and confirmation that TopoPrimer maintains error within 10% under peak seasonal demand where classical and zero-shot models show up to 50% degradation.

The paper positions topology as an orthogonal source of signal for forecasting models; the next milestones are replication on additional datasets and adoption as an adapter on pretrained forecasting backbones.

Authors: Zara Zetlin, Kayhan Moharreri, Maria Safi. Published July 2026.