Paper: arXiv 2511.18578
Authors: Eghbal Rahimikia, Hao Ni, Weiguan Wang
Abstract
Financial time series forecasting is central to trading, portfolio optimization, and risk management, yet it remains challenging due to noisy, non-stationary, and heterogeneous data. Recent advances in time series foundation models (TSFMs), inspired by large language models, offer a new paradigm for learning generalizable temporal representations from large and diverse datasets. This paper presents the first comprehensive empirical study of TSFMs in global financial markets. Using a large-scale dataset of daily excess returns across diverse markets, we evaluate zero-shot inference, fine-tuning, and pre-training from scratch against strong benchmark models. We find that off-the-shelf pre-trained TSFMs perform poorly in zero-shot and fine-tuning settings, whereas models pre-trained from scratch on financial data achieve substantial forecasting and economic improvements, underscoring the value of domain-specific adaptation. Increasing the dataset size, incorporating synthetic data augmentation, and applying hyperparameter tuning further enhance performance.
Complexity vs Empirical Score
- Math Complexity: 6.0/10
- Empirical Rigor: 8.5/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced machine learning architectures (TSFMs) and complex evaluation methodologies, but its primary strength is the large-scale empirical study using real financial data (2 billion observations) with robust backtesting protocols, economic metrics, and benchmarks.
Research Flowchart
flowchart TD A["Research Goal: Evaluate TSFMs in Global Financial Markets"] --> B["Methodology: Zero-Shot, Fine-Tuning, Pre-train from Scratch"] B --> C["Data: Daily Excess Returns<br>Global Equities Markets"] C --> D["Computational Process: Benchmarking vs. TSFMs"] D --> E["Key Findings: Domain-Specific Pre-training<br>Significantly Outperforms Off-the-Shelf TSFMs"] E --> F["Outcomes: Improved Forecasting & Economics via<br>Data Scaling, Synthetic Augmentation, Hyperparameter Tuning"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper conducts the first comprehensive empirical study of time series foundation models (TSFMs) for excess return forecasting in global financial markets.
Source sentence
This paper presents the first comprehensive empirical study of TSFMs in global financial markets.
Assumes 2
- The conditional law of one-step excess returns is identical across assets (homogeneous market structure). market
Source sentence
We further assume that the conditional law of one-step excess returns is identical across assets, that is, P(r(i),ex t+1 |r(i),ex 1:t = r) does not depend on i.
- The time series length T is an integer multiple of the patch length L; otherwise padding is applied. model
Source sentence
For simplicity, in Equation (6), we assume that T is an integer multiple of L; otherwise, the sequence can be padded to satisfy this condition.
Data
- Global daily excess returns market-prices unknown
Source sentence
Section 5.2 broadens the scope to a global context by training all models, including both benchmarks and TSFMs, on an extensive dataset encompassing 94 countries.
- Synthetic data augmentation synthetic unknown
Source sentence
Increasing the dataset size, incorporating synthetic data augmentation, and applying hyperparameter tuning further enhance performance.
JKP Global Factor Data CRSP Compustat Kaggle
Costs
Modeled
Results are evaluated net of transaction costs under four scenarios: no costs (0 bps), fixed costs of 20 bps and 40 bps, and a mixed-cost structure (≈21.3 bps for small‑cap stocks and ≈11.2 bps for large‑cap stocks).
Source sentence
Performance is evaluated net of transaction costs under four scenarios: no costs (0 bps), fixed costs of 20 bps and 40 bps, and a mixed-cost structure where small- and large-cap stocks face different estimated trading costs
Finds 5
- Off‑the‑shelf pre‑trained TSFMs perform poorly in zero‑shot and fine‑tuning settings compared to strong benchmark models. performance
Source sentence
We find that off-the-shelf pre-trained TSFMs perform poorly in zero-shot and fine-tuning settings, whereas models pre-trained from scratch on financial data achieve substantial forecasting and economic improvements.
