* EXAONE Forecast for Finance (EXAONE Finance) is a financial time series foundation model.
* The model was submitted on 4 August 2026.
* EXAONE Finance replaces self-attention with a causal 1D convolution for temporal mixing.
* EXAONE Finance replaces self-attention with a group-aware pooling multi-layer perceptron (MLP) for variate mixing.
* Masked context augmentation is used during training to handle missing data spans.
* Pretraining data includes equities, foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators.
* The model was evaluated using the FinVerse financial forecasting benchmark.
* Performance metrics include point-forecast accuracy, cross-sectional asset ranking, and portfolio profitability.
* EXAONE Finance ranked first across all three FinVerse evaluation tiers.

EXAONE Finance is a specialized foundation model designed to overcome the computational and data-specific hurdles of financial time series forecasting. While general-domain models typically rely on self-attention mechanisms that scale quadratically—creating inefficiencies with long sequences and numerous variables—this architecture utilizes linear-time operators. Specifically, it employs causal 1D convolutions and group-aware pooling MLPs to handle temporal and variate mixing more efficiently.
To address the frequent issue of missing data in financial markets, the model incorporates masked context augmentation during its pretraining phase. The training corpus is broad, spanning multiple asset classes including crypto-assets, commodities, and macroeconomic indicators. Testing on the FinVerse benchmark indicates state-of-the-art performance in accuracy, asset ranking, and overall portfolio profitability, suggesting a significant improvement over previous general-purpose time series models in a financial context.

This scholarship presents a strategic pivot from the "attention-is-all-you-need" paradigm toward linear-time operators for specialized domains. By replacing self-attention with causal 1D convolutions and group-aware pooling MLPs, the authors address the "quadratic bottleneck" that often renders large-scale financial panels computationally prohibitive.
From a methodology standpoint, a peer reviewer would scrutinize the "masked context augmentation." While intended to improve robustness to missing data, the extent to which this simulates real-world "intermittent observation" versus simple noise injection remains a critical point of inquiry. Additionally, the claim of state-of-the-art performance across three distinct tiers of the FinVerse benchmark is bold; the proportionality of these gains—whether they are marginal or transformative—is not detailed in the abstract. The novelty is justified by the specific tailoring to financial dynamics, moving away from the assumption of fully observed inputs common in general TSFMs.
If these findings hold, the implication is a shift toward "domain-specific foundation models" where architectural simplicity (linear-time operators) outperforms general-purpose complexity when paired with high-quality, domain-aligned corpora. This suggests that for high-noise, high-dimensional environments like finance, specialized inductive biases are more valuable than raw attention mechanisms.
To further validate these claims, a study comparing EXAONE Finance against a self-attention model trained on the exact same financial corpus—rather than general-domain models—would isolate whether the success stems from the architecture or the data.
Bridge Questions:
1. How does the model perform during "black swan" events where historical patterns in the pretraining corpus are rendered obsolete?
2. Does the reduction in computational complexity lead to a measurable decrease in training energy or latency without sacrificing long-term predictive horizon?