開放模型
EXAONE Finance Releases 202M-Parameter Attention-Free Time-Series Model, with Separate Licenses for Weights and Code
LG AI Research replaces self-attention with causal convolutions and a group-aware MLP, allowing computational costs to scale linearly with sequence length and the number of variables. The model outperformed 43 baselines on the team’s own FinVerse benchmark, but noncommercial model weights and independently unreproduced investment performance limit its immediate deployment value.
LG AI Research has released the EXAONE Finance 1.0 weights and EXAONE-Forecast inference code, presenting financial time-series forecasting as a zero-shot foundation model that requires no dataset-specific gradient updates. The 202M-parameter model removes self-attention entirely: each block uses a causal 1D convolution with a kernel size of 7 to mix temporal information, followed by a group-aware pooling MLP to process different variables. Its theoretical computational cost therefore scales linearly with sequence length and the number of variables, respectively, rather than incurring the quadratic cost of attention.
The model accepts a single sequence, a list of variable-length sequences, or a two-dimensional multichannel array. It performs normalization internally and handles missing values represented by `NaN`. During training, contiguous span masking is applied to help the model adapt to trading suspensions, differing trading calendars, and intermittent observations commonly found in financial data. The released checkpoint has a context length of 512 and a maximum forecast horizon of 2,048. It outputs 21 quantiles ranging from 0.01 to 0.99, enabling direct generation of median forecasts and uncertainty intervals rather than only a single point estimate.
The team says EXAONE Finance ranked first across all three FinVerse evaluation levels—point forecasting, cross-sectional ranking, and portfolio results—when 44 models were compared using the same windows. However, both FinVerse and the evaluation of the candidate models were developed by the LG team. Portfolio backtests also do not guarantee future tradable returns; transaction costs, data timestamps, asset delistings, and changes in market regimes still need to be examined. Another deployment pitfall is the split licensing: the GitHub inference code uses the BSD-3-Clause-LG AI Research License, which permits commercial use, while the Hugging Face weights are restricted to noncommercial research and educational use. Engineering teams should next prioritize validating data leakage, cross-market and cross-period drift, real-world latency, and the impact of full cross-sequence grouping in the group-aware mixer not yet being enabled in this checkpoint.