EXAONE Finance 1.0

"An Attention-free Time Series Foundation Model for Financial Time Series." LG AI Research's EXAONE Forecast for Finance is a time-series foundation model built for financial forecasting rather than adapted from general-domain TSFMs: it drops the self-attention backbone, whose cost grows quadratically with sequence length and variate count, for causal 1D convolutions with group-aware pooling, and pretrains on a synthetic financial corpus. The report claims the top zero-shot result on GIFT-eval. The first version posted August 4, 2026 and a revision on September 7. Filed late.

The 202M-parameter weights went up on HuggingFace on September 3, 2026 under LG's non-commercial EXAONE license, with inference code in the separately licensed EXAONE Forecast runtime repo. The model card reports first place in all three tiers of LG's financial benchmark suite (point forecasts, cross-sectional ranking, portfolio returns).