EXAONE Finance: A Deep Dive into the Time Series Foundation Model Built for Finance

Introduction: Why General-Purpose Time Series Models Struggle with Financial Forecasting

In recent years, Time Series Foundation Models (TSFMs) have demonstrated strong performance on zero-shot forecasting tasks through large-scale pre-training. These models borrow the pre-train-then-fine-tune paradigm from Large Language Models (LLMs), performing self-supervised pre-training on massive time series data to learn universal temporal representations. Notable examples include Google's TimesFM, Amazon's Chronos, and Salesforce's Moirai. However, when these general-purpose models are applied to the financial domain, they consistently fall short. A new technical report from LG AI Research, titled EXAONE Forecast for Finance (referred to as EXAONE Finance), directly addresses this pain point by introducing a foundation model purpose-built for financial time series forecasting.

Why do general-purpose TSFMs struggle in financial settings? The report identifies three key limitations — each of which EXAONE Finance aims to overcome.

Three Key Limitations of General-Purpose Time Series Models in Finance

Computational Cost Grows Quadratically with Sequence Length

Most mainstream time series foundation models rely on self-attention as their core backbone. Self-attention is the central component of the Transformer architecture, working by computing relevance scores between every position in a sequence and all other positions to capture global dependencies. However, this "all-to-all" computation results in O(N²) time and space complexity, where N is the sequence length. While the research community has proposed approaches like Linformer, Performer, and FlashAttention to mitigate this issue, these are typically approximate optimizations of the attention mechanism rather than fundamental architectural replacements.

Financial panel data is characterized by long sequences and many channels — imagine simultaneously tracking the long-term historical data of hundreds or thousands of assets across stocks, foreign exchange, and commodities. When dealing with thousands of assets, each with years or even decades of daily or higher-frequency data, N can reach tens or even hundreds of thousands, making the computational burden of self-attention architectures prohibitive.

Assumption of Complete Input Data Leads to Poor Real-World Adaptation

General-purpose models typically assume inputs are "fully observable." But real financial markets are riddled with intermittent missingness — trading halts, non-trading days, and data collection interruptions result in large contiguous gaps in financial panel data.

It's worth emphasizing that the missingness patterns in Financial Panel Data are fundamentally different from the random missingness found in sensor or weather data. Financial missingness tends to be "structured" and "intermittent": individual stocks may have no trading data for consecutive days due to suspension; newly listed companies have no data at all before their IPO; different countries' exchanges observe different holidays, making cross-market data alignment difficult; and certain alternative data sources may experience collection interruptions during specific periods. This non-random, block-wise missingness means that simple interpolation methods often introduce bias. Models trained on idealized complete data suffer significant performance degradation when confronted with real-world missing values.

Pre-training Data Disconnected from Financial Market Dynamics

The most fundamental issue is that general-purpose TSFMs are pre-trained primarily on generic time series data, which fails to capture the unique dynamics of financial markets. Financial time series exhibit a set of distinctive statistical properties that set them apart from other domains: Volatility Clustering — first formally described by Nobel laureate Robert Engle in his ARCH model — means that large price swings tend to be followed by more large swings, and calm periods tend to persist; additionally, there are heavy-tailed distributions (extreme values occur far more frequently than a normal distribution would predict), leverage effects (declines trigger greater volatility than advances), alternating mean-reversion and momentum regimes, as well as complex dynamics like cross-asset co-movements and macro factor drivers. These properties are fundamentally different from the statistical characteristics of generic time series data like temperature or electricity load, and this "innate deficiency" at the data level severely constrains the performance of general-purpose models on financial forecasting tasks.

Three Core Innovations of EXAONE Finance

Attention-Free Architecture Achieves Linear Complexity

To address the computational bottleneck, EXAONE Finance boldly discards self-attention in favor of two simple yet efficient linear-time complexity operators:

• Causal 1D Convolution: Handles temporal mixing with linear complexity to capture temporal dependencies. Causal convolution constrains the convolution operation to access only the current and previous time steps (achieved through left-side zero-padding), strictly preserving the causality requirement essential in time series forecasting. Its computational complexity is O(N×k), where k is a fixed kernel size, yielding overall linear complexity O(N). By stacking multiple causal convolution layers or using dilated convolution, the receptive field can be expanded exponentially to capture long-range dependencies — a proven strategy validated by classic architectures like WaveNet and TCN.

• Group-Aware Pooling MLP: Handles variate mixing to efficiently model interactions across multiple channels. In multivariate time series forecasting, "variate mixing" is crucial — cross-asset co-movements in financial markets (e.g., the impact of oil price fluctuations on airline stocks, or the transmission of USD movements to emerging market assets) contain rich predictive signals. The Group-Aware Pooling MLP pools variables by group before performing MLP-based interactions, preserving the ability for cross-asset information exchange while keeping computational complexity at a linear level — an elegant trade-off between efficiency and expressiveness.

