FinCast: A Foundation Model for Financial Time-Series Forecasting Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Information and Knowledge

Zhuohang Zhu, Haodong Chen, Qiang Qu, Vera Chung · 2025

Read FinCast: A Foundation Model for Financial Time-Series Forecasting Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Information and Knowledge by Zhuohang Zhu, Haodong Chen, Qiang Qu, Vera Chung (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-19609.txt.

Introduction

Zhuohang Zhu, Haodong Chen, Qiang Qu, and Vera Chung present a foundation model for forecasting financial series in a 2025 study. Markets shift patterns over time, differ across stocks, commodities, and futures, and arrive at resolutions from seconds to weeks. Models trained narrowly overfit and fail when conditions change.

FinCast aims at one model that handles that variety directly. A foundation design pretrains broadly so new forecasting jobs need less custom work. Stable forecasts support economic stability, policy choices, and investment planning. The study frames the three sources of pattern shift and builds the model to answer each of them.

Catalog record

CREATORZhuohang Zhu, Haodong Chen, Qiang Qu, Vera Chung
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-19609

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

FinCast: A Foundation Model for Financial Time-Series Forecasting Conference: Proceedings of the 34th ACM International Conference on Information and Knowledge Management; November 10–14, 2025; Seoul, Republic of Korea Proceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM ’25), November 10–14, 2025, Seoul, Republic of Korea DOI: 10.1145/3746252.3761261 ISBN: 979-8-4007-2040-6/2025/11 CCS: Computing methodologies Artificial intelligence CCS: Applied computing Economics CCS: Applied computing Forecasting Zhuohang Zhu, Haodong Chen, Qiang Qu, Vera Chung Abstract . Financial time-series forecasting is critical for maintaining economic stability, guiding informed policymaking, and promoting sustainable investment practices. However, it remains challenging due to various underlying pattern shifts. These shifts arise primarily from three sources: temporal non-stationarity (distribution changes over time), multi-domain diversity (distinct patterns across financial domains such as stocks, commodities, and futures), and varying temporal resolutions (patterns differing across per-second, hourly, daily, or weekly indicators). While recent deep learning methods attempt to address these complexities, they frequently suffer from overfitting and typically require extensive domain-specific fine-tuning. To overcome these limitations, we introduce FinCast , the first foundation model specifically designed for financial time-series forecasting, trained on large-scale financial datasets. Remarkably, FinCast exhibits robust zero-shot performance, effectively capturing diverse patterns without domain-specific fine-tuning. Comprehensive empirical and qualitative evaluations demonstrate that FinCast surpasses existing state-of-the-art methods, highlighting its strong generalization capabilities. † † cc-license: by 1. Introduction Forecasting financial time series is crucial for supporting economic stability, guiding investment decisions ( Tsay 2005 ) , and managing financial risk ( Chatfield 2013 ) . Reliable forecasts help allocate capital efficiently, reduce exposure to market shocks, and inform regulatory policy ( Taylor 2011 ) . From central banks setting interest rates to institutional investors managing portfolios, accurate forecasts enable timely, data-driven decisions that influence both short-term market movements ( Franses 1998 ) and long-term economic outcomes ( Cochrane 1997 ) . Despite its importance, financial time-series forecasting remains highly challenging due to various underlying pattern shifts ( Xu and Cohen 2018 ; Tsay 2005 ) . First, financial time series are inherently non-stationary ( Taylor 2011 ) : their distribution shifts over time due to factors such as structural economic changes, shifting investor behavior, policy interventions, and technological disruptions. For example, the distribution of prices for a stock like Apple differs significantly between 2021 and 2025, shaped by both macroeconomic conditions and firm-level developments. Second, forecasting across financial domains poses a core modeling challenge. Each domain, such as stocks, commodities, or currencies, exhibits distinct patterns shaped by diverse factors such as economic mechanisms, regulatory environments, and market structures ( Franses 1998 ) . Third, financial time series occur at varying temporal resolutions, from second-level tick data to weekly or monthly indicators ( Cochrane 1997 ) . For example, high-frequency data reflect rapid, noise-driven fluctuations, while lower-frequency data capture slower, macro-driven trends. These dynamics are often incompatible, and models designed for a single resolution typically fail to generalize across different temporal resolutions ( Chatfield 2013 ) . Existing