- TSFMs pre‑trained from scratch on financial data deliver substantial gains in both predictive and economic performance, especially with longer input windows. performance
Source sentence
TSFMs pre-trained from scratch on financial data deliver substantial gains in both predictive and economic performance and are most competitive with longer input windows.
- Fine‑tuned Chronos (large) achieves positive out‑of‑sample R² of approximately 0.46–0.48. predictive-accuracy
Source sentence
Chronos (large), which, after fine-tuning, achieves positive R2 OOS (R2 OOS ≈0.46–0.48) across window
- TSFMs show slower performance decay over time, indicating greater temporal robustness compared to benchmarks. statistical
Source sentence
TSFMs also display slower performance decay over time, suggesting greater temporal robustness.
- Despite improvements, TSFMs remain weaker in goodness‑of‑fit; zero‑shot models yield poor R² and fine‑tuning offers only modest improvements. negative
Source sentence
TSFMs remain weaker in terms of goodness-of-fit: zero-shot models yield poor R2, fine-tuning offers modest improvements, and finance-native pre-training does not fully close the gap.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 32 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (223)
- A Semiparametric Intraday GARCH Model (2016) — link
- DeepSeek-V3 Technical Report (2024) — link
- Training Language Models to Follow Instructions with Human Feedback (2022)
- Adam: A Method for Stochastic Optimization (2015) — link
- Quant GANs: deep generation of financial time series (2020) — link
- Fin-GAN: Forecasting and Classifying Financial Time Series via Generative Adversarial Networks (2024) — link
- Sig-Wasserstein GANs for Conditional Time Series Generation (2024)
- Generative Learning for Financial Time Series with Irregular and Scale-Invariant Patterns (2024)
and 215 more
- Japanese and Korean voice search (2012)
- Subword Regularization: Improving Neural Network Translation Models with Multiple Subword Candidates (2018)
- Big data analytics (2011)
- Forecasting volatility in financial markets: A review (2003)
- Answering the skeptics: Yes, standard volatility models do provide accurate forecasts (1998)
- Non-Gaussian Ornstein–Uhlenbeck-based models and some of their uses in financial economics (2001)
- A simple approximate long-memory model of realized volatility (2009)
- HAR volatility modelling with heterogeneous leverage and jumps (2009)
- Generalized Method of Integrated Moments for High-Frequency Data (2016)
- Exploiting the errors: A simple approach for improved volatility forecasting (2016)
- Do we need the constant term in the heterogenous autoregressive model for forecasting realized volatilities? (2018)
- Forecasting Realized Volatility: A Review (2018) — link
- Forecasting realized volatility: A review (2018)
- Order book slope and price volatility (2008)
- Good volatility, bad volatility: Signed jumps and the persistence of volatility (2015)
- Dynamics of the limit order book and the volume-volatility relation (2018)
- Sequence classification of the limit order book using recurrent neural networks (2018)
- Modelling high-frequency limit order book dynamics with support vector machines (2015)
- Benchmark dataset for mid-price forecasting of limit order book data with machine learning methods (2018)
- Convolutional neural networks applied to high-frequency market microstructure forecasting (2017)
- Universal features of price formation in financial markets: perspectives from deep learning (2018)
- Deep learning for limit order books (2019)
- Prediction of Hidden Liquidity in theLimit Order Book of GLOBEX Futures (2013)
- Forecasting of jump arrivals in stock prices: New attention-based network architecture using limit order book data (2018) — link
- Multivariate realised kernels: consistent positive semi-definite estimators of the covariation of equity prices with noise and non-synchronous trading (2011)
- A tale of two time scales: Determining integrated volatility with noisy high-frequency data (2005)
- Volatility forecast comparison using imperfect volatility proxies (2011)
- Forecasting the volatility of crude oil futures using HAR-type models with structural breaks (2016)
- Forecasting volatility of the US oil market (2014)
