cs.AISep 25, 2026
SaveEXAONE Demand 1.0: A Time Series Foundation Model for Demand Forecasting
Organizations: LG AI Research
Abstract
Time series foundation models (TSFMs) are pretrained on series from diverse domains, where demand series make up only a small fraction. Demand data has properties that such corpora rarely contain: Short histories, frequent zeros, censoring by stock-outs, and exogenous events that the series does not record. To this end, we propose EXAONE Demand, built on 1) a demand-specific corpus and 2) a demand-aware adapter. For the corpus, we assemble 11.3M series and 48.4B observations from 73 sources, and a synthetic generator supplies the behaviour that open demand data under-represents. For the adapter, we attach low-rank branches to a frozen general-domain backbone, one for each of the four demand classes (smooth, intermittent, erratic, and lumpy), and a router that reads eight scale-free statistics of the input series decides how much each branch contributes. We build EXAONE Demand in two versions, one trained on real-world and synthetic demand together and one trained on the synthetic corpus alone. On 22 held-out datasets, both versions outperform 36 TSFMs, and real-world demand adds a gain over synthetic data alone.
Figures & tables
Figure 3: Overview of EXAONE Demand.
Figure 4: Pipeline for demand dataset construction. 1) Collection: Real sources are grouped by their published role, and the synthetic corpus forms a group of its own below. 2) Preprocessing: The two corpora run on separate tracks. The real track filters, converts, and verifies each source before removing leaked series, while the synthetic track generates, validates, and labels its own. 3) Train-test split: Of the five roles, only the evaluation suites reach the test side, and the other four are used for training, together with every one of the synthetic series.
| Domain | Sources | Representative content |
| Retail and e-commerce | 24 | Store and SKU sales, grocery baskets, fashion and online retail |
| Energy | 23 | Household and industrial electricity load, national demand |
| Transport | 14 | Taxi and ride-hailing pickups, bike share, transit ridership, port calls |
| Healthcare | 6 | Prescription dispensing at practice, pharmacy, and hospital level |
| Service and hospitality | 4 | Restaurant covers, hotel and booking demand |
| Telecommunications | 3 | Cell-grid traffic |
Table 1: Collected sources by domain. The count is over the 93 collected sources, and the last two rows are source types rather than domains, since each of them holds many datasets under a single published name.
| Published role | Sources | Series | Sources in this group |
| Pretrain only | 8 | 8,542,433 | Time-300B [ 63 ] , GIFT-Eval pretrain [ 2 ] , Chronos [ 4 ] , LOTSA [ 72 ] , EIA-930 [ 68 ] , IMF PortWatch [ 32 ] , NYC TLC [ 53 ] , NYC 311 [ 56 ] |
| No split | 67 | 999,102 | Goiener smart meters [ 28 ] , UCI electricity load [ 67 ] , Citi Bike [ 12 ] , Divvy [ 19 ] , NHS dispensing [ 54 ] , Online Retail II [ 9 ] , OPSD [ 57 ] , AEMO [ 6 ] , and 59 others |
| Train only (Held-out private) | 3 | 815,098 | Web Traffic [ 35 ] , Instacart [ 64 ] , Kaggle sales forecasting [ 59 ] |
| Train and test | 11 | 50,489 | M4 [ 50 ] , M5 [ 51 ] , Favorita [ 36 ] , VN1 [ 69 ] , Rossmann [ 34 ] , Walmart sales forecast [ 1 ] , Walmart recruiting [ 33 ] , Store Item Demand [ 37 ] , Panama load (two releases) [ 49 ] , Food demand [ 38 ] |
| Evaluation | 3 | 899,618 | GIFT-Eval [ 2 ] , fev-bench [ 62 ] , BOOM [ 18 ] |
Table 2: Collected sources by published role. The role is a property of the release rather than a decision of ours, and what a source finally contributes is settled by the leakage check rather than by the role alone.
Figure 6: Corpus composition by series count. Every series carries one domain of the converted corpus.
| Series | Time points | |||
| Value | Share | Value | Share | |
| Train | 10,049,114 | 88.88% | 47,419,572,862 | 98.06% |
| Test | 899,618 | 7.96% | 810,403,280 | 1.68% |
| Removed | 358,008 | 3.17% | 125,461,105 | 0.26% |
Table 4: Train and test separation. Removed is what the leakage check takes out of training, and the shares are over the whole corpus.
| Family | Property | Generating mechanism |
| Counts | Discreteness | Negative-binomial sampling from a multiplicative intensity |
| Overdispersion | Dispersion parameter sampled per series | |
| Intermittency | Bernoulli mask over the intensity, targeted to a demand class | |
| Calendar | Day-of-week profile | Multiplicative weekday factors |
| Seasonality | Fourier components at annual and weekly periods | |
| Moving holidays | Lunar New Year, Chuseok, Easter, and Thanksgiving as shifting dates |
Table 5: Design of the synthetic generator by demand properties. Each property has a component of its own in the sampling process, and each component is turned on or off for every series with its strength drawn at random, so that the properties combine freely rather than in the fixed groups that a fixed recipe would give.
