EXAONE Demand 1.0

"A Time Series Foundation Model for Demand Forecasting." LG AI Research's second domain-specific time-series foundation model, after EXAONE Finance. General TSFMs see little demand data during pretraining, and demand series have properties those corpora rarely contain: short histories, frequent zeros, censoring by stock-outs, and exogenous events such as promotions that the series does not record. EXAONE Demand addresses this with two components. The first is a demand corpus of 11.3M series and 48.4B observations from 73 sources, plus a synthetic generator for behaviour that open demand data under-represents: intermittency, moving holidays (Lunar New Year, Chuseok, Easter, Thanksgiving), promotions with a build-up, peak and post-promotion dip, cannibalisation, price elasticity, product launches and discontinuations, and stock-out censoring under an (s,S) inventory policy. The second is a demand-aware adapter. A frozen general-domain backbone (24 blocks, width 1,024, 8,192-step context, pretrained only on KernelSynth Gaussian-process series) gets five low-rank branches in each of its attention and feed-forward projections: one shared branch and one for each of the four classical demand classes (smooth, intermittent, erratic, lumpy). A router reads eight scale-free statistics of the input series to weight the four class branches. Only the branches and the router are trained (9,000 steps on 8 GPUs).

LG built two versions: one trained on real and synthetic demand, and one trained on the synthetic corpus alone, because open demand data carries licences that a model trained on it inherits. The authors ran 36 TSFMs themselves under one protocol on 22 held-out demand datasets. The full model has the best geometric-mean MASE (1.0667) and wins 91.9% of its 836 pairwise comparisons. Next come the synthetic-only version (1.0742, 88.0%), TiRex-1.1 (1.0818, 82.5%), Chronos-2 (1.0885, 80.0%) and TimesFM-2.5 (1.1073, 76.2%). The frozen backbone alone scores 1.1050. Adding real-world data lowers error on 16 of the 22 datasets and raises it on 6.

All twelve authors are at LG AI Research, and Wonbin Ahn is the corresponding author. The report is on arXiv under CC BY-NC-ND 4.0. The weights are not public: the Hugging Face repository the paper links (LG-AI-Research/EXAONE-Demand-1.0) returned HTTP 401 at filing. The GitHub repository it cites as code, LGAI-Research/EXAONE-Forecast (6 stars), held only the EXAONE Finance inference runtime and the EXAONE Forecast general-model report at filing, with no demand code.