IND-Eval Preprint Introduces Domain-Focused Benchmark for Industrial Time Series Models
arXiv preprint from October 12 2026 presents IND-Eval, a public benchmark emphasizing fine-resolution multivariate industrial time series. Evaluation of Chronos-2, TiRex-2 and…
- TimesFM-3 achieves the lowest source-balanced nMASE and nCRPS, leading seven of nine sources in the IND-Eval preprint.
- Source-specific exceptions demonstrate the importance of domain-representative evaluation for industrial model selection.
- A 50/50 mix of general and oscillator-augmented simulation data balances target-domain gains against lost generality on univariate tasks.
IND-Eval Benchmark Design
The preprint introduces IND-Eval as a public benchmark that groups industrial time series by source into forecasting tasks. Relative to prior suites such as GIFT-Eval, it stresses finer sampling rates and a higher share of multivariate prediction problems typical in industrial applications.
This evidence is limited to the abstract; the full paper is not reviewed here.
Foundation Model Evaluation Results
The authors test Chronos-2, TiRex-2 and TimesFM-3 on IND-Eval. TimesFM-3 records the lowest source-balanced normalized Mean Absolute Scaled Error and normalized Continuous Ranked Probability Score, outperforming the others on seven of nine sources for each metric.
Notable source-specific exceptions appear, reinforcing that aggregate scores may not predict suitability for a particular industrial domain.
Simulation and Domain Alignment Case Study
Because real industrial data are often scarce, the work examines whether a domain's statistical structure can shape simulator design. In a case study on two offshore well-monitoring regimes, an oscillator-augmented pipeline improves a general simulator's coverage of the target domain.
This change lowers forecasting error on the target data but degrades performance on the univariate subset of GIFT-Eval. A balanced 50/50 mixture of the general and augmented recipes yields the lowest error across four event-metric pairs in the case study while recovering most of the lost GIFT-Eval performance.
Practical Takeaways for Model Selection
The preprint, dated 2026-10-12, shows how domain-representative benchmarks can inform both model choice and the blending of synthetic pretraining data. Builders of industrial forecasting systems may benefit from testing open-weight candidates on data that mirrors their target distribution rather than depending only on broad leaderboards.
The observed trade-off between specialization and generality suggests that mixing simulation strategies offers a practical route to balanced performance when real data is limited. This abstract-only capture does not replicate the experiments or verify implementation details.
