EXAONE Finance Sets State-of-the-Art Results in Financial Forecasting

Summary

The technical report introduces EXAONE Forecast for Finance, or EXAONE Finance, a time series foundation model designed for financial forecasting. The authors argue that general-purpose time series models are poorly suited to finance because self-attention becomes expensive for long, many-channel sequences, the models usually expect fully observed inputs, and their pretraining data does not reflect financial market dynamics. EXAONE Finance replaces self-attention with two linear-time components: a causal one-dimensional convolution for temporal mixing and a group-aware pooling MLP for mixing information across variables. It also uses masked context augmentation, training on contiguous missing spans so the model is more robust to the intermittent observations common in financial panels. Pretraining covers equities, foreign exchange, commodities, crypto-assets, fixed income and macroeconomic indicators. On the FinVerse benchmark, which spans multiple asset classes, the model ranks first in point-forecast accuracy, cross-sectional asset ranking and portfolio profitability. The report presents these results as state-of-the-art performance across all three evaluation tiers, while the available abstract does not provide numerical scores or details of the baselines.