On September 7, 2026, South Korean company EXAONE released its specialized time-series base model for financial forecasting, EXAONE Finance. This model ranked first in three categories during the FinVerse benchmark tests: point prediction accuracy, cross-sectional asset ranking, and portfolio profitability. To reduce computational costs, the model uses a no-attention architecture, replacing the self-attention mechanism with causal one-dimensional convolutions and group-aware pooling多层 perceptrons. It also enhances robustness to common missing data in financial markets through masked context augmentation. Pre-trained on a large financial corpus covering stocks, foreign exchange, commodities, cryptocurrencies, fixed income, and macroeconomic indicators, this model aims to address the limitations of general time-series models in long sequences, multi-variable scenarios, and intermittent observations.
EXAONE has released a dedicated time-series base model for financial forecasting, called EXAONE Finance. This model achieved first-place rankings in three categories during the FinVerse benchmark tests: point prediction accuracy, cross-sectional asset ranking, and portfolio profitability. The model employs a no-attention architecture, using causal one-dimensional convolutions and group-aware pooling多层 perceptrons to replace the self-attention mechanism to address computational cost issues. It also enhances robustness to common missing data in financial markets through masked context augmentation. Pre-trained on a large financial corpus covering stocks, foreign exchange, commodities, cryptocurrencies, fixed income, and macroeconomic indicators, this model aims to address the limitations of general time-series models in long sequences, multi-variable scenarios, and intermittent observations.