EXAONE Demand 1.0 is a time series foundation model purpose-built for demand forecasting, addressing properties like short histories, frequent zeros, stock-out censoring, and unrecorded exogenous events that general TSFMs under-represent. It pairs a demand-specific corpus of 11.3M series (48.4B observations from 73 sources) with a demand-aware adapter of low-rank branches and a statistics-based router. Both a real+synthetic and a synthetic-only version outperform 36 TSFMs across 22 held-out datasets.
A paper on arXiv introduces Aurora-X, a billion-scale time series foundation model trained with a progressive curriculum (channel-independent pretraining, then midtraining with cross-variable dependencies, covariates, and varied context/horizon lengths). Variable-resolution post-training enables test-time scaling under a fixed token budget. It reports state-of-the-art results on GIFT-Eval, TIME, FEV-Bench, TFB, and DAG-Bench against both pretrained TSFMs and supervised models.
A paper on arXiv introduces Aurora-X, a billion-scale time series foundation model trained with a progressive curriculum (channel-independent pretraining, then midtraining with cross-variable dependencies, covariates, and varied context/horizon lengths). Variable-resolution post-training enables test-time scaling under a fixed token budget. It reports state-of-the-art results on GIFT-Eval, TIME, FEV-Bench, TFB, and DAG-Bench against both pretrained TSFMs and supervised models.
A new arXiv paper introduces CEDAR, a forecasting framework that decomposes demand into residual components to support controlled, event-driven predictions for e-commerce planning. It addresses the limitations of passive time series forecasting by enabling merchants to evaluate outcomes under future action sequences like budget schedules.
A new arXiv paper introduces CEDAR, a forecasting framework that decomposes demand into residual components to support controlled, event-driven predictions for e-commerce planning. It addresses the limitations of passive time series forecasting by enabling merchants to evaluate outcomes under future action sequences like budget schedules.
A new arXiv paper introduces QUALS, a large-scale time series corpus equilibrium framework for training zero-shot forecasting foundation models. It combines pattern quantization with learnability synchronization to rebalance heterogeneous training data. Benchmarks show superior zero-shot performance even with substantially reduced training budgets.
A new arXiv paper introduces QUALS, a large-scale time series corpus equilibrium framework for training zero-shot forecasting foundation models. It combines pattern quantization with learnability synchronization to rebalance heterogeneous training data. Benchmarks show superior zero-shot performance even with substantially reduced training budgets.
A new arXiv report introduces $t_0$, a family of open-weights time-series foundation models that condition forecasts on past covariates and known-future covariates without task-specific retraining. The first two releases, t0-alpha (102M) and t0-beta (256M), rank third on GIFT-Eval among zero-shot TSFMs and perform strongly on fev-bench and an independent Macrocosm evaluation.
A new arXiv report introduces $t_0$, a family of open-weights time-series foundation models that condition forecasts on past covariates and known-future covariates without task-specific retraining. The first two releases, t0-alpha (102M) and t0-beta (256M), rank third on GIFT-Eval among zero-shot TSFMs and perform strongly on fev-bench and an independent Macrocosm evaluation.
Researchers propose ORBIT, a training paradigm for time series foundation models that makes pre-training distributions explicit and controllable. It combines Bootstrap Multi-Level Sampling to control dataset exposure and sample records, target variables, and more. The approach addresses under-explored training regimes in TSFMs.
Researchers propose ORBIT, a training paradigm for time series foundation models that makes pre-training distributions explicit and controllable. It combines Bootstrap Multi-Level Sampling to control dataset exposure and sample records, target variables, and more. The approach addresses under-explored training regimes in TSFMs.