An arXiv preprint dated 20 August 2026 reports that no single early time-series classification method led across all of its controlled tests. End-to-end approaches had the lowest cumulative cost in both incremental drift scenarios, with DQeND ranking first; under abrupt drift, reinforcement-learning methods adapted most effectively when location changed, while ELECTS performed best overall when class changed abruptly.
The joint-adaptive approach was not consistently stable: under repeated abrupt drifts, DQeND’s policy progressively collapsed toward early predictions.
What the systems were being asked to do
Early time-series classification requires two linked decisions: assign a class from a partial sequence and decide whether enough of the sequence has been seen to stop. The study asked whether optimizing those decisions together would be more robust and adaptable as the data distribution changed.
DQeND uses an encoder for raw time-series prefixes and a reinforcement-learning decision module that jointly determines when to stop and which class to predict. Its training is offline on fully labelled time series; during deployment, the encoder and decision module adapt together from decision feedback.
The comparison also included separable Alert pipelines, built with an ELM encoder, a linear classifier and an Alert trigger, as well as end-to-end ELECTS and EARLIEST and bronze and silver controls.
A controlled test of drift
Here, drift means that the pattern in incoming data changes over time. The benchmark was synthetic, binarized MNIST-1D, described as a ten-class problem with univariate series fixed at 100 time points. Researchers varied template location, Gaussian noise and class shuffling through incremental and abrupt changes.
Researchers began with 1,000 examples: 700 to fit models and 300 for an original-distribution hold-out. At each deployment step, they generated 1,500 new drifted examples, using 500 for online updates and 1,000 as a drift hold-out. The deployment ran for 100 steps, and results were averaged over five random seeds.
An eval-then-update protocol scored each batch before the model learned from it. The study tracked cumulative cost and two hold-out AvgCost measures—one on unseen drifted data and one on a fixed original-distribution hold-out—with the weighted AvgCost parameter set to α = 0.8.
The leading method changed with the drift
In the two incremental scenarios, end-to-end approaches were among the lowest-cumulative-cost methods, with DQeND first. DQeND and ELECTS were also reported to preserve the original concept best in the incremental location case.
Abrupt changes produced a different ranking. Reinforcement-learning methods adapted most effectively to abrupt location drift, while ELECTS was best overall in abrupt class drift. Under repeated abrupt drifts, DQeND’s policy progressively collapsed toward early predictions.
Hold-out AvgCost varied by scenario as well. For DQeND, the drift-holdout means ± standard deviation were 0.16 ± 0.03 in incremental location, 0.36 ± 0.01 in incremental noise, 0.19 ± 0.01 in abrupt location and 0.33 ± 0.04 in abrupt class. On the original-distribution hold-out, the corresponding values were 0.09 ± 0.04, 0.31 ± 0.01, 0.11 ± 0.02 and 0.4 ± 0.04, in the same order.
An ablation—tests that freeze one component at a time—favored coordinated updating. Jointly updating the encoder and decision module was reported as best across most scenarios; the largest degradation appeared in the comparison where the decision module was frozen, while encoder freezing had a more limited impact in abrupt settings.
The results were reported as means with standard deviations over five random seeds. That captures variation across those runs, but the rankings remain tied to this benchmark and its selected drift scenarios.
Evidence with a narrow test
These are synthetic-benchmark results. The study tested controlled changes in template location, noise and class composition, so it does not establish how the methods would generalize to real-world streaming data or to other forms of non-stationarity.
Computation also differed. In the CPU-only comparison, ELECTS and EARLIEST were fastest, while DQeND had the highest training and update times and a longer deployment time.
Within the tested scenarios, the ablation favored updating the encoder and decision module together. That result sits alongside DQeND’s instability under repeated abrupt drifts and its slower CPU training and updates.
The paper states that its experiments are fully reproducible and that source code is available from an anonymous repository.
Paper data and sources
Original title: End-to-end Early Classification of Time Series in Non-Stationary Environments
Authors: Aurélien Renault, Alexis Bondu, Antoine Cornuéjols, Vincent Lemaire
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text