LG trained demand AI on 11.3 million series, but the missing promotion still wins

THE SIGNAL IN ONE SENTENCE

LG AI Research has built a forecasting model for the series that make normal forecasting systems sweat: the spare part ordered twice a year, the medicine that suddenly disappears from shelves, the train platform that fills after a concert, and the winter power spike that was not visible last Tuesday. EXAONE Demand 1.0 is described in a September 25 preprint from the South Korean laboratory. The paper assembles 11,306,740 demand series containing 48,355,437,247 observations from 73 converted sources, adds synthetic data for rare demand patterns, and attaches a demand-specific adapter to a frozen general forecasting model. On 22 held-out datasets, both the full version and a synthetic-only version rank ahead of the 36 time-series foundation models the authors tested. That is a serious research result. It is also the beginning of the useful conversation, not the end. Demand is not simply a line waiting to be extended. A promotion can move the line. A holiday can bend it. A product launch can break it. A stockout can hide the very demand a retailer needs to estimate because sales fall to zero when the shelf is empty, not when customers stop wanting the item. The model only sees those outside events through the traces they leave in the history. LG's authors name that limitation directly. Their router reads the context window, not a calendar of promotions, prices and holidays. Stockout censoring is represented in the synthetic generator, but the system does not recover the hidden demand lost behind real out-of-stock sales. The missing promotion still wins because no amount of clever pattern recognition can read a fact that never entered the room. The paper starts by treating demand as its own species of time series. General forecasting models train across weather, finance, traffic, energy, health and many other domains. Actual demand can be a small part of that mixture. It is often short, full of zeros and divided across thousands of individual products, stores or services. A national electricity curve can be long and smooth. A replacement component for one machine can be mostly silence followed by one urgent order. Training one generic behavior across both can make the rare, commercially painful cases disappear inside the average. LG's corpus reaches across retail and e-commerce, energy, transport, healthcare, hospitality, telecommunications, economics, web traffic and computing. The team collected 93 sources and converted 73 into one schema. It excluded sources that were non-tabular, duplicated material already collected or required unavailable credentials. The converted collection has a heavily skewed shape. The lower quartile contains only 128 observations per series, while the median has 1,373. By the paper's classification, about 6.34 million series are smooth, 4.14 million are erratic, 727,598 are intermittent and 97,558 are lumpy. The difficult intermittent and lumpy cases remain a minority, so the synthetic generator deliberately produces more of them. The four labels come from established demand-forecasting practice. Smooth demand appears often and varies relatively little. Erratic demand appears often but jumps in size. Intermittent demand arrives infrequently with more stable sizes. Lumpy demand is both infrequent and highly variable, which is forecasting's polite term for a series that likes to throw furniture. EXAONE Demand does not force every series into one hard box. It computes soft membership across the four behaviors, then teaches a small router to produce weights from eight scale-free statistics. A series near the boundary can be partly intermittent and partly lumpy instead of receiving a brittle label. The adapter has one shared low-rank branch that is always active and four routed branches tied to the demand classes. Those branches modify a frozen general-domain backbone without retraining all of its weights. Each layer can mix the branches differently for the same series. The result is specialization without building four separate forecasting models and a fifth one to choose among them. The training data is also handled with more care than a giant-number headline suggests. Published time-series collections often recycle the same underlying series under different names. A model can look excellent if a renamed copy of its test set appeared during training. LG compares raw values, not merely dataset titles, and removes matching training candidates. The paper reports 358,008 series removed from the training side through that leakage check. The held-out side contains 899,618 series from evaluation suites. The authors also state an important asymmetry: competing models were pretrained before these benchmarks and their original training data was not necessarily filtered with the same procedure. That does not invalidate the comparison, but it belongs in the receipt. LG reran 36 foundation-model baselines and the frozen EXAONE backbone under one protocol instead of copying favorable numbers from other papers. The list includes versions of Chronos, TimesFM, Moirai, TiRex, Sundial, Toto, TabPFN-TS, Granite, YingLong and others. No baseline was selected according to its performance on the 22 datasets, and the paper says no system was tuned on this evaluation suite. EXAONE Demand ranks first in the main table with an average rank of 4.09 across datasets, a 91.9 percent pairwise win rate and a mean absolute scaled error, or MASE, of 1.0667. The synthetic-only version ranks second with an average rank of 5.55, an 88.0 percent win rate and MASE of 1.0742. TiRex-1.1 ranks third in the table. The full version also reports lower normalized deviation, weighted quantile loss, percentage error and absolute error than the synthetic-only version, while the synthetic-only version has a slightly lower mean scaled interval score. In other words, one neat winner depends on which part of the forecast a decision actually needs. Then comes the number that should be taped above every procurement meeting. The paper explains that MASE compares a model with a seasonal naive forecast. A value above one means repeating the last season would have been the better option. Every