This design reduces computational complexity from quadratic to linear, enabling the model to comfortably handle the "long sequences, many channels" data characteristics typical of the financial domain — a critical capability when simultaneously modeling massive numbers of assets in real-world scenarios.

Masked Context Augmentation Strategy for Handling Missing Data

To address the pervasive missingness in financial data, EXAONE Finance introduces a Masked Context Augmentation training strategy. This approach draws inspiration from masked language modeling in NLP (such as BERT) and masked autoencoders (MAE) in computer vision, but with a crucial domain-specific adaptation: unlike the random masking used in those methods, EXAONE Finance specifically employs contiguous block masking to precisely simulate the intermittent missingness patterns that actually occur in financial data.

During training, the model actively encounters contiguous missing segments and is forced to generate accurate predictions even when portions of the input are masked. This "lossy input, complete output" training paradigm compels the model to learn to reason with incomplete contextual information, thereby developing robust prediction capabilities even under data incompleteness. This "train on missingness to handle missingness" philosophy directly addresses the ubiquitous intermittent observation problem in financial markets.

Large-Scale Financial Pre-training Corpus Covering Multiple Asset Classes

For pre-training data, EXAONE Finance uses a large-scale financial corpus with coverage extending far beyond equities to include:

• Foreign exchange
• Commodities
• Crypto-assets
• Fixed income
• Macroeconomic indicators

This broad asset coverage enables the model to truly understand cross-asset dynamics in financial markets, rather than remaining confined to surface-level patterns within a single domain. For example, fixed income data helps the model learn how interest rate environments affect equity valuations, macroeconomic indicators provide leading signals of economic cycles, and crypto-asset data exposes the model to more extreme volatility patterns and complex market microstructure. This multi-asset pre-training gives the model the opportunity to internalize the intricate transmission mechanisms between various asset classes in financial markets.

FinVerse Benchmark: SOTA Across All Three Metrics

EXAONE Finance was evaluated on FinVerse, a comprehensive financial forecasting benchmark covering multiple asset classes, achieving state-of-the-art (SOTA) results. FinVerse is a comprehensive benchmark suite specifically designed to assess time series model performance in the financial domain. Its very existence marks a paradigm shift in financial time series forecasting research — from "competing on generic metrics" to "competing on practical value" — since traditional time series benchmarks (such as ETTh, Weather, Traffic, etc.) primarily come from energy, meteorology, and transportation domains and cannot reflect the unique challenges of financial data.

Notably, EXAONE Finance ranked first across all three core evaluation dimensions:

1. Point-Forecast Accuracy: Measures the direct numerical prediction accuracy for future values, corresponding to signal generation needs in actual trading;
2. Cross-Sectional Asset Ranking: Measures the ability to rank the relative performance of different assets, corresponding to the most critical stock selection and asset allocation scenarios in quantitative investing — in practice, accurately determining "which stock will outperform another" is often more valuable than predicting "how much a particular stock will rise tomorrow";
3. Portfolio Profitability: Directly evaluates the model's practical value from an investment return perspective, serving as the ultimate test of whether predictive capability can translate into real economic returns.

From numerical accuracy to asset ranking to ultimate investment returns, EXAONE Finance leads across the entire chain, meaning it not only excels on academic metrics but also demonstrates the potential for deployment in real-world investment decision-making.

Trends and Outlook for Domain-Specific Foundation Models

The value of EXAONE Finance lies in its refusal to blindly follow the mainstream approach of "bigger attention, more parameters." Instead, it starts from the actual constraints of the financial domain and makes three targeted architectural and data choices: replacing self-attention with linear operators to solve the efficiency problem, using masked augmentation to solve the missingness problem, and using multi-asset corpus to solve the domain adaptation problem.

This also reflects a broader industry trend: as foundation models enter vertical industries, "general-purpose large models" may not be the optimal solution, and Domain-Specific Foundation Models are demonstrating unique advantages. This trend is already evident across multiple fields: Bloomberg's BloombergGPT in finance, Google's Med-PaLM and BioMedLM in healthcare, and Meta's Galactica in scientific computing. The underlying logic is that different vertical domains each possess unique data distributions, specialized knowledge systems, and compliance constraints, making the vision of "one model to rule all scenarios" difficult to realize. Domain-specific models can more precisely address the right problems through targeted customization at three levels: architecture, data, and training strategy.

Of course, as a technical report, questions remain about EXAONE Finance's true generalization capability, its performance in live trading environments, and whether it will be open-sourced and how it might be deployed. Regardless, EXAONE Finance provides an inspiring technical blueprint for "how time series foundation models can serve financial forecasting" and brings fresh thinking to the development direction of financial AI models.

Key Takeaways