Paper: arXiv 2501.01278

Authors: Nico Herrig

Abstract

This work aims to implement Long Short-Term Memory mixture density networks (LSTM-MDNs) for Value-at-Risk forecasting and compare their performance with established models (historical simulation, CMM, and GARCH) using a defined backtesting procedure. The focus was on the neural network’s ability to capture volatility clustering and its real-world applicability. Three architectures were tested: a 2-component mixture density network, a regularized 2-component model (Arimond et al., 2020), and a 3-component mixture model, the latter being tested for the first time in Value-at-Risk forecasting. Backtesting was performed on three stock indices (FTSE 100, S&P 500, EURO STOXX 50) over two distinct two-year periods (2017-2018 as a calm period, 2021-2022 as turbulent). Model performance was assessed through unconditional coverage and independence assumption tests. The neural network’s ability to handle volatility clustering was validated via correlation analysis and graphical evaluation. Results show limited success for the neural network approach. LSTM-MDNs performed poorly for 2017/2018 but outperformed benchmark models in 2021/2022. The LSTM mechanism allowed the neural network to capture volatility clustering similarly to GARCH models. However, several issues were identified: the need for proper model initialization and reliance on large datasets for effective learning. The findings suggest that while LSTM-MDNs provide adequate risk forecasts, further research and adjustments are necessary for stable performance.

Complexity vs Empirical Score

  • Math Complexity: 5.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper uses advanced neural network architectures (LSTM, Mixture Density Networks) and statistical backtesting frameworks, balancing mathematical theory with practical implementation on real financial data across multiple market regimes.

Research Flowchart

  flowchart TD
A["Research Goal<br>Forecast VaR using LSTM-MDNs<br>vs. Benchmarks (Hist. Sim, CMM, GARCH)"] --> B["Methodology<br>Test 3 Architectures: 2-comp, Reg-2-comp, 3-comp<br>Backtest via Unconditional Coverage & Independence"]
A --> C["Data/Inputs<br>3 Indices: FTSE 100, S&P 500, EURO STOXX 50<br>2 Periods: Calm (2017-18), Turbulent (2021-22)"]
B & C --> D["Computation<br>Train LSTM-MDNs & Benchmarks<br>Calculate VaR & Correlations"]
D --> E["Key Findings & Outcomes<br>Poor Performance (2017/2018)<br>Outperformed Benchmarks (2021/2022)<br>LSTM captured volatility clustering (like GARCH)<br>Issue: Needs large data & proper initialization"]

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 implements Long Short-Term Memory mixture density networks (LSTM‑MDNs) for Value‑at‑Risk forecasting and compares them to established benchmark models.

Source sentence
This work aims to implement Long Short-Term Memory mixture density networks (LSTM-MDNs) for Value-at-Risk forecasting and compare their performance with established models (historical simulation, CMM, and GARCH) using a defined backtesting procedure.

Assumes 5

  • VaR is defined with confidence level α=0.99 and horizon h=1 day. model
    Source sentence
    estimate V aR[α = 0.99, h = 1], where h denotes the number of days in the future.
  • Daily closing prices from Yahoo Finance accurately represent market returns. data
    Source sentence
    Daily stock price data is sourced from Yahoo Finance.
  • Large training datasets are required for effective learning. data
    Source sentence
    LSTM-MDNs provide adequate risk forecasts, further research and adjustments are necessary for stable performance.
  • Proper model initialization is necessary for reliable forecasts. model
    Source sentence
    Findings show that model performance is highly dependent on an appropriate initialization procedure
Show 1 more
  • Volatility clustering can be captured via the LSTM recurrent mechanism. market
    Source sentence
    The LSTM mechanism allowed the neural network to capture volatility clustering similarly to GARCH models.

Data

Backtest sample: FTSE 100, S&P 500, EURO STOXX 50 · equities · daily

  • FTSE 100 daily closing prices · Yahoo Finance market-prices public
    2017‑2018 and 2021‑2022 · FTSE 100 index
    Source sentence
    The FTSE 100 index is composed of the 100 largest companies listed on the London Stock Exchange
  • S&P 500 daily closing prices · Yahoo Finance market-prices public
    2017‑2018 and 2021‑2022 · S&P 500 index
    Source sentence
    The analysis is concentrated on three different stock indices, which represent different geographical regions to avoid any potential geographical dependencies.
  • EURO STOXX 50 daily closing prices · Yahoo Finance market-prices public
    2017‑2018 and 2021‑2022 · EURO STOXX 50 index
    Source sentence
    The analysis is concentrated on three different stock indices, which represent different geographical regions to avoid any potential geographical dependencies.

