EXAONE Forecast for Finance: A Linear-Time Foundation Model for Financial Time Series

EXAONE Forecast for Finance: A Linear-Time Foundation Model for Financial Time Series

Submitted on 4 Aug 2026 | arXiv:2609.04239 (cs.AI) | Subjects: Artificial Intelligence (cs.AI), Machine Learning (cs.LG)

Authors: Seunghan Lee, Jaehoon Lee, Jun Seo, Tae Yoon Lim, Dongwan Kang, Hwanil Choi, Minjae Kim, Sungdong Yoo, Junhyeok Kang, Sangjun Han, Soonyoung Lee, Wonbin Ahn

Abstract

This technical report introduces EXAONE Forecast for Finance (EXAONE Finance), a financial time series (TS) foundation model (TSFM) optimized for financial forecasting. Recent TSFMs achieve strong zero-shot performance via large-scale pretraining, yet they are typically built for general-domain data and rely on self-attention backbones, whose computational cost scales quadratically with sequence length and variate count. Additionally, these models assume fully observed inputs and are pretrained on corpora that do not capture the unique dynamics of financial markets. Such limitations impede their applicability in finance, where long, many-channel, intermittently observed panels are the norm.

To overcome these hurdles, EXAONE Finance adopts an attention-free architecture, replacing self-attention with two straightforward but powerful linear-time operators: a causal 1D convolution for temporal mixing and a group-aware pooling multi-layer perceptron (MLP) for variate mixing. Furthermore, a masked context augmentation mechanism exposes the model to contiguous missing spans during training, bolstering robustness against the pervasive missingness in financial markets. EXAONE Finance is pretrained on a large-scale financial corpus that spans equities, foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators. On FinVerse—a comprehensive financial forecasting benchmark covering diverse asset classes—EXAONE Finance achieves state-of-the-art performance, ranking first across all three evaluation tiers: point-forecast accuracy, cross-sectional asset ranking, and portfolio profitability.

Context and Motivation

The financial industry increasingly relies on machine learning to forecast asset prices, manage risk, and optimize portfolios. As of 2026, the demand for models that handle complex, high-dimensional, and irregular time series is more pronounced than ever, given the proliferation of high-frequency trading and alternative data sources. General-purpose TSFMs, while powerful, falter on financial data due to their architectural inefficiencies and training data biases.

Key challenges:

• Scalability: Self-attention mechanisms incur quadratic costs, making them impractical for the long sequences and large variate counts typical in finance.
• Missingness: Financial data often have gaps due to trading halts, holidays, or reporting delays, yet most models assume complete observations.
• Domain specificity: Financial markets exhibit non-stationary, high-noise dynamics that are underrepresented in generic pretraining corpora.

Architectural Innovations

EXAONE Finance’s architecture is designed to address these challenges head-on:

1. Causal 1D Convolution for Temporal Mixing: Instead of full attention, the model uses a causal convolution to capture local temporal dependencies efficiently. This linear-time operator processes long sequences without the quadratic overhead, making it suitable for high-frequency data.
2. Group-Aware Pooling MLP for Variate Mixing: To handle interactions among multiple financial series (e.g., different stocks or assets), the model employs a pooling MLP that operates in linear time. This mechanism groups related variates, pooling information across them in a way that preserves scalability, even with many channels.
3. Masked Context Augmentation: During training, the model is exposed to artificially induced missing spans, teaching it to infer absent values from context. This boosts its resilience to the intermittent observation patterns common in financial markets.



Pretraining and Data




EXAONE Finance is pretrained on a diverse financial corpus that includes:

• Equities (stock prices, indices)
• Foreign exchange (currency pairs)
• Commodities (e.g., oil, gold)
• Crypto-assets (e.g., Bitcoin, Ethereum)
• Fixed income (bonds, yields)
• Macroeconomic indicators (GDP, inflation, employment)



This breadth ensures the model learns cross-market patterns and can generalize to unseen assets or regions.




Performance Evaluation




On FinVerse, a benchmark designed to test financial forecasting across multiple asset classes, EXAONE Finance excels:

• Point-Forecast Accuracy: Achieves the lowest error metrics, demonstrating precise predictions.
• Cross-Sectional Asset Ranking: Correctly orders assets by expected performance, crucial for portfolio construction.
• Portfolio Profitability: In simulated trading, the model generates higher risk-adjusted returns than competing approaches, highlighting its practical utility.



Conclusion and Outlook




EXAONE Finance represents a significant step forward in applying TSFMs to finance. Its linear-time architecture enables scalability without sacrificing accuracy, while its focus on missingness and domain-specific data addresses long-standing issues in the field. As we move further into 2026, such models are poised to become integral tools for analysts, traders, and risk managers, enabling more robust and efficient decision-making in increasingly complex markets.




Future directions could include extending the model to multi-horizon forecasting, integrating unstructured data (e.g., news sentiment), or refining the masking strategy to mimic real-world data gaps even more closely.




Cite as: arXiv:2609.04239 [cs.AI]

via ArXiv AI