Paper: arXiv 2309.02072

Authors: Chen Liu, Minh-Ngoc Tran, Chao Wang, Richard Gerlach, Robert Kohn

Abstract

Neural networks have revolutionized many empirical fields, yet their application to financial time series forecasting remains controversial. In this study, we demonstrate that the conventional practice of estimating models locally in data-scarce environments may underlie the mixed empirical performance observed in prior work. By focusing on volatility forecasting, we employ a dataset comprising over 10,000 global stocks and implement a global estimation strategy that pools information across cross-sections. Our econometric analysis reveals that forecasting accuracy improves markedly as the training dataset becomes larger and more heterogeneous. Notably, even with as little as 12 months of data, globally trained networks deliver robust predictions for individual stocks and portfolios that are not even in the training dataset. Furthermore, our interpretation of the model dynamics shows that these networks not only capture key stylized facts of volatility but also exhibit resilience to outliers and rapid adaptation to market regime changes. These findings underscore the importance of leveraging extensive and diverse datasets in financial forecasting and advocate for a shift from traditional local training approaches to integrated global estimation methods.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced econometric and neural network formulations with formal likelihood specifications, while backing them with a massive 10,000-stock dataset, out-of-sample generalization tests, and a provided implementation notebook, bridging theoretical depth with rigorous empirical validation.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Global vs. Local NNs in Financial Time Series"] --> B{"Key Methodology"}
  
  B --> B1["Dataset: 10k+ Global Stocks"]
  B --> B2["Focus: Volatility Forecasting"]
  B --> B3["Strategy: Global Estimation<br>Pooling Cross-sectional Info"]
  
  B1 & B2 & B3 --> C{"Computational Process"}
  C --> C1["Train Neural Network<br>on Large Heterogeneous Data"]
  
  C1 --> D["Key Findings / Outcomes"]
  
  D --> D1["Accuracy ↑ as<br>Training Data Increases"]
  D --> D2["Robust Predictions<br>on Unseen Stocks/Portfolios"]
  D --> D3["Captures Stylized Facts<br>Adapts to Regime Changes"]

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 global training approach for volatility forecasting that pools information across thousands of stocks.

Source sentence
By focusing on volatility forecasting, we employ a dataset comprising over 10,000 global stocks and implement a global estimation strategy that pools information across cross-sections.

Assumes 5

  • Conditional return variance follows a Gaussian error distribution. model
    Source sentence
    In volatility modeling, the key quantity of interest is the conditional return variance, σ2 t = var(yt | Ft−1), where Ft−1 contains a set of available information up to time t −1, which, in our case, is the past returns. A local volatility model, assuming Gaussian errors, is specified as yt = σtϵt, ϵt i.i.d ∼N(0, 1),
  • All series share a common conditional volatility function f* in the global model. model
    Source sentence
    Global training uses several time series to estimate a shared set of parameters, forming a global model.
  • Stocks with fewer than 1,260 observations are excluded from training. data
    Source sentence
    To ensure the local models can be effectively trained, stocks with fewer than 1,260 (five years) observations during the training period are excluded, leaving a total of 11,771 stocks.
  • Training, validation, and testing periods are non‑overlapping calendar years. data
    Source sentence
    period spans from 2014 to 2019, the validation period spans from 2020 to 2021, and the testing period spans from 2022 to 2023.
Show 1 more
  • When evaluating supervised forecasting for the data‑scaling study, only the 10 smallest‑training‑set stocks are used. execution
    Source sentence
    For supervised forecasting, evaluation is restricted to the 10 stock series in the smallest training set rather than all 10,240 training series.

Data

  • Global stock dataset market-prices unknown
    2014-2023 · 11,771 global stocks
    Source sentence
    Of the 11,771 stock series in our dataset, 10,240 are allocated for training and

Recognised sources: Refinitiv / LSEG (Datastream, Eikon, Tick History)

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Forecasting accuracy improves markedly as the training dataset becomes larger and more heterogeneous. predictive-accuracy
    Source sentence
    Our econometric analysis reveals that forecasting accuracy improves markedly as the training dataset becomes larger and more heterogeneous.
  • Even with as little as 12 months of data, globally trained networks deliver robust predictions for individual stocks and portfolios not in the training set. predictive-accuracy
    Source sentence
    Notably, even with as little as 12 months of data, globally trained networks deliver robust predictions for individual stocks and portfolios that are not even in the training dataset.
  • Global models trained on a similarly sized but more diverse dataset generally outperform group‑specific models. performance
    Source sentence
    The results indicate that the global model, trained on a similarly sized but more diverse dataset, generally outperforms the group-specific models.
  • The global training approach and adequate data size are essential for the success of neural‑network models in finance. theoretical
    Source sentence
    Our findings show that the global training approach and adequate data size are essential for the success of NN models.
  • When employing flexible models such as neural networks, stock series should be effectively treated as homogeneous and modeled together for optimal performance. theoretical
    Source sentence
    Moreover, the results again highlight that, when employing flexible models such as neural networks, stock series should be effectively treated as homogeneous and modeled together for optimal performance.

Checked by others

Supported by

  • Empirical analysis in limit order book modeling for Nikkei 225 Stocks with Cox-type intensities (2023)
    On the question: Does using more frequent or larger training data improve forecasting accuracy of limit‑order‑book models?
    Compare the findings

    This paper: Forecasting accuracy improves markedly as the training dataset becomes larger and more heterogeneous.

    That paper: Calibration every 1‑2 weeks (instead of daily) can increase accuracy from ~77% to ~78%, indicating benefit from more training data.

    Both findings show accuracy gains with more data: A’s 1‑2‑week calibration raises accuracy from ~77% to ~78%; B reports marked improvement as training set grows and diversifies, despite different markets.

  • A hidden Markov model for statistical arbitrage in international crude oil futures markets (2023)
    On the question: How does training sample size affect model performance in statistical arbitrage and neural‑network forecasting?
    Compare the findings

    This paper: Forecasting accuracy improves markedly as the training dataset becomes larger and more heterogeneous.

    That paper: A minimum sample size is required for accurate estimation; smaller training samples weaken cointegration and reduce performance.

    Both agree larger training sets improve performance; A warns a minimum size is needed to preserve cointegration, B shows accuracy rises with bigger, more heterogeneous data.

  • Risk forecasting using Long Short-Term Memory Mixture Density Networks (2025)
    On the question: Does increasing the training dataset size improve forecasting accuracy of neural network models?
    Compare the findings

    This paper: Forecasting accuracy improves markedly as the training dataset becomes larger and more heterogeneous.

    That paper: Model performance is highly dependent on proper initialization and large training datasets.

    Both findings agree that larger, more heterogeneous training sets boost performance; B additionally stresses proper initialization.

Qualified by

  • Which Past Matters? Training History and Multi-Forecast Horizon in Equity Transformers (2026)
    On the question: Does increasing the size and heterogeneity of training data improve forecasting accuracy?
    Compare the findings

    This paper: Forecasting accuracy improves markedly as the training dataset becomes larger and more heterogeneous.

    That paper: Increasing training quantity does not uniformly improve accuracy; the benefit of larger samples is conditional on the sampling family and horizon.

    A finds accuracy improves markedly with larger, more heterogeneous data; B finds larger samples help only conditionally, depending on sampling family and forecast horizon.

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.