forecasting models often fail under real-world conditions: they struggle to generalize across distribution shifts, financial domains, and temporal resolutions. Models trained on one financial domain or temporal resolution typically perform poorly when applied elsewhere, and their accuracy degrades rapidly when distributional properties shift. A central reason is that most approaches rely on supervised learning with strong assumptions about the stability of underlying patterns. Their architectures are often tailored to specific financial domains (e.g., stocks) or temporal resolutions (e.g., daily data), which limits their applicability beyond the original setting. Trained on fixed datasets and optimized for narrow tasks, these models overfit to historical patterns and are unable to generalize beyond their original context. To address the limitations of existing approaches, we introduce FinCast , a foundation model for financial time-series forecasting. Intuitively, a large-capacity model, trained on sufficiently diverse and large-scale financial data, can learn a broad spectrum of temporal patterns, domain-specific dynamics, and resolution-dependent behaviors. FinCast is implemented as a large decoder-only transformer and trained on over 20 billion time points across a wide range of financial domains and temporal resolutions. To enable this generalization in practice, we introduce three key design choices. First, Point-Quantile loss (PQ-loss), which jointly optimizes point forecasts and quantile-based probabilistic estimates to model uncertainty across the distribution, enhances robustness to temporal shifts and prevents forecast collapse. Second, a token-level sparse Mixture-of-Experts (MoE) mechanism that increases capacity efficiently and enables experts to specialize across domains. Third, learnable frequency embeddings that encode temporal characteristics at varying resolutions, improving the capture of cyclic and seasonal patterns. This unified framework balances high capacity with robustness, allowing FinCast to learn both shared and domain-specific dynamics across financial time series. Empirical evaluations validate the effectiveness of our approach. FinCast consistently outperforms state-of-the-art methods across both zero-shot and supervised financial forecasting benchmarks, achieving best results without task-specific fine-tuning. Our experiments span a wide range of financial domains, including stocks, cryptocurrencies, forex, and futures, capturing the diversity and non-stationarity of real-world markets. Complementary qualitative analyses further show that FinCast adapts well to shifting patterns across domains and temporal resolutions. Our contributions can be summarised as follows: • We introduce the first foundation model for financial time-series forecasting, a decoder-only transformer with 1B parameters, trained on 20B+ time points across diverse financial domains and temporal resolutions. • We propose a novel Point-Quantile Loss that combines point forecasts with quantile-based probabilistic estimates to enhance robustness under temporal non-stationarity. • We design a learnable frequency embedding that encodes temporal resolution, enhancing adaptability across various temporal resolutions. Combined with a token-level sparse Mixture-of-Experts mechanism, this increases model’s capacity efficiently and enables expert specialization across financial domains. • FinCast consistently outperforms state-of-the-art methods, achieving reductions in forecasting error by an average of 20% and 23% respectively. 2. Related Work Traditional financial time-series forecasting has historically relied on statistical models such as ARIMA ( Ariyo et al. 2014 ) , GARCH ( Franses and Van Dijk 1996 ) , and other domain-specific techniques ( Wang et al. 2020 ) . They often fall short in capturing nonlinear dynamics and abrupt domain shifts. The emergence of deep learning introduced recurrent neural networks ( Medsker et al. 2001 ) , particularly Long Short-Term Memory ( Gers et al. 2000 ) architectures, as popular tools for modeling temporal dependencies in financial time series ( Zhu et al. 2024 ; Du et al. 2024 ; You et al. 2021 ) . While these models can learn short- and medium-term dependencies, they tend to struggle with long-range correlations and suffer from vanishing gradients ( Hou et al. 2021 ) . To account for inter-series relationships and interpretability, graph-based models have gained traction ( Liu et al. 2024 ; Chang et al. 2020 ; Ding et al. 2020 ) . By representing stocks as nodes and their dependencies as edges, these methods incorporate relational inductive biases via explicit or learned topologies. Recent work has attempted to integrate graph-based models with sequence or variational frameworks ( Jia et al. 2024 ; Cheng et al. 2024 ) . Transformer architectures ( Vaswani et al. 2017 ) , originally developed for NLP ( Achiam et al. 2023 ) and later adapted to vision tasks ( Dosovitskiy et al. 2021 ) , have shown promise in time-series forecasting

Read the complete text