EXAONE Forecast

"A Universal Zero-shot Time-Series Foundation Model." The general-purpose member of LG AI Research's EXAONE Forecast family of time-series models. The family's repository also serves the finance model EXAONE Finance, and the EXAONE Demand report cites the same repository as its code. It is a 312M-parameter encoder-only Transformer (24 layers, width 1,024, 16 heads, 8,192-step context, 16-step patches, 21 output quantiles) that handles univariate, multivariate and covariate-informed forecasting in one model. Covariates enter as extra variates, and cross-variate group attention runs only in the last block, following Toto-2.0's report that variate attention is needed in only a few layers. A query-aware scalable softmax keeps attention sharp over long contexts. It emits up to 512 steps in one forward pass.

The report's second contribution is CKD-GP (Composite Kernel Decomposition for GP generation). CKD-GP expands the composite Gaussian-process kernels of the KernelSynth and CauKer generators into product terms and samples each term in closed form in O(T log T) time, with no dense covariance factorization. At 8,192 steps it is up to 2,645× (KernelSynth) and 1,089× (CauKer) faster than dense sampling. The pretraining corpus combines real series from the Chronos corpus, GIFT-Eval's pretraining split and QuitoBench with 78.9M synthetic, 53.1M augmented and 41.7M synthetically coupled multivariate series. GIFT-Eval's TrainTest source (the train splits of its benchmark datasets) and mixup data derived from it are excluded. Training used schedule-free AdamW on eight NVIDIA RTX Pro 6000 GPUs for about 402 hours, with the context lengthened from 2,048 to 4,096 to 8,192 steps.

On GIFT-Eval (97 configurations from 23 datasets), LG reports an aggregate MASE of 0.673 and CRPS of 0.460. It claims first place in the benchmark's zero-shot category on both aggregates and on both rank-based scores; the next model in its comparison scores 0.681 and 0.466. The results LG submitted to GIFT-Eval's public repository on August 3 reproduce those aggregates. By the time the report went public, two later zero-shot submissions scored better on both aggregates: TimesFM 3.0 (0.667 / 0.456, submitted August 31) and Alibaba Cloud's Lingjiang 2.0 (0.666 / 0.457, September 28). EXAONE Forecast-Agent is a learned router over eight TSFMs (Chronos-2, TimesFM-2.5, FlowState, Toto-2.0 and its fine-tune, TiRex-2, Timer-S1 and EXAONE Forecast) plus retrieval of related historical series, where an LLM only narrows the search space. It scores 0.6099 / 0.4185 and still led GIFT-Eval's agentic category at filing. On fev-bench (100 tasks), EXAONE Forecast ranks fourth by skill score (0.445, after Chronos-2, TimesFM-2.5 and TiRex-2) and fourth by win rate (0.742). The report's conclusion describes the model as performing on par with recent TSFMs. The PDF is dated August 2026 and was posted to GitHub on September 29, 2026. It lists 14 LG AI Research authors alphabetically, with Wonbin Ahn as corresponding author. At filing there was no arXiv version and no released weights or model code for the general model.