Jev and time series: clear shapes work, raw numbers sit near chance

The Jev decision model is built for questions about text. A new preprint looks at what happens when you send it a series of numbers instead: sensor readings, stock prices or an event log. Authors from the Hong Kong University of Science and Technology (Guangzhou) turned time-series tasks into questions with fixed options and tested jev-1.13.0 on them. S69 The result is uneven: the model picks up something on pronounced curve shapes and on event types, but on most numeric series it answers at roughly chance level. S69 For a Slovak online shop thinking about using Jev to predict sales or price movements, this is a useful warning.

How the authors set up the tasks

The preprint was written by Jianye Wang, Kexin Zhang, Jiayin Xiong and Hao Xue. S69 They arranged the observations chronologically in the state, with channel names and timestamps where needed, and added a fixed question to each task. The model reads the numbers as ordinary tokens. S69 Two-way tasks were asked as Noul, multi-class tasks as Choice. S69 The paper does not say exactly what the input looked like, for example the rounding or the window length. S69

They tested six groups of tasks: 41 two-class and 86 multi-class univariate datasets from the UCR archive, 17 multivariate UEA datasets, five EasyTPP event-sequence datasets, eight ETT settings and daily closing prices of three stocks (AAPL, BA, CAT). S69 For ETT, the question was whether the oil temperature would rise within 24 hours. For the stocks, it was whether the price would rise after a fixed number of trading steps. S69

The authors report results in two metrics, accuracy (ACC) and AUC. S69 AUC shows whether the model gives higher probabilities to correct answers than to wrong ones. A value of 0.5 corresponds to random guessing.

Where Jev held up and where it did not

Datasets with clear shapes did best. On SyntheticControl Jev had an accuracy of 0.81 and an AUC of 0.9978, on SmoothSubspace an accuracy of 0.90 and an AUC of 0.9815. S69 On the EasyTPP event sequences the mean AUC was 0.7028, and on the Taobao dataset it reached an accuracy of 0.6343 and an AUC of 0.8241. S69

Stocks and industrial measurements sit at the other end. Across the three stocks the mean AUC was 0.5040 and the accuracy 0.5034, which is practically chance. S69 On ETT the mean AUC was 0.5900, and according to the authors accuracy stayed around 0.5 in most settings. S69 Most numeric classification tasks were also close to chance: the median AUC was 0.5047 on the multi-class UCR datasets and 0.4986 on UEA. S69 Physiological signals such as ECG also did poorly. The authors name limited domain knowledge and insufficient capture of signal features as possible explanations, but they do not settle on a definite cause. S69

Good ranking does not yet mean a good decision

The gap between AUC and accuracy is worth noting. On the GunPoint dataset the AUC was 0.7215, yet accuracy was only 0.52. On BasicMotions the AUC was 0.7671 and accuracy 0.30. S69 So the probabilities sometimes rank correct answers higher, but the final pick derived from them still goes wrong. The authors see this as a reason to examine thresholds for two-way questions and how comparable the scores are across classes. They did not measure this directly in the preprint. S69

What this means for a Slovak online shop

The tempting idea goes like this: send Jev the daily sales of a product for the past few weeks and ask whether they will rise next week. The stock and oil-temperature tasks had the same form, a numeric series plus a question about whether the value will rise, and the results there were close to chance. S69 JEV.sk's own suggestion: compute trend, seasonality and year-on-year comparison in code or with a standard statistical method. Leave Jev the questions about text, for example whether a customer's message reports a delayed parcel. If you still want to use a numeric signal, put an already computed description in the state ("sales fell by 30% over 14 days") instead of raw numbers. That too is only a suggestion; the preprint did not try this approach.

The gap between AUC and accuracy has one more practical consequence. Anyone who uses Jev's probabilities for ranking should set a threshold and check it on their own data rather than relying only on the most probable option. The confidence and calibration guide describes how to work with confidence and thresholds. The Choice, Score and Noul page explains how the question types differ.

Limits of the preprint

This is a preprint, and only one model was tested. The authors did not compare Jev with any other model or with standard time-series methods, so we do not know how simple statistics would have fared. S69 They did not measure price or response speed. S69 The version tested was jev-1.13.0. Since 14 October 2026 jev-1.14.0 has been available, and the jev-latest alias now points to it. S01 Whether the newer version handles time series better is not examined in the preprint, and JEV.sk has not measured it.