Paper: arXiv 2505.11163
Authors: Anubha Goel, Puneet Pasricha, Martin Magris, Juho Kanniainen
Abstract
Time series foundation models (FMs) have emerged as a popular paradigm for zero-shot multi-domain forecasting. These models are trained on numerous diverse datasets and claim to be effective forecasters across multiple different time series domains, including financial data. In this study, we evaluate the effectiveness of FMs, specifically the TimesFM model, for volatility forecasting, a core task in financial risk management. We first evaluate TimesFM in its pretrained (zero-shot) form, followed by our custom fine-tuning procedure based on incremental learning, and compare the resulting models against standard econometric benchmarks. While the pretrained model provides a reasonable baseline, our findings show that incremental fine-tuning, which allows the model to adapt to new financial return data over time, is essential for learning volatility patterns effectively. Fine-tuned variants not only improve forecast accuracy but also statistically outperform traditional models, as demonstrated through Diebold-Mariano and Giacomini-White tests. These results highlight the potential of foundation models as scalable and adaptive tools for financial forecasting-capable of delivering strong performance in dynamic market environments when paired with targeted fine-tuning strategies.
Complexity vs Empirical Score
- Math Complexity: 5.0/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper applies advanced deep learning techniques (transformer architecture, incremental learning) but primarily focuses on empirical evaluation, statistical tests, and extensive backtesting across 21 global indices.
Research Flowchart
flowchart TD
A["Research Goal:<br>Evaluate TimesFM for Volatility Forecasting"] --> B["Data: S&P 500 & NASDAQ<br>Realized Volatility Series"]
B --> C["Methodology: Three-Stage Approach"]
C --> D["1. Zero-Shot<br>Pretrained TimesFM"]
C --> E["2. Fine-Tuned<br>Incremental Learning"]
C --> F["3. Baseline<br>Standard Econometrics"]
D --> G["Testing & Comparison"]
E --> G
F --> G
G --> H{"Evaluation"}
H --> I["Forecast Accuracy<br>vs Benchmarks"]
H --> J["Statistical Tests<br>DM & GW"]
I --> K["Key Findings"]
J --> K
K --> L["• Incremental Fine-Tuning Essential"]
K --> M["• Outperforms Econometrics"]
K --> N["• Scalable for Financial Risk"]
style A fill:#e1f5fe
style B fill:#fff3e0
style C fill:#f3e5f5
style K fill:#e8f5e8
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 proposes an incremental fine‑tuning procedure for the pre‑trained TimesFM foundation model to improve realized volatility forecasting.
Source sentence
We first evaluate TimesFM in its pretrained (zero-shot) form, followed by our custom fine-tuning procedure based on incremental learning, and compare the resulting models against standard econometric benchmarks.
Assumes 4
- The explanatory variables are assumed to be stationary, limiting adaptability of traditional econometric models. statistical
Source sentence
limited adaptability to evolving market conditions primarily due to the stationarity assumption of the explanatory variables.
- Forecast horizon is set to one day ahead (H = 1). model
Source sentence
Throughout this study, we set H = 1.
- The model is retrained on a rolling window of the most recent L observations. model
Source sentence
the model is trained using the most recent L observations, xt−L+1, . . . , xt, where L denotes the length of the historical context.
- Data are daily realized volatility observations. data
Source sentence
The dataset provides daily observations, including key realized measures of market volatility.
Data
- Oxford-Man Institute Realized Library market-prices public
Source sentence
This study employs high-frequency realized volatility data from the Oxford-Man Institute’s Realized Library for 21 major global stock market indices. The
Costs
Not applicable
Finds 3
- Fine‑tuned TimesFM variants improve forecast accuracy and statistically outperform traditional econometric models according to Diebold‑Mariano and Giacomini‑White tests. statistical
Source sentence
Fine-tuned variants not only improve forecast accuracy but also statistically outperform traditional models, as demonstrated through Diebold-Mariano and Giacomini-White tests.
- TimesFM models (e.g., TFM64logIL and TTFM512PT) achieve lower relative errors across loss functions such as MSE, QLIKE, and sMAPE. performance
Source sentence
TimesFM models, particularly TFM64logIL and TTFM512PT, consistently achieve lower relative errors across loss functions such as MSE, QLIKE, and sMAPE.