- Volatility transmission in emerging European foreign exchange markets (2011)
- Quantile forecasts of daily exchange rate returns from forecasts of realized volatility (2008)
- Relative liquidity and future volatility (2015)
- Limit orders, depth, and volatility: Evidence from the stock exchange of Hong Kong (2001)
- Does the open limit order book matter in explaining informational volatility? (2009)
- LOBSTER: Limit order book reconstruction system (2011)
- Realized kernels in practice: Trades and quotes (2009)
- Realized variance and market microstructure noise (2006)
- Order book characteristics and the volume–volatility relation: Empirical evidence from a limit order market (2006)
- Measuring stock illiquidity: An investigation of the demand and supply schedules at the TASE (2004)
- Time-series classification using neural bag-of-features (2017)
- Order aggressiveness in limit order book markets (2004)
- Canceling liquidity (2015)
- A simple nonparametric test of predictive performance (1992)
- Estimating the persistence and the autocorrelation function of a time series that is measured with error (2014)
- Unstructured data and the 80 percent rule (2008)
- News implied volatility and disaster concerns (2017)
- When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks (2011)
- Textual Analysis in Accounting and Finance: A Survey (2016)
- Predicting Returns With Text Data (2019)
- Giving Content to Investor Sentiment: The Role of Media in the Stock Market (2007)
- Can stock market forecasters forecast? (1933)
- The impact of sentiment and attention measures on stock market volatility (2020)
- Swisscheese at semeval-2016 task 4: Sentiment classification using an ensemble of convolutional neural networks with distant supervision (2016)
- Leveraging large amounts of weakly supervised data for multi-language sentiment classification (2017)
- Predicting financial markets: Comparing survey, news, twitter and search engine data (2011) — link
- Distillation of news flow into analysis of stock reactions (2016)
- The Twitter myth revisited: Intraday investor sentiment, Twitter activity and individual-level stock return volatility (2018)
- Intraday online investor sentiment and return patterns in the US stock market (2017)
- Is All That Talk Just Noise? The Information Content of Internet Stock Message Boards (2004)
- Stock return predictability and investor sentiment: A high-frequency perspective (2016)
- Power and bipower variation with stochastic volatility and jumps (2004)
- A reality check for data snooping (2000)
- MFE MATLAB function reference financial econometrics (2009)
- The stationary bootstrap (1994)
- Comparing predictive accuracy (2002)
- Yahoo! for Amazon: Sentiment extraction from small talk on the web (2007)
- Measuring and testing the impact of news on volatility (1993)
- Long short-term memory (1997)
- Learning to forget: Continual prediction with LSTM (1999)
- Deep learning in asset pricing (2019)
- Deep learning in asset pricing (2020)
- Deep learning (2016)
- Improving neural networks by preventing co-adaptation of feature detectors (2012) — link
- Dropout: a simple way to prevent neural networks from overfitting (2014)
- Adam: A Method for Stochastic Optimization (2014) — link
- Asymptotic inference about predictive ability (1996)
- Evaluating volatility and correlation forecasts (2009)
- Alternative Data for Realised Volatility Forecasting: Limit Order Book and News Stories (2020)
- Efficient Estimation of Word Representations in Vector Space (2013) — link
- Advances in Pre-Training Distributed Word Representations (2018)
- Enriching word vectors with subword information (2017)
- Simlex-999: Evaluating semantic models with (genuine) similarity estimation (2015)
- A study on similarity and relatedness using distributional and wordnet-based approaches (2009)
- Convolutional Neural Networks for Sentence Classification (2014) — link
- Measuring and hedging geopolitical risk (2020)
- The common and specific components of dynamic volatility (2006)
- Text as Data (2019)
- Empirical Asset Pricing via Machine Learning (2020)
- Autoencoder asset pricing models (2021)
- (Re-) Imag (in) ing Price Trends (2020)
- Machine learning for realised volatility forecasting (2020)
- The structure of economic news (2020)
