NVExplain's flow-only form performed strongly across the four tested datasets. Its faithfulness pass rate (FPR), a lag-replacement check comparing forecast changes after removing highly ranked lags with changes after removing lower-ranked or random lags, was 99.1% on ETTh1, 88.9% on Exchange, 67.1% on ILI and 96.3% on Weather. It matched the best listed result on ETTh1 and exceeded every listed baseline on Exchange and Weather; Integrated Gradients was highest on ILI at 76.1%.
That advantage did not carry through unchanged when the full pipeline added structure-preserving perturbations and sparse local surrogate models. Its FPR fell to 96.8% on ETTh1, 59.9% on Exchange, 50.6% on ILI and 72.1% on Weather. The authors suggest that approximation error in highly nonlinear temporal dynamics helps explain the weaker results on Exchange, ILI and Weather.
Following the model's internal path
NVExplain is presented as a model-agnostic framework built in four stages. It constructs latent trajectories—sequences of internal numerical states—from rolling context windows, measures changes between consecutive states as semantic flow, aggregates them into a lag-by-horizon attribution map, and fits sparse local surrogate models, or simple stand-ins, for each horizon.
To generate perturbed contexts, the method uses block-bootstrap resampling and Fourier-domain amplitude perturbations intended to preserve continuity and seasonal structure. Comparisons included Random, Attention, Integrated Gradients and TimeSHAP, alongside the flow-only and full NVExplain variants.
The evaluation used the MOMENT forecasting model on four Nixtla benchmark datasets: ETTh1, Exchange, ILI and Weather. The model used 512 time points of context and forecast 72 steps ahead, while explanations covered the most recent 128 lags.
Fast scores, smoother maps
On AOPC and lag-removal Gap, two measures based on forecast changes after ranked lags are removed, the flow-only version was highest on ETTh1, Exchange and Weather, but not ILI. Its reported runtime was 1.952 seconds on ETTh1, 5.611 on Exchange, 10.305 on ILI and 15.133 on Weather, lower than Integrated Gradients and TimeSHAP on all four datasets.
On Weather, the semantic-flow attribution map was described as smoother and more concentrated than TimeSHAP's, with importance focused on recent lags and decaying smoothly across horizons. Its total variation, a roughness measure for the map, was 0.010 versus 0.474 for TimeSHAP.
Faithfulness did not guarantee stability
The stability results did not line up neatly with faithfulness. Mean embedding sensitivity—the change in the model's representation under perturbation—was 0.0084 on Weather, 0.0100 on ILI and 0.0187 on Exchange; the corresponding 95th-percentile values were 0.0101, 0.0128 and 0.0237. ETTh1 was an instability outlier at 0.3683 on average and 0.4841 at the 95th percentile, even though its flow-only FPR was 99.1%.
Trajectory diagnostics showed a similar split: Weather had the lowest zero-crossing rate and jitter—measures of direction changes and relative wobble—at 0.0701 and 0.3533, while ETTh1 had the highest, at 0.2477 and 0.8182. The authors describe these measures as indicators of reliable or cautionary interpretation regimes.
A benchmark, not a verdict
The evidence is limited to MOMENT on four named benchmark datasets, so it does not establish how NVExplain would perform with other forecasting architectures or data. The reported outcomes concern explanation faithfulness, embedding and trajectory stability, and runtime, rather than forecast accuracy. The document is an arXiv version 1 preprint dated 25 August 2026.
Paper data and sources
Original title: NVExplain: Explaining Time Series Forecasting with Latent Trajectory Analysis and Structure-Preserving Surrogates
Authors: Muyan Anna Li, Manikandan Ravikiran, Aditi Gautam
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text