LG Brings Foundation Models to Financial Time Series With EXAONE Forecast

For three years the race in time-series foundation models has been a race to be general: one model, any domain, zero shot. LG AI Research has published a 17-page argument that in finance, generality is the problem.

EXAONE Forecast for Finance — posted to arXiv on August 4, 2026 as arXiv:2609.04239 and surfacing in the September cs.AI listings — is a 202-million-parameter forecasting model from a twelve-person team at LG AI Research led by Seunghan Lee. It is small by 2026 standards, it contains no attention layers at all, and on the benchmark LG released the same day it outranks all 43 publicly available time-series foundation models (TSFMs) it was measured against, including Google’s TimesFM, Amazon’s Chronos, Salesforce’s Moirai and IBM’s Tiny Time Mixers.

The pitch starts with a complaint about the field’s architecture of choice. Today’s TSFMs, the report argues, “largely rely on self-attention backbones whose computational cost grows quadratically with sequence length and variate count,” and they “assume fully observed inputs” while being pretrained on corpora that miss how markets actually behave. Financial panels break all three assumptions at once: long, thousand-channel, and full of holes — halted tickers, market closures, quarterly fundamentals sampled against daily prices.

An attention-free backbone, on purpose

EXAONE Finance throws out self-attention for two linear-time operators: a causal 1D convolution for mixing across time, and a group-aware pooling MLP for mixing across variates. The complexity argument is the whole design — temporal mixing costs O(L·d·c·k) and variate mixing O(C·d), against the O(L²d) time-axis and O(C²d) variate-axis attention used by the incumbents. The released configuration is 12 blocks, hidden dimension 1024, patch size 16, maximum input length 512, and 21 quantile heads. It decodes an entire horizon in a single forward pass — no autoregressive roll-out, and no key/value cache, because there is no attention to cache.

The second intervention is masked context augmentation, which deliberately blanks contiguous spans of history during pretraining so the model learns to forecast through gaps rather than choke on them. The pretraining corpus spans 13 financial subsets across four scopes, dominated by 96,522 US individual-asset series (167.4 million observations) and 19,538 sector-ETF series (30.3 million), with foreign exchange, commodities, crypto-assets, fixed income and macro indicators filling out the rest, blended with synthetic KernelSynth data and a general-domain slice of GIFT-Eval. The report never states a headline total; summing its own figures yields roughly 117,000 series and 200 million observations.

Evaluation runs on FinVerse, LG’s companion benchmark: 116,897 series and 171.1 million observations, of which 60,232 series and 17.4 million observations are scored, with testing beginning January 1, 2020. FinVerse grades on three tiers — point-forecast accuracy (hit ratio and MASE variants), cross-sectional ranking (information coefficient), and long-only portfolio backtests (return, Sharpe, drawdown). EXAONE Finance places first in all three, a perfect rank sum of 3, with average ranks of 6.53, 7.45 and 7.43. The nearest baselines, Chronos-2 (Synthetic) and Reverso Small, tie at 14; TimesFM-2.5-200M lands at 32, Chronos-2 at 30, Moirai-1.1-R Large at 78, and IBM’s Tiny Time Mixers variants at 99, 101 and 117. Head-to-head, EXAONE beats every one of the 43 baselines, with per-opponent win rates from 0.51 to 0.90.

The most quotable finding is about size. “EXAONE Finance sits on the Pareto frontier: at only 202M parameters it attains the best overall rank, outperforming models more than an order of magnitude larger,” the authors write, pointing at billion-scale Toto-2.0 variants. Then the sharper claim: “a larger model size does not reliably improve the rank, suggesting that the scaling law common in general-domain forecasting does not yet hold clearly in finance.”

Why It Matters

This is the domain-specialization thesis stated about as cleanly as it can be. The general-TSFM camp bets that enough heterogeneous data plus scale produces a forecaster that transfers anywhere. LG is betting that finance is different enough — non-stationary, regime-switching, low signal-to-noise, structurally full of missing values — that a purpose-built 202M model beats a 2.5B generalist. The company is explicit that this is strategy, not just science. “The focus of global tech companies is rapidly shifting from general-purpose language models toward industry-specific AI that can solve problems in actual business settings,” Lim Woo-hyung, co-head of LG AI Research, told the Korea Herald in August, adding that “data accumulated in manufacturing and other industries can serve as a differentiated competitive advantage for Korea in AI.”

But the result deserves the standard caveat about home-field benchmarks. FinVerse and EXAONE Finance were submitted the same day by the same twelve authors. FinVerse is a genuine contribution — it reports only a 0.40 Pearson correlation between its rankings and GIFT-Eval’s, itself an argument that general benchmarks mislead in finance — but a lab that designs both the exam and the student earns extra scrutiny, and no independent replication exists yet.

There is also a reporting gap. Despite training with a quantile loss and emitting 21 quantiles, the report publishes no absolute MASE, CRPS, Sharpe or return figures for its own model; the results table contains ranks only, and CRPS is never mentioned. Nor are there measured latency numbers — the efficiency case is entirely asymptotic. FinVerse, to its credit, admits its own hardest problem: it does not model point-in-time data availability, so “the benchmark may treat information as available earlier than it would have been in a real forecasting setting” — the look-ahead bias that has flattered quantitative backtests for decades.

Watch three things. First, whether LG releases weights; EXAONE has an open-weight track record, but a finance model with commercial partners may not follow it. Second, whether anyone reproduces the FinVerse result independently, or whether the Chronos and TimesFM teams answer with finance-specific variants. Third, whether ranks convert into money: LG says it is already working with Koscom and the London Stock Exchange Group on Korean and US market analysis, and running proofs of concept on loan delinquency forecasting and suspicious-transaction detection. Benchmark first place is cheap. A Sharpe ratio that survives contact with a live book is not.