← News

ResearchSources: arXiv · 20 Nov 2026

Researchers propose falsifiable compositional hypotheses for time series forecasting

An arXiv paper argues that top-scoring forecast models often fail when data regimes shift, and suggests testing stated assumptions instead of searching historical data for winners.

In plain words

Time series models predict future values from past patterns. The authors say models that win on standard historical tests can still fail when the underlying data-generating patterns change. Their proposed fix: state assumptions about the data up front, then run tests designed to disprove them.

What was announced

The paper reports that models ranking highest on historical validation splits often degrade across non-stationary regime transitions, meaning situations where the statistical behaviour of the data changes over time. The authors attribute this to treating historical backtesting as an open-ended search and model-selection objective, rather than as a way to check assumptions stated in advance. They argue that searching large pools of architectures, hyperparameters, or reasoning rules biases selection toward shortcuts that only hold for the historical period.

Using theoretical analysis plus diagnostic experiments on a controlled multi-primitive benchmark and German day-ahead electricity prices, the authors found that expanding validation search did not improve accuracy outside the observed regime. Decomposing data into primitives without an upfront falsification test also broke down when residual correlations shifted. Their proposal combines neural primitive encoders, explicit coupling constraints, and falsification tests calibrated during training.

What it does not show

  • The evidence is the paper's abstract only; full methods, results and peer review status are not available in the supplied passages.
  • The reported experiments cover a controlled benchmark and German day-ahead electricity prices; performance in other domains is not established by these passages.
  • The proposed approach's effectiveness is claimed by the authors and has not been independently confirmed.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.