Paper: arXiv 2503.12648

Authors: Andreas Teller, Uta Pigorsch, Christian Pigorsch

Abstract

Forecasting the volatility of financial assets is essential for various financial applications. This paper addresses the challenging task of forecasting the volatility of financial assets with limited historical data, such as new issues or spin-offs, by proposing a multi-source transfer learning approach. Specifically, we exploit complementary source data of assets with a substantial historical data record by selecting source time series instances that are most similar to the limited target data of the new issue/spin-off. Based on these instances and the target data, we estimate linear and non-linear realized volatility models and compare their forecasting performance to forecasts of models trained exclusively on the target data, and models trained on the entire source and target data. The results show that our transfer learning approach outperforms the alternative models and that the integration of complementary data is also beneficial immediately after the initial trading day of the new issue/spin-off.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced statistical and machine learning techniques (transfer learning, DTW, neural networks) and includes rigorous empirical testing with specific datasets, forecasting horizons, and comparative model evaluations.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Forecast RV for New Issues/Spin-Offs<br>with limited data"] --> B["Input Data<br>Source Assets: Extensive RV history<br>Target Assets: Limited RV history"]
  
  B --> C["Methodology<br>Multi-Source Transfer Learning"]
  
  C --> D{"Select Similar Instances<br>Find historical periods in source data<br>matching target asset characteristics"}
  
  D --> E["Model Estimation<br>Train Linear & Non-linear RV models<br>on combined data"]
  
  E --> F["Computational Process<br>Forecast Volatility<br>vs Baselines: Target-only & Full-source models"]
  
  F --> G["Key Findings<br>Transfer learning outperforms alternatives<br>Beneficial immediately after trading day 1"]

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 a model‑agnostic, instance‑based multi‑source transfer learning approach that selects source subsequences via DTW similarity to improve realized volatility forecasts for new issues and spin‑offs.

Source sentence
we propose an instance-based multi-source transfer learning approach that selects subsequences from complementary financial assets based on the dynamic time warping (DTW) distance between the new issue/spin-off and complementary data series.

Assumes 5

  • The main limitation for forecasting new issues is the lack of historical data at the beginning of an asset’s life cycle. data
    Source sentence
    It can be assumed that the main reason for this shortage in the existing research is the lack of historical data at the beginning of a financial asset’s life cycle.
  • The transfer learning method can be applied to any forecasting model (model‑agnostic). model
    Source sentence
    Our proposed transfer learning approach has the advantage of being model-agnostic, meaning it can be applied to any type of model one might consider for forecasting realized volatility.
  • Target and source data contain labeled volatility observations suitable for supervised learning. data
    Source sentence
    When predicting the 1-day-ahead volatility of new issues/spin-offs based on past volatility observations, we encounter data sets that inherently contain labeled data.
  • Dynamic Time Warping distance is an appropriate similarity metric for selecting relevant source subsequences. statistical
    Source sentence
    After computing the DTW distance between the most recent target observations and each subsequence within U, we selectively transfer into the training data set DT only those subsequences whose DTW distance ranks below a predefined percentile ϵ.
Show 1 more
  • Only a minimal amount of target data (as little as the first trading day) is sufficient when combined with selected source data. data
    Source sentence
    our approach necessitates only a minimal amount of target data and can be combined with any realized variance forecasting model.

Data

  • Target new‑issue/spin‑off intraday price data · FirstRate Data LLC market-prices commercial
    2010-2020 · 10 new issues and spin‑offs that entered the public market and were part of the S&P 500 on
    Source sentence
    As our primary target data set, we use 1-minute intraday stock price data obtained from FirstRate Data LLC on 10 new issues and spin-offs that entered the public market between 2010 and 2020 and were part of the S&P 500 index on November 1, 2022.
  • Source S&P 500 intraday price data · FirstRate Data LLC market-prices commercial
    66 additional S&P 500 stocks selected by market‑cap within each GICS sector
    Source sentence
    Furthermore, for our source data set, we consider the 1-minute intraday price data of 66 additional S&P 500 stocks. Again, the data is retrieved from FirstRate Data LLC.

Recognised sources: Cboe Yahoo Finance

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • The multi‑source transfer learning approach outperforms models trained only on target data or on pooled target‑plus‑all‑source data. performance
    Source sentence
    The results show that our transfer learning approach outperforms the alternative models and that the integration of complementary data is also beneficial immediately after the initial trading day of the new issue/spin-off.
  • Combining the transfer learning method with XGBoost yields the best forecasting performance among all model classes. predictive-accuracy
    Source sentence
    our proposed transfer learning method in combination with XGBoost models shows superior forecasting performance compared to any other model class or training data set approach.
  • In the first 100‑day sample period, MTL models improve average MSE by up to 44.2% and average MAE by up to 23.2% relative to the naive‑pooling (NF) benchmark. predictive-accuracy
    Source sentence
    In the first sample period, the MTL models yield improvements in forecast accuracy of up to 44.2% in the average MSE and up to 23.2% in the average MAE.
  • MTL models dominate across all individual 100‑day evaluation periods, consistently providing more accurate realized variance predictions than the NF approach. performance
    Source sentence
    Every MTL model provides, on average, more accurate realized variance predictions than the NF in each sample period, considering both MSE and MAE.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 35 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 (75)

