EXAONE Demand 1.0: a specialised approach to forecasting demand from short histories

The challenge of demand in foundation models

Time-series foundation models (TSFMs) are generally pretrained on data from a range of domains, but demand series make up only a small part of those corpora. This creates a mismatch: demand data has characteristics rarely found in other types of time series, including short histories, frequent zeros, censoring caused by stockouts, and external events that the series itself does not record.

To address this gap, the paper proposes EXAONE Demand, an architecture intended to capture demand-specific dynamics without relying solely on the generality of existing models. Its premise is that a model trained on generic data may fail to capture the irregular nature of sales, particularly in sectors with high variability or slow-moving products.

Architecture: a specialised corpus and an adapter

The proposal has two main components. The first is a demand-specific corpus containing 11.3 million series and 48.4 billion observations from 73 sources. A synthetic generator is also used to cover behaviours that open demand data underrepresents.

The second component is a demand-aware adapter attached to a general-domain model whose backbone remains frozen during training. The adapter uses low-rank branches, one for each of four demand classes—smooth, intermittent, erratic, and lumpy—and a router. The router reads eight scale-free statistics from the input series to determine how much each branch contributes to the final output.

Training strategy and model versions

The authors developed two versions of EXAONE Demand to assess the contribution of different data. The first is trained on a combination of real-world demand and synthetic data; the second is trained only on the synthetic corpus. This setup is intended to isolate the contribution of real data relative to simulation.

Keeping the general-domain model frozen suggests a computationally efficient strategy: preserve the model’s general knowledge and train only the demand-specific adaptation layer. This could make deployment easier in settings where fully retraining a foundation model would be costly. The paper’s abstract, however, does not establish the actual computational cost or deployment requirements.

Results reported in the preprint

The technical report says that both EXAONE Demand versions outperform 36 TSFMs on 22 held-out datasets. The authors also report that adding real-world demand data improves results compared with training on synthetic data alone.

These are results reported in a preprint, not independently established conclusions. The abstract names the number of models and test datasets but does not provide the metrics, the size of the gains, or how consistently they occur across sectors. The reported advantage from including real-world data suggests that observed market behaviour may matter, but the information available here is not enough to assess the practical scale of that contribution.

Limitations and the need for external validation

The work is listed as a technical report on arXiv. The available record does not establish that it has undergone peer review. Nor does the abstract confirm whether the code, data, or model weights are available for independent reproduction, which would be important for checking the performance claims.

The construction of the synthetic data and its influence on the reported results also need closer examination. The stated evaluation covers 22 held-out datasets, but the abstract does not show how well the model performs across different real-world sectors or under extreme external events. Whether the approach is useful in practice will depend in part on further detail and independent testing in relevant settings.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.