row in the reported comparison, including EXAONE Demand, sits above one on the present evaluation suite. LG's model is the best-performing foundation model in that table while still losing, on the aggregate MASE interpretation supplied by the authors, to an extremely simple baseline. That is not a gotcha. It is excellent scientific context. A new model can improve the frontier among its peers and still fail the cheapest test a planner should run. The useful deployment question is not whether EXAONE beat TimesFM. It is whether it beats last season, a store-specific statistical model and the current human process after every cost is counted. The model trained with real and synthetic demand performs better than the synthetic-only version on 16 of the 22 evaluation datasets and worse on six. The authors report that the gains on helped datasets outweigh the losses on the others, which supports the aggregate advantage of real data. It also warns against buying one universal score. A retailer, grid operator, transit agency and pharmacy do not share the same error costs. Underestimating insulin demand is not the mirror image of ordering too many seasonal decorations. A forecast can have lower average error while making the expensive mistake more often. EXAONE Demand emits nine quantiles, which can express a distribution instead of one point. That is useful because a planner can see a likely range and choose a service level. The paper evaluates weighted quantile loss and the quality of an 80 percent interval alongside point metrics. Production teams still have to connect those distributions to consequences. What is the cost of a stockout? What expires on the shelf? How quickly can supply respond? Which customers are harmed first? What level of reserve is legally required? A probability band becomes a decision only after someone writes the loss function in human terms. The paper includes six selected forecast plots where EXAONE Demand follows collapses to zero or sharp peaks more closely than two strong baselines. The authors explicitly say those panels were chosen to show the adaptation when it works and that selected curves cannot carry the aggregate claim. That sentence should be standard equipment in every AI release. A pretty chart is an example. A held-out protocol is evidence. A live operating result is something else again. Nothing in the preprint establishes that EXAONE Demand has reduced waste, prevented a stockout, improved staffing, cut reserve margins or saved money in an LG business or another South Korean organization. No production deployment, intervention study, service-level comparison, planner workflow, operating cost or independent replication is presented. The public evidence is a company-authored preprint and benchmark package. That is enough to take the method seriously. It is not enough to let the model order the warehouse. The first practical deployment should run in shadow mode. Feed the model the same historical series available to planners, freeze each forecast before the outcome arrives and compare it with the seasonal naive baseline, the current production system and human adjustments. Keep promotion, holiday, price, stockout and lifecycle data separate at first so the team can see what the univariate model misses. Score by product, location, horizon and demand class. Report the mean and the tail. A model that trims average error but creates catastrophic misses on scarce medicines, transformer parts or peak electricity days is not an improvement. Next, record every override. If a planner raises the forecast because a campaign begins Friday, capture the reason and the source of the information. If the model beats the override, keep that too. Over time, the override log becomes the covariate roadmap the paper already points toward. Promotions, holidays and prices can be added as explicit features only after the organization proves their definitions, availability and timing. A promotion entered after sales begin is not a forecast input. It is hindsight wearing an employee badge. Stockouts need their own treatment. Observed sales are censored when inventory reaches zero. Training on those zeros can teach a model that customers wanted nothing precisely when demand exceeded supply. Teams need inventory availability, unfilled orders, substitutions, wait lists and replenishment timing to estimate lost demand. The paper leaves this recovery problem open. A production pilot should label it openly instead of replacing every zero with a guess. Sensitive domains also need access and purpose controls. Pharmacy demand, transportation usage and household energy can reveal patterns about communities and individuals even when the forecasting target looks operational. The paper combines open datasets across domains, but an adopter still has to inventory its own data, minimize identifiers, set retention periods and restrict who can drill from a forecast into the underlying series. More data is not automatically a better reason to collect it. South Korea has a strong practical reason to care about this work. Manufacturing, logistics, retail, transport, energy and healthcare all depend on forecasts made across enormous catalogs and networks. A domestic research lab building a reusable demand model can lower the effort required to test specialized forecasting across those sectors. The significance beyond Korea is the architecture: build a domain corpus, verify source counts, remove value-level leakage, preserve a frozen backbone and route small adapters according to interpretable characteristics. That recipe can travel. The benchmark victory should travel with its caveats. The plain signal is that EXAONE Demand makes demand look less like one generic line and more like the family of awkward behaviors planners actually meet. Its corpus is large, its routing idea is readable and its evaluation is unusually explicit. The paper also tells us the model cannot see a promotion that is not recorded, does not recover hidden demand from real stockouts and remains above the seasonal naive benchmark on aggregate MASE. That honesty makes the work more useful, not less. Put it beside the planner, not over the planner. Make it beat last season before asking it to beat the world.