  1. A Simple Approximate Long-Memory Model of Realized Volatility (2009) — link
  2. Forecasting realized volatility: The role of implied volatility, leverage effect, overnight returns, and volatility of realized volatility (2021) — link
  3. A transfer forecasting model for container throughput guided by discrete PSO (2014) — link
  4. Transfer learning for time series anomaly detection (2017) — link
  5. Transfer learning for time series classification (2018) — link
  6. Xgboost: A scalable tree boosting system (2016) — link
  7. A decision-theoretic generalization of on-line learning and an application to boosting (1997) — link
  8. Forecasting of realised volatility with the random forests algorithm (2018) — link
and 67 more
  1. Adam: A Method for Stochastic Optimization (2015) — link
  2. Adaptive Subgradient Methods for Online Learning and Stochastic Optimization (2011) — link
  3. Long short-term memory (1997) — link
  4. Short-to Long-Term Realized Volatility Forecasting using Extreme Gradient Boosting (2022) — link
  5. A Machine Learning Approach to Volatility Forecasting (2022) — link
  6. Convolutional Networks for Images, Speech, and Time Series (1995)
  7. Learning representations by back-propagating errors (1986) — link
  8. Deep learning using rectified linear units (relu) (2019) — link
  9. Implementing transfer learning across different datasets for time series forecasting (2021) — link
  10. Attention is All you Need (2017) — link
  11. The Distribution of Realized Exchange Rate Volatility (2001) — link
  12. On adaptive control processes (1959) — link
  13. The model confidence set (2011)
  14. Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation (1982) — link
  15. Generalized autoregressive conditional heteroskedasticity (1986) — link
  16. A comprehensive look at financial volatility prediction by economic variables (2012) — link
  17. Empirical evidence on the importance of aggregation, asymmetry, and jumps for volatility prediction (2015) — link
  18. Stock market volatility: Identifying major drivers and the nature of their impact (2015) — link
  19. A Survey on Transfer Learning (2010) — link
  20. Realized Volatility Forecasting with Neural Networks (2020) — link
  21. Approximation by superposition of sigmoidal functions (1989) — link
  22. Gradient and newton boosting for classification and regression (2021) — link
  23. Financial returns modelled by the product of two stochastic processes-a study of the daily sugar prices 1961-75 (1982)
  24. Global industry classification standard methodology (2023) — link
  25. Overnight stock returns and realized volatility (2013) — link
  26. Using neural networks for forecasting volatility of S&P 500 Index futures prices (2004) — link
  27. Keratinocytic Skin Cancer Detection on the Face Using Region-Based Convolutional Neural Network (2020) — link
  28. Fast and Robust Neural Network Joint Models for Statistical Machine Translation (2014) — link
  29. HD-CNN: Hierarchical Deep Convolutional Neural Networks for Large Scale Visual Recognition (2015) — link
  30. Greedy Function Approximation: A Gradient Boosting Machine (2001)
  31. A survey of transfer learning (2016) — link
  32. Transfer Learning With Time Series Data: A Systematic Mapping Study (2021) — link
  33. The Pricing of Options and Corporate Liabilities (1973) — link
  34. A novel deep transfer learning framework with adversarial domain adaptation: application to financial time-series forecasting (2023) — link
  35. Domain Adaptation Techniques for EEG-Based Emotion Recognition: A Comparative Study on Two Public Datasets (2019) — link
  36. DALSA: Domain Adaptation for Supervised Learning From Sparsely Annotated MR Images (2016) — link
  37. Active multi-kernel domain adaptation for hyperspectral image classification (2018) — link
  38. Data Scaling Effect of Deep Learning in Financial Time Series Forecasting (2024) — link
  39. Real-Time Measurement of Business Conditions (2008) — link
  40. Measuring Economic Policy Uncertainty* (2016) — link
  41. Econometric Analysis of Realized Volatility and its Use in Estimating Stochastic Volatility Models (2002) — link
  42. Answering the skeptics: Yes, standard volatility models do provide accurate forecasts (1998) — link
  43. Roughing it up: Including jump components in the measurement, modeling, and forecasting of return volatility (2007) — link
  44. Discrete-time volatility forecasting with persistent leverage effect and the link with continuous-time volatility modeling (2012) — link
  45. Good volatility, bad volatility: Signed jumps and the persistence of volatility (2015) — link
  46. Instance Weighting for Domain Adaptation in NLP (2007) — link
  47. Domain Adaptation with Structural Correspondence Learning (2006) — link
  48. Cross-Domain Co-Extraction of Sentiment and Topic Lexicons (2012) — link
  49. Inter-subject transfer learning with an end-to-end deep convolutional neural network for EEG-based BCI (2019) — link
  50. Boosting for transfer learning (2007) — link
  51. Multi-Task Feature Learning (2006) — link
  52. Evaluation of Transfer Learning for Human Activity Recognition Among Different Datasets (2019) — link
  53. Characterizing and Avoiding Negative Transfer (2019) — link
  54. SpotTune: Transfer Learning Through Adaptive Fine-Tuning (2019) — link
  55. Long memory versus structural breaks in modeling and forecasting realized volatility (2010) — link
  56. Realized Volatility Forecast: Structural Breaks, Long Memory, Asymmetry, and Day-of-the-Week Effect (2014) — link
  57. Instance-based deep transfer learning with attention for stock movement prediction (2023) — link
  58. Multi-source Transfer Learning with Ensemble for Financial Time Series Forecasting (2020) — link
  59. Forecasting realized volatility in turbulent times using temporal fusion transformers (2023) — link
  60. Time Series Prediction Method of Industrial Process With Limited Data Based on Transfer Learning (2023) — link
  61. Parameter-efficient transfer learning for NLP (2019) — link
  62. A Study on CNN Transfer Learning for Image Classification (2019) — link
  63. 3-Month Treasury Bill Secondary Market Rate, Discount Basis [DTB3] (2023) — link
  64. Aruoba-Diebold-Scotti Business Conditions Index (2023) — link
  65. Economic Policy Uncertainty Index (2023) — link
  66. Historical Data for Cboe VIX Index and Other Volatility Indices (2023) — link
  67. Everything you know about dynamic time warping is wrong (2004) — link

0 of 75 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.