- Distributed representations of words and phrases and their compositionality (2013) — link
- Exploiting similarities among languages for machine translation (2013) — link
- Hierarchical probabilistic neural network language model. (2005)
- FastText.zip: Compressing text classification models (2016) — link
- Normalized (pointwise) mutual information in collocation extraction (2009)
- Axiomatic attribution for deep networks (2017)
- Natural language processing (almost) from scratch (2011)
- A Unified Approach to Interpreting Model Predictions (2017) — link
- Learning important features through propagating activation differences (2017)
- Modelling Volatility Cycles: the (MF)^ 2 GARCH Model (2021)
- Predictably unequal? the effects of machine learning on credit markets (2021)
- Enhancing time-series momentum strategies using deep neural networks (2019)
- Deep reinforcement learning for trading (2020)
- A picture is worth a thousand words: Measuring investor sentiment by combining machine learning and photos from news (2021)
- Machine learning and portfolio optimization (2018)
- Forecasting the equity premium: mind the news! (2020)
- Bond risk premiums with machine learning (2021)
- Corrigendum: Bond risk premiums with machine learning (2021)
- The role of corporate culture in bad times: Evidence from the COVID-19 pandemic (2020)
- Venture capital communities (2020)
- A cross-sectional machine learning approach for hedge fund return prediction and selection (2020)
- Measuring news sentiment (2020)
- Deeplob: Deep convolutional neural networks for limit order books (2019)
- Universal features of price formation in financial markets: perspectives from deep learning (2019)
- DeepLOB: Deep Convolutional Neural Networks for Limit Order Books (2019)
- Multi-Horizon Forecasting for Limit Order Books: Novel Deep Learning Approaches and Hardware Acceleration using Intelligent Processing Units (2021) — link
- Building Cross-Sectional Systematic Strategies by Learning to Rank (2021)
- Slow Momentum with Fast Reversion: A Trading Strategy Using Deep Learning and Changepoint Detection (2021) — link
- Shap values for explaining cnn-based text classification models (2020) — link
- Reducing duplicate filters in deep neural networks (2017)
- A machine learning approach to volatility forecasting (2021)
- Textual factors: A scalable, interpretable, and data-driven approach to analyzing unstructured information (2019)
- A primer on neural network models for natural language processing (2016)
- Modeling price volatility based on a genetic programming approach (2019)
- Measuring firms’ market orientation using textual analysis of 10-K filings (2020)
- Identifying linguistic markers of CEO hubris: a machine learning approach (2022)
- Forecasting realized volatility with machine learning: Panel data perspective (2023)
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018) — link
- RoBERTa: A Robustly Optimized BERT Pretraining Approach (2019) — link
- Language Models are Unsupervised Multitask Learners (2019)
- Language Models are Few-Shot Learners (2020) — link
- LLaMA: Open and Efficient Foundation Language Models (2023) — link
- Llama 2: Open Foundation and Fine-Tuned Chat Models (2023) — link
- The Llama 3 Herd of Models (2024) — link
- Realised Volatility Forecasting: Machine Learning via Financial Word Embedding (2021) — link
- FinBERT: A Large Language Model for Extracting Information from Financial Text (2023)
- Expected Returns and Large Language Models (2022)
- How Much Can Machines Learn Finance from Chinese Text Data? (2024)
- Lookahead Bias in Pretrained Language Models (2024)
- Financial Statement Analysis with Large Language Models — → 2407.17866
- Bloated Disclosures: Can ChatGPT Help Investors Process Information? (2024)
- Time Machine GPT (2024)
- Deep Contextualized Word Representations (2018) — link
- Well-Read Students Learn Better: On the Importance of Pre-training Compact Models (2019) — link
- ZeRO: Memory optimizations Toward Training Trillion Parameter Models (2020)
- Decoupled Weight Decay Regularization (2017) — link
- On Persistence in Mutual Fund Performance (1997)
- Trading Costs of Asset Pricing Anomalies (2012)
- Estimating Number of Factors by Adjusted Eigenvalues Thresholding (2022)
- Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models — → 2304.07619
- ChatGPT, Stock Market Predictability and Links to the Macroeconomy (2023)
- Is There a Replication Crisis in Finance? (2023)
- A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix (1986)
- Chronos: Learning the Language of Time Series (2024) — link
- Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (2020)
- A Decoder-Only Foundation Model for Time-Series Forecasting (2024)
- MOMENT: A Family of Open Time-series Foundation Models (2024) — link
- Kairos: Towards Adaptive and Generalizable Time Series Foundation Models (2025) — link
- Attention Is All You Need (2017)
- Foundation Models for Time Series Analysis : A Tutorial and Survey (2024) — link
- Unified Training of Universal Time Series Forecasting Transformers (2024) — link
- CatBoost: Unbiased Boosting with Categorical Features (2018)
- XGBoost: A Scalable Tree Boosting System (2016) — link
- LightGBM: A Highly Efficient Gradient Boosting Decision Tree (2017) — link
- TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning (2025) — link
- xLSTM 7B: A Recurrent LLM for Fast and Efficient Inference (2025) — link
- Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting (2023) — link
- Simplified State Space Layers for Sequence Modeling (2022) — link
- FlowState: Sampling Rate Invariant Time Series Forecasting (2025) — link
- Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series (2024)
- Sundial: A Family of Highly Capable Time Series Foundation Models (2025) — link
- Toto: Time Series Optimized Transformer for Observability (2024) — link
- Empirical Asset Pricing: The Cross Section of Stock Returns (2016)
- Returns to Buying Winners and Selling Losers : Implications for Stock Market Efficiency (1993) — link
- Evidence of Predictable Behavior of Security Returns (1990) — link
- Foundations of Technical Analysis : Computational Algorithms , Statistical Inference , and Empirical Implementation (2000) — link
- (Re-) Imag (in) ing price trends (2023)
- Charting by Machines (2024) — link
- Machine Learning from a ``Universe’’ of Signals: The Role of Feature Engineering (2025) — link
- Market Efficiency in the Age of Big Data (2022)
- Forecasting the Equity Risk Premium : The Role of Technical Indicators (2014) — link
- Simple Technical Trading Rules and the Stochastic Properties of Stock Returns (1992)
- Machine Learning in the Chinese Stock Market (2022) — link
- Re(Visiting) Large Language Models in Finance (2024)
- Chronologically Consistent Large Language Models — → 2502.21206
- Deep Learning in Asset Pricing (2024)
- Artificial Intelligence Asset Pricing Models (2025)
- Isambard-AI: A Leadership-Class Supercomputer Optimised Specifically for Artificial Intelligence (2024)
- Time Series Momentum (2012)
- Value and Momentum Everywhere (2013)
- Risks and Returns of Cryptocurrency (2021)
- Momentum Has Its Moments (2015)
- Currency Momentum Strategies (2012)
- Factor Momentum and the Momentum Factor (2022)
- Factor Momentum Everywhere (2018)
- Machine Learning from a ``Universe’’ of Signals: The Role of Feature Engineering (2025)
- The Virtue of Complexity in Return Prediction (2024)
- Quant GANs: Deep Generation of Financial Time Series (2020)
- Ridge Regression: Biased Estimation for Nonorthogonal Problems (1970)
- Regression Shrinkage and Selection via the Lasso (1996)
- Regularization and Variable Selection via the Elastic Net (2005)
- Principal Components Regression in Exploratory Statistical Research (1965)
- Robust Statistics (2011)
- Simplified State Space Layers for Sequence Modeling (2022) — link
- Understanding Transformers for Time Series: Rank Structure, Flow-of-ranks, and Compressibility (2025) — link
- Simplified: A Closer Look at the Virtue of Complexity in Return Prediction (2025)
- Seemingly Virtuous Complexity in Return Prediction (2025)
- Understanding The Virtue of Complexity (2025)
- The Limited Virtue of Complexity in a Noisy World (2025)
- Comment on “The Virtue of Complexity in Return Prediction” (2023)
- Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Experts (2024) — link
- Training Compute-Optimal Large Language Models (2022) — link
- Neural Machine Translation of Rare Words with Subword Units (2016)
3 of 223 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.