EXAONE Finance Unveils Attention‑Free Time‑Series Model for Market Forecasting

A recent arXiv technical report (arXiv:2609.04239v1) announces EXAONE Finance, a foundation model specifically engineered for financial time‑series forecasting. While recent time‑series foundation models (TSFMs) have demonstrated strong zero‑shot capabilities after large‑scale pretraining, they are largely built for general‑domain data and rely on self‑attention mechanisms whose computational cost grows quadratically with sequence length and the number of variates.

The report highlights three key limitations of existing TSFMs for finance: (1) they assume fully observed inputs, (2) they are pretrained on corpora that do not reflect the unique dynamics of financial markets, and (3) their self‑attention backbones become prohibitively expensive for the long, many‑channel, intermittently observed panels typical in finance.

To overcome these challenges, EXAONE Finance replaces self‑attention with a linear‑time, attention‑free architecture. The new design uses two simple linear operations that keep computational complexity proportional to sequence length, making it suitable for the high‑dimensional, irregularly sampled data common in financial applications.

The authors position EXAONE Finance as a domain‑specific alternative that retains the strong zero‑shot performance of large TSFMs while addressing the scalability and data‑availability constraints of financial forecasting. The full report is available on arXiv.

*Source: arXiv cs.AI, 2026-09-07*