Figure 7: Representative synthetic demand series. One series per generator. Spare parts and B2B orders realise the intermittent and lumpy regimes that open data supplies only sparsely, new product realises a full launch-to-discontinuation life-cycle, and e-commerce carries explicit promotional responses of its own.
Figure 8: EXAONE Demand on the four demand classes. (1) Raw context : The window that the model receives, with no rescaling of its values. (2) 8 Statistics : The scale-free summary that the window is reduced to. (3) Router : One weight per routed branch, with the always-on shared branch added at a fixed weight.
| Class | Source | 8 Statistics | |||||||
| Zero | ADI | CV 2 | Trend | AC(1) | Up | CV | Length | ||
| Smooth | Citi Bike | 0.00 | 0.23 | 0.04 | 0.16 | 0.36 | 0.87 | 0.10 | 0.56 |
| Intermittent | Port calls | 0.98 | 1.26 | 0.00 | 0.09 | 0.02 | 0.95 | 0.67 | 0.70 |
| Erratic | SKU orders | 0.03 | 0.24 | 0.32 | 0.04 | 0.10 | 0.66 | 0.28 | 0.51 |
| Lumpy | Retail sales | 0.86 | 0.69 | 0.14 | 0.01 | 0.00 | 0.79 | 0.47 | 0.63 |
Table 6: Example router input for four demand series. Each row is the vector of Equation 4 as the router receives it, and the four series are drawn from four different real-world sources and are of four different lengths.
| Smooth Frequency , Variability | Erratic Frequency , Variability | |
| Intermittent Frequency , Variability | Lumpy Frequency , Variability |
Table 7: Classical demand classes. A low ADI means that demand arrives often, and a high CV 2 means that its size varies widely.
Figure 11: What the router produces on real demand series. Each panel holds a context window and the weight that the gate gives to each of the four routed branches, against the target that supervision asks for, where the four bars carry the smooth, intermittent, erratic, and lumpy branches of the mixture in that same order.
| # | Model | Ranking | Error metrics | ||||||
| Avg. rank | Win rate | MASE | ND | WQL | MAPE | MAE | MSIS | ||
| 1 | EXAONE Demand | 4.09 | 91.9% | 1.0667 | 0.1409 | 0.1138 | 0.2565 | 60.24 | 10.70 |
| 2 | EXAONE Demand (Synthetic) | 5.55 | 88.0% | 1.0742 | 0.1420 | 0.1147 | 0.2583 | 60.73 | 10.63 |
| 3 | TiRex-1.1 [ 5 ] | 7.64 | 82.5% | 1.0818 | 0.1451 | 0.1161 | 0.2911 | 62.06 | 10.75 |
| 4 | Chronos-2 [ 3 ] | 8.59 | 80.0% | 1.0885 | 0.1497 | 0.1406 | 0.3010 | 64.02 | 14.48 |
| 5 | EXAONE Backbone (Zero-shot) | 7.95 | 81.7% | 1.1050 | 0.1445 | 0.1169 | 0.2609 | 61.82 | 11.41 |
Table 8: Comparison with TSFMs. Both of our models are shown against every baseline we ran and against the frozen backbone that we adapt, on the 22 evaluation datasets, with the rows ordered and numbered (#) by MASE.
Figure 12: Forecasts on six demand series. Each line is the median forecast of one model, and EXAONE Demand reaches the lowest MASE of the three on all six series. On Car parts (M) and Electricity (D), where demand falls to zero, TiRex-1.1 and Chronos-2 keep forecasting a positive level, while EXAONE Demand follows the zeros.
Figure 13: What the training data contributes. [Left] The share of synthetic series in the mixture, where each point is the median of three seeds except the synthetic-only point at , which is the median of its own runs. [Right] The change in MASE from adding real-world demand to the synthetic corpus, on the 22 datasets in the evaluation suite.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Group | Setting | Value |
| Backbone (frozen) | Blocks | 24 |
| Model width | 1,024 | |
| Embedding width | 2,048 | |
| Attention heads | 16 | |
| Context length | 8,192 | |
| Input and output patch | 16 |
Table 10: Hyperparameters of EXAONE Demand. The backbone group describes the frozen model that we adapt, and the remaining three groups describe what we add to it and how it is trained on the demand corpus.
| Domain | Series | Share | Observations | Share |
| Energy | 4,507,842 | 39.87% | 32,646,256,620 | 67.51% |
| Retail | 1,942,682 | 17.18% | 641,490,811 | 1.33% |
| Transport | 1,578,373 | 13.96% | 9,742,740,721 | 20.15% |
| Compute and cloud ops | 1,133,431 | 10.02% | 3,686,572,872 | 7.62% |
| Web traffic | 880,755 | 7.79% | 1,348,375,729 | 2.79% |
| Health | 817,901 | 7.23% | 32,301,783 | 0.07% |
Table 11: Corpus composition by domain. Shares are over the whole converted corpus, including the synthetic portion. Series that arrived inside the general-purpose corpora are counted under the domain of the dataset they came from, not under the corpus that delivered them, so that the shares reflect the data rather than its packaging.