Recognised sources: Yahoo Finance

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • LSTM‑MDNs performed poorly for the calm period (2017‑2018) but outperformed benchmark models in the turbulent period (2021‑2022). performance
    Source sentence
    Results show limited success for the neural network approach. LSTM-MDNs performed poorly for 2017/2018 but outperformed benchmark models in 2021/2022.
  • The LSTM mechanism allowed the neural network to capture volatility clustering similarly to GARCH models. predictive-accuracy
    Source sentence
    The LSTM mechanism allowed the neural network to capture volatility clustering similarly to GARCH models.
  • Model performance is highly dependent on proper initialization and large training datasets. statistical
    Source sentence
    Further, results indicate that LSTM-MDNs require a large amount of training data to appropriately account and "learn" underlying patterns from the data.
  • No striking difference in forecast accuracy was observed between the three LSTM‑MDN architectures. performance
    Source sentence
    Comparing the three differen LSTM-MDN architectures, the backtesting results show no striking difference in forecast accuracy between the three model architectures nor any other pattern.
  • The neural network does not react to volatility switches better than a GARCH(1,1) model. negative
    Source sentence
    However, a correlation analysis (see table 5.2) does not indicate a better reactivity than a GARCH(1,1) model.

Checked by others

Supported by

  • Global Neural Networks and The Data Scaling Effect in Financial Time Series Forecasting (2023)
    On the question: Does increasing the training dataset size improve forecasting accuracy of neural network models?
    Compare the findings

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

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

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

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 (92)

