EXAONE Finance tops FinVerse with attention-free architecture
Researchers have presented EXAONE Finance, a financial time series foundation model built specifically for financial forecasting rather than adapted from general-purpose time series models. The paper argues that existing time series foundation models fall short in finance for two structural reasons: most rely on self-attention backbones, whose computational cost grows quadratically with both sequence length and the number of variates, and most are pretrained on general-domain corpora that do not reflect the specific dynamics of financial markets, including the long, many-channel, intermittently observed panels that are typical of trading and macroeconomic data.
To address the compute problem, EXAONE Finance replaces self-attention entirely with two linear-time operators: a causal 1D convolution that handles temporal mixing, and a group-aware pooling multi-layer perceptron (MLP) that handles mixing across variates (assets, indicators, or channels). Because both operators scale linearly rather than quadratically, the architecture is designed to remain tractable on the long, high-dimensional panels common in finance. To address missing data, a common feature of financial time series, the model is trained with a masked context augmentation technique that deliberately exposes it to contiguous missing spans during training, aiming to make it more robust when it later encounters similarly missing data in practice.
EXAONE Finance is pretrained on a large-scale financial corpus that spans multiple asset classes: equities, foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators. The authors evaluate it on FinVerse, a financial forecasting benchmark that itself covers diverse asset classes, and report that EXAONE Finance achieves state-of-the-art performance there, ranking first across all three evaluation tiers the benchmark measures: point-forecast accuracy, cross-sectional asset ranking, and portfolio profitability. The paper does not report specific numeric scores, name the competing models it was compared against, or give author names, institutional affiliations, model size, or training data volume.
Key facts
- EXAONE Finance is a financial time series foundation model built specifically for financial forecasting.
- It replaces self-attention with two linear-time operators: a causal 1D convolution for temporal mixing and a group-aware pooling MLP for variate mixing.
- A masked context augmentation trains the model on contiguous missing spans to improve robustness to the intermittent data common in financial markets.
- It is pretrained on a corpus spanning equities, foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators.
- On the FinVerse benchmark it ranks first across all three evaluation tiers: point-forecast accuracy, cross-sectional asset ranking, and portfolio profitability.
Why it matters
Financial time series differ from general-domain data in ways that break the assumptions most time series foundation models are built on: panels are long, span many correlated variates such as assets or indicators, and are frequently missing data outright. EXAONE Finance is pitched as a model designed around those specific constraints rather than a general-purpose model repurposed for finance, swapping the standard self-attention backbone for operators whose cost scales linearly instead of quadratically with sequence length and variate count.
Who it affects
The paper targets financial forecasting broadly rather than a single asset class: the pretraining corpus and the FinVerse evaluation both span equities, foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators. That breadth suggests the intended users are forecasting and quantitative research teams working across multiple markets rather than a single-instrument use case.
How to use it
The paper is a technical report describing the model's architecture and training setup rather than a product announcement. No release details, licence terms, pricing, or availability are given in the source material, so there is nothing yet to point to for actually running or accessing EXAONE Finance.
How solid is it
The central evidence is a first-place result across all three evaluation tiers of the FinVerse benchmark: point-forecast accuracy, cross-sectional asset ranking, and portfolio profitability. The source does not report the actual numeric scores behind that ranking, does not name which competing models were benchmarked against it, and does not describe the size or composition of FinVerse beyond calling it diverse across asset classes, so the claim currently rests on the paper's own reported ranking rather than independently verifiable figures.
Risks and caveats
The abstract gives no author names or institutional affiliations, no model size or parameter count, no training data volume, and no release date, which limits how much can currently be checked about who built the model and on what scale. Attention-free architectures trade the flexibility of self-attention for computational efficiency, and how that trade-off holds up outside the FinVerse benchmark, on live market data or under distribution shift, is not addressed in the available material.