Authors describe EXAONE Finance forecasting model

A technical report on arXiv presents EXAONE Forecast for Finance, described by its authors as a financial time-series foundation model built for forecasting across equities, foreign exchange, commodities, crypto-assets, fixed income and macroeconomic indicators. The paper says existing time-series foundation models are often aimed at general data and can struggle with the long, many-variable and intermittently missing sequences common in finance.

To address that, the authors replace self-attention with two linear-time components: a causal 1D convolution for temporal mixing and a group-aware pooling multilayer perceptron for mixing across variables. They also describe a masked context augmentation method intended to make the model more robust when chunks of market data are missing.

Claims and limits

The authors say the model was pretrained on a large-scale financial corpus and ranks first on all three tiers of the FinVerse benchmark: point-forecast accuracy, cross-sectional asset ranking and portfolio profitability. If those results hold up, the design could matter for firms that want broader market coverage without the heavier compute costs associated with some attention-based systems.

The limits are important. This is an arXiv preprint and technical report, not a peer-reviewed paper. The evidence here is the abstract alone, so details such as training data construction, benchmark design, implementation choices and comparison baselines are not independently tested in the supplied material. In short, promising on paper, which is where many things are impeccable.