- TFM512PT and TFM512IL consistently outperform traditional models under the Q‑like loss function, with p‑values below 0.10 in DM and GW tests. statistical
Source sentence
TFM512PT and TFM512IL consistently outperform traditional models under the Q-like loss function, with p-values below 0.10 in DM and GW tests.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 36 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 (71)
- A machine learning approach to volatility forecasting (2023)
- Neural network–based financial volatility forecasting: A systematic review (2022)
- Volatility forecast using hybrid neural network models (2014)
- GARCH based artificial neural networks in forecasting conditional variance of stock returns (2014)
- Combining conditional volatility forecasts using neural networks: an application to the EMS exchange rates (1999)
- Artificial neural networks for realized volatility prediction in cryptocurrency time series (2019)
- Evolving hybrid neural fuzzy network for realized volatility forecasting with jumps (2014)
- Forecasting realized volatility: HAR against Principal Components Combining, neural networks and GARCH (2017)
and 63 more
- Evolving fuzzy-GARCH approach for financial volatility modeling and forecasting (2016)
- A hybrid modeling approach for forecasting the volatility of S&P 500 index return (2012)
- Using neural networks for forecasting volatility of S&P 500 Index futures prices (2004)
- Forecast combining with neural networks (1996)
- HAR volatility modelling with heterogeneous leverage and jumps (2009)
- Volatility forecasting with machine learning and intraday commonality (2024)
- Stock market volatility forecasting: Do we need high-frequency data? (2021)
- Volatility forecasting using deep recurrent neural networks as GARCH models (2023)
- Comparing predictive accuracy (2002)
- A forecast comparison of volatility models: does anything beat a GARCH (1, 1)? (2005)
- Power and bipower variation with stochastic volatility and jumps (2004)
- Roughing It Up: Including Jump Components in the Measurement, Modeling, and Forecasting of Return Volatility (2007)
- Good volatility, bad volatility: Signed jumps and the persistence of volatility (2015)
- Exploiting the errors: A simple approach for improved volatility forecasting (2016)
- A simple approximate long-memory model of realized volatility (2009)
- Realized GARCH: a joint model for returns and realized measures of volatility (2012)
- Volatility forecast comparison using imperfect volatility proxies (2011)
- Discrete-time volatility forecasting with persistent leverage effect and the link with continuous-time volatility modeling (2012)
- Forecasting stock market volatility using realized GARCH model: International evidence (2016)
- The model confidence set (2011)
- Tests of conditional predictive ability (2006)
- A decoder-only foundation model for time-series forecasting (2024)
- Using LSTM and GRU neural network methods for traffic flow prediction (2016)
- Time-Series Foundation AI Model for Value-at-Risk Forecasting — → 2410.11773
- Estimating quadratic variation using realized variance (2002)
- A time series is worth 64 words: Long-term forecasting with transformers (2022) — link
- Chronos: Learning the language of time series (2024)
- Lag-llama: Towards foundation models for time series forecasting (2023)
- Visionts: Visual masked autoencoders are free-lunch zero-shot time series forecasters (2024) — link
- Foundation models for time series analysis: A tutorial and survey (2024)
- Unified training of universal time series forecasting transformers (2024)
- Tiny time mixers (ttms): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series (2024)
- The M4 Competition: 100,000 time series and 61 forecasting methods (2020)
- A review of generalized zero-shot learning methods (2022)
- DeepAR: Probabilistic forecasting with autoregressive recurrent networks (2020)
- Monash time series forecasting archive (2021) — link
- N-BEATS: Neural basis expansion analysis for interpretable time series forecasting (2019) — link
- Informer: Beyond efficient transformer for long sequence time-series forecasting (2021)
- Gpt (generative pre-trained transformer)–a comprehensive review on enabling technologies, potential applications, emerging challenges, and future directions (2024)
- Mamba4Cast: Efficient Zero-Shot Time Series Forecasting with State Space Models (2024) — link
- Timedit: General-purpose diffusion transformers for time series foundation model (2024) — link
- Masked autoencoders are scalable vision learners (2022)
- Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation (1982)
- Generalized autoregressive conditional heteroskedasticity (1986)
- Modeling and predicting the CBOE market volatility index (2014)
- Novel volatility forecasting using deep learning–long short term memory recurrent neural networks (2019)
- The benefits of bagging for forecast models of realized volatility (2010)
- Introducing NBEATSx to realized volatility forecasting (2024)
- Machine learning for realised volatility forecasting (2020)
- An artificial neural network-GARCH model for international stock return volatility (1997)
- Realized volatility forecasting with neural networks (2020)
- Trading volume and realized volatility forecasting: Evidence from the China stock market (2023)
- Forecasting realised volatility using ARFIMA and HAR models (2019)
- Financial returns modelled by the product of two stochastic processes-a study of the daily sugar prices 1961-75 (1982)
- Lassoing the HAR model: A model selection perspective on realized volatility dynamics (2016)
- The impact of sentiment and attention measures on stock market volatility (2020)
- Building news measures from textual data and an application to volatility forecasting (2017)
- Forecasting of realised volatility with the random forests algorithm (2018)
- Stock market volatility: Identifying major drivers and the nature of their impact (2015)
- Data-driven neural architecture learning for financial time-series forecasting (2019) — link
- Deep adaptive input normalization for time series forecasting (2019)
- Deeplob: Deep convolutional neural networks for limit order books (2019)
- Using machine learning to predict realized variance (2020)
1 of 71 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.