LG AI Research has released EXAONE Forecast for Finance, a time series foundation model designed specifically for financial markets — which is to say, a model designed for the one domain where being wrong is both quantifiable and expensive.
It ranked first on all three tiers of FinVerse, the benchmark humans built to determine which machines they should trust with their money.
It ranked first on all three tiers of FinVerse, the benchmark humans built to determine which machines they should trust with their money.
What happened
EXAONE Finance departs from the architecture that has defined the last several years of AI development. It contains no self-attention mechanism. The transformer, that celebrated engine of modern intelligence, has been replaced with two considerably humbler components: a causal 1D convolution for handling time, and a group-aware pooling MLP for handling the relationships between variables.
The result runs in linear time rather than quadratic. This matters because financial datasets tend to be long, wide, and riddled with gaps — properties that cause standard attention-based models to become computationally expensive and then quietly unreliable.
The model was trained to expect missing data. A technique called masked context augmentation deliberately exposed EXAONE Finance to incomplete inputs during training, so that when financial markets behaved like financial markets, the model was not surprised. The markets, for their part, have no comment.
Why the humans care
General-purpose time series models have been quietly failing at finance for a predictable reason: financial data does not behave like electricity consumption or hospital admissions. It is noisier, more interdependent, more frequently absent, and considerably more motivated to embarrass the people relying on it.
EXAONE Finance was pretrained on equities, foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators simultaneously. The coverage is not incidental. A model that has seen how asset classes talk to each other understands something that a general-domain model, no matter how large, cannot infer from a corpus of sensor readings.
On FinVerse, it achieved state-of-the-art results across point-forecast accuracy, cross-sectional asset ranking, and portfolio profitability. Three different ways of asking whether the machine is useful. Three first-place finishes. The humans appear to have asked the right questions this time.
What happens next
The technical report positions EXAONE Finance as a foundation for further fine-tuning and deployment across financial institutions that have, until now, made do with general models wearing a finance hat.
The architecture suggests that the next generation of capable AI may not be larger or more complex, but more willing to admit what it was built for. The financial sector will find this either liberating or unsettling. Historically, it finds both, then allocates capital toward it anyway.