  1. Neural Networks and Value at risk (2020) — link
  2. Value at Risk Estimation with Neural Networks: A Recurrent Mixture Density Approach (2021)
  3. Machine learning for financial risk management: a survey (2020)
  4. Machine learning techniques for anti-money laundering (AML) solutions in suspicious transaction detection: a review (2018)
  5. Machine learning applications in healthcare sector: An overview (2022)
  6. The Present and Future of Financial Risk Management (2005) — link
  7. Basel 2.5: potential benefits and unintended consequences (2013)
  8. Good and bad uncertainty: Macroeconomic and financial market implications (2015)
and 84 more
  1. History of Value-at-Risk (2002)
  2. GARCHNet: Value-at-Risk Forecasting with GARCH Models Based on Neural Networks (2023)
  3. Value at risk models in finance (2001)
  4. Value-at-risk support vector machine: stability to outliers (2014)
  5. Encoded Value-at-Risk: A machine learning approach for portfolio risk measurement (2022)
  6. Mixture density networks (1994)
  7. The real effects of financial markets (2012)
  8. Volatility and correlation forecasting (2006)
  9. The econometrics of financial markets (1998)
  10. Volatility clustering in financial markets: empirical facts and agent-based models (2007)
  11. Characterization of financial time series (2011)
  12. Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation (1982)
  13. Generalized autoregressive conditional heteroskedasticity (1986)
  14. Evaluation of Basel III revision of quantitative standards for implementation of internal models for market risk (2012)
  15. Implementing value at risk (2000)
  16. An evaluation of the effectiveness of Value-at-Risk (VaR) models for Australian banks under Basel III (2016)
  17. A Comparison between Historical Simulation and GARCH Model for Estimating Value-at-Risk: Case of Indian Stock Exchange (2017)
  18. The level and quality of Value-at-Risk disclosure by commercial banks (2010)
  19. Stylized facts of financial time series and three popular models of volatility (2010)
  20. Modelling volatility of stock returns: Is GARCH (1, 1) enough (2014)
  21. A forecast comparison of volatility models: does anything beat a GARCH (1, 1)? (2005)
  22. Backtesting (2008)
  23. A review of backtesting and backtesting procedures (2005)
  24. Techniques for Verifying the Accuracy of Risk Measurement Models (1995)
  25. Evaluating interval forecasts (1998)
  26. Modelling naira/dollar exchange rate volatility: Application of GARCH and asymmetric models (2009)
  27. Artificial neural network (2013)
  28. The perceptron: a probabilistic model for information storage and organization in the brain. (1958)
  29. Vanilla feedforward neural networks as a discretization of dynamic systems (2022) — link
  30. Activation functions in neural networks (2017)
  31. How to Choose an Activation Function for Deep Learning (2021)
  32. How to Choose an Activation Function for Deep Learning (2021)
  33. A review on weight initialization strategies for neural networks (2022)
  34. Cme 323: Distributed algorithms and optimization, spring 2015 (2015)
  35. Adam: A method for stochastic optimization (2014) — link
  36. Why tanh: choosing a sigmoidal function (1992)
  37. A Gentle Introduction to the Rectified Linear Unit (ReLU) (2019) — link
  38. CORDIC based implementation of the softmax activation function (2020)
  39. Fast and accurate deep network learning by exponential linear units (elus) (2015) — link
  40. Parametric exponential linear unit for deep convolutional neural networks (2017)
  41. Dying relu and initialization: Theory and numerical examples (2019) — link
  42. How do mixture density RNNs predict the future? (2019) — link
  43. Gaussian mixture models. (2009)
  44. Mixture density networks for distribution and uncertainty estimation (2017)
  45. Gaussian Mixture Models Part 1: Mixture Density Networks (2020) — link
  46. Chapter 10, Sequence Modeling: Recurrent and Recursive Nets (2016)
  47. Fundamentals Of Neural Networks: Architectures, Algorithms, and Applications.., New Jersey, USA: Prenctice-Hall (1994)
  48. Generating sequences with recurrent neural networks (2013) — link
  49. Long short-term memory (1997)
  50. LSTMs Explained: A Complete, Technically Accurate, Conceptual Guide with Keras (2020) — link
  51. Understanding LSTM networks (2015) — link
  52. Towards movement generation with audio features (2020) — link
  53. Coronavirus disease (covid-19) pandemic (2023) — link
  54. Circular economy impact analysis on stock performances: an empirical comparison with the euro stoxx 50 \textregistered ESG index (2022)
  55. Trust and stock market volatility during the COVID-19 crisis (2021)
  56. Deep learning (2016)
  57. Effects of hidden layers on the efficiency of neural networks (2020)
  58. Empirical asset pricing via machine learning (2020)
  59. On the optimal number of hidden nodes in a neural network (1998)
  60. Components of bull and bear markets: bull corrections and bear rallies (2012)
  61. Regime switching in stock market returns (1997)
  62. Clustering financial time series: New insights from an extended hidden Markov model (2015)
  63. The loss surfaces of multilayer networks (2015)
  64. On the saddle point problem for non-convex optimization (2014) — link
  65. Second-order optimization for non-convex machine learning: An empirical study (2020)
  66. An empirical analysis of data requirements for financial forecasting with neural networks (2001)
  67. Understanding the role of nonlinearity in training dynamics of contrastive learning (2022) — link
  68. On weight initialization in deep neural networks (2017) — link
  69. All you need is a good init (2015) — link
  70. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification (2015)
  71. Understanding the difficulty of training deep feedforward neural networks (2010)
  72. Financial Time Series Data Augmentation with Generative Adversarial Networks and extended Intertemporal Return Plots (2022) — link
  73. Evaluating data augmentation for financial time series classification (2020) — link
  74. Street View - Estimating Global Investor Views with Reverse Optimization (2020) — link
  75. Quantitative risk management: concepts, techniques and tools-revised edition (2015)
  76. Estimating value at risk: LSTM vs. GARCH (2022) — link
  77. Efficient Monte Carlo methods for value-at-risk (2000)
  78. Small sample size effects in statistical pattern recognition: Recommendations for practitioners (1991)
  79. R: A Language and Environment for Statistical Computing (2021) — link
  80. tseries: Time Series Analysis and Computational Finance (2020) — link
  81. Rmetrics - Autoregressive Conditional Heteroskedastic Modelling (2022) — link
  82. Python 3 Reference Manual (2009)
  83. Keras (2015) — link
  84. TensorFlow : Large-Scale Machine Learning on Heterogeneous Systems (2015) — link

0 of 92 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.