Source-linked AI summary
EXAONE Forecast for Finance
Seunghan Lee, Jaehoon Lee, Jun Seo, Tae Yoon Lim, Dongwan Kang, Hwanil Choi, Minjae Kim, Sungdong Yoo, Junhyeok Kang, Sangjun Han, Soonyoung Lee, Wonbin Ahn
TL;DR
Existing TSFMs are poorly matched to financial forecasting because they incur costly attention, assume fully observed inputs, and underrepresent financial dynamics in pretraining. EXAONE Finance addresses these gaps with attention-free linear-time mixing, missing-span augmentation, and broad financial pretraining, ranking first across all three FinVerse evaluation tiers.
Problem
Financial forecasting involves long, many-channel, intermittently observed panels, while existing TSFMs face quadratic attention cost, fully observed-input assumptions, and financial domain mismatch.
Method
EXAONE Finance uses causal convolution for temporal mixing, group-aware pooling for variate mixing, masked-context augmentation, and a broad financial pretraining corpus.
Results
EXAONE Finance ranks #1 across point accuracy, cross-sectional ordering, and portfolio performance on FinVerse, with rank-sum 3 versus 14 for the strongest baseline.
Takeaways & Limitations
The model provides a finance-specific TSFM that combines linear-time processing and missing-data handling with strong performance across diverse financial forecasting objectives.
Takeaways & Limitations
Evaluation is zero-shot on a financial benchmark; parameter-efficient fine-tuning, ensembling, and broad general-domain benchmarking remain untested or future work.
Abstract
from arXiv · showhide
This technical report presents EXAONE Forecast for Finance (EXAONE Finance), a financial time series (TS) foundation model (TSFM) tailored to financial forecasting. Recent TSFMs achieve strong zero-shot performance through large-scale pretraining. However, they are primarily developed for general-domain TS and largely rely on self-attention backbones whose computational cost grows quadratically with sequence length and variate count. Moreover, they assume fully observed inputs and are pretrained on corpora that fail to capture the unique dynamics of financial markets. These limitations hinder their applicability to finance, where long, many-channel, intermittently observed panels are common. To address these challenges, EXAONE Finance adopts an attention-free architecture, replacing self-attention with two simple yet effective linear-time operators: 1) a causal 1D convolution for temporal mixing and 2) a group-aware pooling multi-layer perceptron (MLP) for variate mixing. Furthermore, a masked context augmentation exposes the model to contiguous missing spans during training, improving robustness to the missingness pervasive in financial markets. EXAONE Finance is pretrained on a large-scale financial corpus covering not only equities but also foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators. On FinVerse, a financial forecasting benchmark covering diverse asset classes, EXAONE Finance attains state-of-the-art performance, ranking first across all three evaluation tiers---point-forecast accuracy, cross-sectional asset ranking, and portfolio profitability.
1 Introduction
Financial forecasting requires models that handle diverse, long, intermittently observed asset panels, while existing TSFMs face computational, missingness, and domain-mismatch gaps. EXAONE Finance addresses these needs with finance-specific design and achieves strong benchmark performance across evaluation tiers.
- Challenges: Financial TSFMs face quadratic self-attention cost, missing observations, and pretraining corpora that underrepresent financial dynamics.These gaps are especially relevant for long, many-channel financial panels.
- Scope: Financial forecasting spans equities, ETFs, foreign exchange, commodities, crypto-assets, fixed income, and macroeconomic indicators with heterogeneous dynamics.These assets are often observed jointly as multi-channel panels.
- Positioning: EXAONE Finance is the only compared TSFM that is both attention-free and pretrained on financial data.The comparison identifies its backbone and pretraining domain as distinct from the other listed TSFMs.
- Approach: EXAONE Finance combines attention-free linear-time temporal and variate mixing with masked-context training and a financial pretraining corpus.Its design uses causal convolutions, a group-aware pooling MLP, contiguous-span masking, and broad financial data.
- Evaluation: EXAONE Finance outperforms five representative TSFMs on the FinVerse benchmark.The benchmark comparison is summarized in Figure 1.
2 Related Works
Related TSFM research scales zero-shot forecasting through large heterogeneous corpora and varied architectures, but most models rely on self-attention. EXAONE Finance instead removes attention for finance-oriented forecasting.
- Time-series foundation models: Recent TSFMs pretrain on large heterogeneous time-series corpora to support zero-shot forecasting across domains.Examples include Chronos, TimesFM, Moirai, Moirai-MoE, and Toto-2.0.
- Attention-based models: Most contemporary TSFMs rely on self-attention, whose cost grows quadratically with sequence length and variate count.EXAONE Finance departs from this common design while retaining a patch-based approach.
- Efficient alternatives: Convolutional sequence models use causal convolutions for large receptive fields at linear cost and can compete with transformers on long-horizon forecasting.MLP-based models also provide strong simple baselines through linear or token-/channel-mixing layers.
- Pretraining corpora: TSFM pretraining corpora combine real archives, synthetic generators, or both, ranging from multi-domain collections to hundreds of billions of time points.KernelSynth supplies synthetic patterns, while sources such as LOTSA, GIFT-Eval, and Time-300B provide real data at different scales.
3 Preliminaries
The forecasting setup maps recent context windows, including grouped and partially missing financial series, to future horizons and probabilistic forecasts. EXAONE Finance processes this context through patch-based attention-free blocks with masked-context training.
- Forecasting setup: A univariate forecast uses the latest L observations as context to predict the next H values in the forecast horizon.The context is xctx = (x_t−L+1,...,x_t), and the horizon is y = (x_t+1,...,x_t+H).
