LG-AI-Research/EXAONE-Finance-1.0

EXAONE Finance

EXAONE Finance is a time series foundation model (TSFM) built 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 — so cost grows linearly in both sequence length and variate count. A masked-context augmentation makes it robust to the missing spans that are pervasive in financial data, and it is pretrained on a synthetic financial corpus whose generator reproduces heavy tails, volatility clustering, jumps, regime shifts, and cross-asset dependence.

This repository holds the model weights. The inference code lives at LGAI-Research/EXAONE-Forecast and is released under a different license — see License.

• 📄 Technical report: arXiv:2609.04239 (also bundled here as EXAONE_Finance_v1.0_Technical_Report.pdf)
• 💻 Code: https://github.com/LGAI-Research/EXAONE-Forecast

Overview

How EXAONE Finance differs from other time series foundation models, in the two choices that set a TSFM's cost and its inductive bias — how it mixes information along time, and along variates — and in what it was pretrained on.

Model	Temporal mixing	Variate mixing	Financial-domain
Chronos-2, Moirai, Moirai-MoE, Toto, UniTS	Attention	Attention	❌
Chronos, Chronos-Bolt, TimesFM, Lag-Llama, TimeGPT, Timer, MOMENT, Sundial, TEMPO, ROSE, Time-MoE, PatchTST-FM, Kairos, YingLong, CleanTS, TabPFN-TS, VisionTS	Attention	—	❌
TiRex (xLSTM), TempoPFN, FlowState, Reverso	RNN / SSM	—	❌
TTM (TinyTimeMixer)	MLP	MLP	❌
EXAONE Finance (Ours)	CNN	MLP	✅

EXAONE Finance is the only attention-free convolutional TSFM in this comparison, and the only one pretrained on financial-domain data, which we generate synthetically. Both of its mixers are linear-time, so cost grows linearly in sequence length and in variate count rather than quadratically. A dash (—) means the model has no variate-mixing operator and forecasts each series independently.

_{CNN: convolutional neural network · RNN: recurrent neural network · SSM: state-space model · MLP: multi-layer perceptron}

Model Configuration

Parameters	202M
Encoder blocks	12
Hidden dimension / FFN	1024 / 4096
Temporal mixing	Causal 1D CNN — 512 channels, kernel 7, 2 conv layers per block
Variate mixing	Group-aware pooling MLP — hidden dimension 512
Patch size (input = output)	16
Context length	512
Maximum horizon	2048
Output	21 quantile levels (0.01, 0.05, 0.10, …, 0.95, 0.99)
Attention layers	none
Precision	float32

Evaluation Results

Zero-shot results on FinVerse, a financial forecasting benchmark spanning FX, commodities, crypto-assets, fixed income, equities, ETFs, and macroeconomic indicators. Models are ranked per (tier, spec, metric) cell; Avg. Rank is the geometric mean of per-cell ranks and Rank Sum is the sum of the three per-tier placements (lower is better). 44 models were scored on identical windows.

#	Model	Tier 1 (Point)	Avg. Rank	Tier 2 (IC)	Avg. Rank	Tier 3 (Portfolio)	Avg. Rank	Rank Sum
1	EXAONE Finance	1	6.53	1	7.45	1	7.43	3
2	Chronos-2 (Synthetic)	9	11.68	3	8.74	2	9.90	14
3	Reverso (Small)	6	9.60	2	8.64	6	13.04	14
4	TiRex-1.1	2	6.98	4	9.71	14	15.76	20
5	Chronos-2	8	10.14	13	16.12	9	15.17	30

The three tiers score complementary objectives: Tier 1 point accuracy (MASE, hit rate), Tier 2 cross-sectional ranking skill (information coefficient), and Tier 3 realized portfolio performance (return, Sharpe, volatility, maximum drawdown). EXAONE Finance ranks first in all three for a perfect rank sum of 3, ahead of the strongest baseline at 14, and is the only model that wins its head-to-head comparison against all 43 baselines (per-opponent win rate 0.51–0.90). At 202M parameters it sits on the Pareto frontier, outperforming models more than an order of magnitude larger. Full protocol and results are in the technical report.

Requirements

• Python ≥ 3.9
• PyTorch ≥ 2.0 (a CUDA build is recommended; CPU works for small workloads)

Quickstart

pip install "exaone-forecast[finance] @ git+https://github.com/LGAI-Research/EXAONE-Forecast.git"

import numpy as np
from exaone_forecast.finance import from_pretrained

fc = from_pretrained(device="cuda:0")     # downloads these weights on first use

series = [np.random.randn(500).cumsum() + 100 for _ in range(4)]

q = fc.predict(series, horizon=20)                       # (4, 21, 20) quantiles
yhat = fc.point(series, horizon=20)                      # (4, 20) median forecast
lo, hi = fc.interval(series, horizon=20, lower=0.1, upper=0.9)

Inputs may be a single 1D series, a list of 1D arrays of differing lengths, or a 2D (n_series, T) array. Feed values on their natural scale — the model normalizes internally and reads only the last context_length (512) points. Missing observations are supported: encode them as np.nan.

Available checkpoints

Name	File	Parameters
default	exaone-finance-1.0.safetensors	202M

config.json accompanies the weights and describes the architecture.

Intended use

Zero-shot probabilistic forecasting of financial and other real-valued time series, for research and educational purposes. Each series is forecast from its own recent history; no per-dataset training is required.

Limitations

• Results are reported on a financial benchmark; broad general-domain benchmarking is future work.
• Only zero-shot performance is reported. Fine-tuning and ensembling are untested.
• In this release every series is treated as an individual channel. The group-aware mixer supports richer cross-series grouping, which is not exercised by this checkpoint.
• Forecasts are statistical outputs, not financial advice. Do not use them as the sole basis for an investment decision.

License

The model weights in this repository are released under the EXAONE AI Model License Agreement 1.2 - NC (LICENSE), which limits use to non-commercial research and education.

The inference code at LGAI-Research/EXAONE-Forecast is licensed separately under the BSD-3-Clause-LG AI Research License, which permits commercial use. Installing that package does not grant commercial rights to the weights it downloads.

Third-party open source components and their licenses are listed in Notice.md.

Citation

@article{lgai2026exaonefinance,
  title   = {EXAONE Finance 1.0: An Attention-free Time Series Foundation Model for Financial Time Series},
  author  = {Lee, Seunghan and Lee, Jaehoon and Seo, Jun and Lim, Tae Yoon and
             Kang, Dongwan and Choi, Hwanil and Kim, Minjae and Yoo, Sungdong and
             Kang, Junhyeok and Han, Sangjun and Lee, Soonyoung and Ahn, Wonbin},
  journal = {arXiv preprint arXiv:2609.04239},
  year    = {2026}
}

The technical report introduces this model as EXAONE Forecast for Finance in its abstract. EXAONE Finance is the name used everywhere else, in the report, on the Hub, and in the code.

Contact

LG AI Research — https://www.lgresearch.ai