- Multivariate series: Grouped multivariate series contain C aligned variates sharing timestamps, with group identifiers specifying which variates may exchange information.A univariate series is the special case C = 1.
- Missingness: An observation mask marks missing or padded entries, and the model forecasts through missing spans without imputation.Financial missingness can arise from market closures, trading halts, and unreported periods.
- Probabilistic forecasting: The probabilistic objective produces Q quantile forecasts for specified quantile levels and future horizons.Each predicted value estimates the corresponding q-th quantile of the h-step-ahead outcome.
- Model pipeline: EXAONE Finance normalizes and patches context inputs, then applies attention-free encoder blocks and a probabilistic forecasting head.Masked-context augmentation trains the model to forecast through contiguous missing spans encountered at inference.
4 EXAONE Finance
EXAONE Finance combines an attention-free input-to-forecast pipeline with causal temporal mixing, group-aware variate mixing, and training designed for missing spans. Its linear-time mixers, probabilistic forecasting head, and single-pass inference target long, multivariate financial series.
- Input pipeline: The pipeline normalizes and transforms the context, patches it with values and observation masks, embeds the patches, applies encoder blocks, and decodes forecasts.Instance normalization uses observed entries, while an arcsinh transform compresses extreme values and preserves sign.
- Temporal mixing: Temporal mixing replaces self-attention with stacked causal 1D convolutions, so each token depends only on current and past tokens.Left-only padding preserves causality, and a residual connection follows the convolution stack.
- Variate mixing: Variate mixing uses a group-aware pooling MLP that exchanges information only within groups through each variate’s group mean.The operator is permutation-invariant within groups and reduces to a per-variate transformation for singleton groups.
- Masked-context augmentation: Masked-context augmentation randomly removes contiguous input spans during training and excludes them from normalization and loss calculations.The masking fraction is ρ = 0.25, and the augmentation is applied only during training; unlike reconstruction masking, it masks inputs rather than targets.
- Forecasting and configuration: The forecasting head predicts Q quantile trajectories with pinball loss, while loss masking permits multivariate inputs but scores designated target channels.The released configuration uses 21 quantile levels and patch size 16.
- Inference and complexity: Both mixers have linear complexity in their respective axes, and inference decodes the full horizon in one forward pass rather than autoregressively.The convolution costs O(L d c k), group-aware pooling costs O(C d), and the model supports arbitrary horizons by varying output tokens.
5 Experiments
FinVerse evaluates forecasting across diverse financial assets using three tiers covering point accuracy, cross-sectional ordering, and portfolio value. EXAONE Finance ranks first across all three tiers and remains strongest across model-size and asset-scope analyses.
- 5.1 Pretraining Dataset: The pretraining corpus contains 13 financial subsets across four scopes and spans asset classes beyond equities.The financial source is combined with synthetic KernelSynth and general-domain GIFT-Eval data.
- 5.2 Evaluation Protocol: Tiers and Metrics: FinVerse evaluates scale-free point accuracy, cross-sectional asset ordering, and realized portfolio performance across three tiers.Tier 1 uses change-based accuracy and hit-rate metrics; Tier 2 uses Spearman information coefficient; Tier 3 backtests top-K%, equal-weight, long-only portfolios.
- 5.2 Evaluation Protocol: Tiers and Metrics: The benchmark uses frequency-aware windows and change-based targets so predicted and realized changes share the same future point.Tiers 1–2 vary the change period and forecast position, while the same changes underlie all three tiers.
- 5.3 Baselines and Aggregation: Geometric-mean rank aggregation rewards consistent performance across heterogeneous evaluation cells and penalizes a single poor rank proportionally.The headline rank-sum adds the three per-tier placements, with lower values better.
- 5.4 Main Results: EXAONE Finance ranks #1 in all three tiers with rank-sum 3, ahead of the strongest baseline at rank-sum 14.The lead spans point accuracy, cross-sectional ordering, and realized portfolio performance rather than a single metric.
- 5.4 Main Results: At 202M parameters, EXAONE Finance attains the best overall rank and outperforms models more than an order of magnitude larger.It sits on the Pareto frontier, while larger model size does not reliably improve rank in the plotted finance results.
- 5.4 Main Results: EXAONE Finance wins head-to-head against all 43 baselines, with per-opponent win rates ranging from 0.51 to 0.90.The narrowest margins are against TiRex-2-pretrain and TiRex-1.1, which still trail it on every tier.
- 5.5 Per-Asset Analysis: EXAONE Finance scores in the 95–100 percentile on ten of eleven populated scope×tier cells and is top-3 in ten of eleven scope×tier panels.The per-scope comparison covers point accuracy, cross-sectional IC, and portfolio backtest across four data scopes.
6 Conclusion
EXAONE Finance combines linear-time temporal and variate mixing with masked-context training for financial forecasting. It ranks first across all three FinVerse tiers, while broader grouping, fine-tuning, ensembling, and general-domain evaluation remain future work.
- EXAONE Finance replaces self-attention with causal 1D convolution and group-aware pooling MLPs, achieving linear-time cost in sequence length and variate count.The model retains a patch-based, multi-quantile decoding interface.
- Masked-context augmentation makes EXAONE Finance robust to missingness pervasive in financial markets.
- EXAONE Finance ranks #1 in all three tiers of the FinVerse benchmark and is the only model winning head-to-head comparisons against every baseline.
- Future work includes richer variate grouping, parameter-efficient fine-tuning and ensembling, and broader general-domain evaluation.The reported results are zero-shot and evaluated on a financial benchmark.