Paper: arXiv 2512.12250

Authors: Anna Perekhodko, Robert Ślepaczuk

Abstract

Accurate volatility forecasting is essential in banking, investment, and risk management, because expectations about future market movements directly influence current decisions. This study proposes a hybrid modelling framework that integrates a Stochastic Volatility model with a Long Short Term Memory neural network. The SV model improves statistical precision and captures latent volatility dynamics, especially in response to unforeseen events, while the LSTM network enhances the model’s ability to detect complex nonlinear patterns in financial time series. The forecasting is conducted using daily data from the S and P 500 index, covering the period from January 1 1998 to December 31 2024. A rolling window approach is employed to train the model and generate one step ahead volatility forecasts. The performance of the hybrid SV-LSTM model is evaluated through both statistical testing and investment simulations. The results show that the hybrid approach outperforms both the standalone SV and LSTM models and contributes to the development of volatility modelling techniques, providing a foundation for improving risk assessment and strategic investment planning in the context of the S and P 500.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced statistical methodology (Stochastic Volatility modeling, statistical tests like Diebold-Mariano) combined with complex deep learning (LSTM networks), while demonstrating high empirical rigor through a backtest-ready design using long-term S&P 500 data, a rolling window approach, and providing open-source code on GitHub.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Forecast S&P 500 Volatility"] --> B["Data Input:<br>Jan 1998 - Dec 2024 Daily Data"]
  B --> C{"Hybrid Methodology"}
  C --> D["Stochastic Volatility SV<br>Captures Latent Dynamics"]
  C --> E["LSTM Neural Network<br>Identifies Nonlinear Patterns"]
  D & E --> F["Rolling Window Training<br>One-Step Ahead Forecasts"]
  F --> G{"Evaluation"}
  G --> H["Statistical Testing<br>Superior Accuracy"]
  G --> I["Investment Simulation<br>Enhanced Risk Management"]

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 hybrid modelling framework that combines a Stochastic Volatility model with a Long Short‑Term Memory neural network for S&P 500 volatility forecasting.

Source sentence
This study proposes a hybrid modelling framework that integrates a Stochastic Volatility model with a Long Short Term Memory neural network.

Assumes 4

  • Log‑returns are stationary, which is required for volatility modelling. statistical
    Source sentence
    enabling stationarity in the return series, a critical assumption for volatility modeling.
  • The SV model parameters are estimated with weakly informative priors rather than strong priors. model
    Source sentence
    we adopt a data‑driven approach by employing weakly informative default priors, as inspired by Kim et al. (1998), rather than imposing strong prior distributions
  • Rolling‑window training uses the most recent 504 days to forecast the next day. execution
    Source sentence
    Model training and evaluation are conducted across 11 overlapping rolling windows covering the period from 2000 to 2024. For each window, the model is trained on the most recent 504 days and generates forecasts for the next day (t + 1)
  • Trading signals are generated at the daily close price with a flat transaction cost of 0.1% per side. execution
    Source sentence
    Transaction costs of 0.1% are applied at both entry and exit.

Data

Backtest sample: 2014-01 – 2024-12 · S&P 500 · equities · daily

  • S&P 500 daily closing prices · Yahoo Finance market-prices public
    1998-01-01 to 2024-12-31 · S&P 500 index
    Source sentence
    For the modeling of S&P 500 index volatility, the dataset consisted of daily closing prices obtained from Yahoo Finance, spanning January 1, 1998, to December 31, 2024.
  • VIX monthly futures contracts · CBOE options public
    2014-01 to 2024-12 · VIX monthly futures
    Source sentence
    The selected data consisted of VIX monthly futures contracts, collected from the CBOE website for the entire testing period, from January 2014 to December 2024.

Recognised sources: Yahoo Finance Cboe

Costs

Modeled Costs are part of the reported headline results.

A flat fee of 0.1% is applied at both entry and exit of each VIX futures trade.

Source sentence
Transaction costs of 0.1% are applied at both entry and exit.

Finds 3

Reported performance (Hybrid SV-LSTM): Sharpe -0.56 · annual return -14.6% · total return -82.0% · max drawdown 85.2%

  • The hybrid SV‑LSTM model achieves a Mean Absolute Percentage Error of 4.75% on out‑of‑sample 21‑day rolling volatility forecasts (Jan 2014‑Dec 2024). predictive-accuracy
    Source sentence
    The Mean Absolute Percentage Error (MAPE) of 4.75% indicates high relative accuracy and reflects the hybrid’s ability to closely track proportional volatility changes.
  • The hybrid model outperforms both standalone SV and LSTM models according to statistical tests (Diebold‑Mariano and Wilcoxon). statistical
    Source sentence
    The error metrics demonstrated the superiority of the SV‑LSTM model over both the LSTM and SV models. Additionally, the Diebold‑Mariano (DM) test confirmed that the SV‑LSTM model exhibits better predictive accuracy than both benchmark models.
  • The investment‑strategy simulation using VIX futures shows that the SV‑LSTM signals generate cumulative equity curves that exceed naive long‑only and short‑only benchmarks. performance
    Source sentence
    The investment strategy simulation further validated the practical relevance of the SV‑LSTM model by assessing its performance on VIX futures data spanning from 2014 to 2024 in a realistic market environment.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 26 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (23)

  1. Deep Learning for Financial Time Series Forecasting (2019) — link
  2. The Pricing of Options and Corporate Liabilities (1973) — link
  3. Generalized Autoregressive Conditional Heteroskedasticity (1986) — link
  4. The Detection and Estimation of Long Memory in Stochastic Volatility (1998) — link
  5. Realized Volatility Forecasting with Neural Networks (2020) — link
  6. Comparing Predictive Accuracy (2002) — link
  7. Autoregressive Conditional Heteroskedasticity with Estimates of the Variance of United Kingdom Inflation (1982) — link
  8. GARCH 101: The Use of ARCH/GARCH Models in Applied Econometrics (2001) — link
and 15 more
  1. Application of machine learning in algorithmic investment strategies on global stock markets (2023) — link
  2. A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options (1993) — link
  3. Long Short-Term Memory (1997) — link
  4. Stochastic Volatility: Likelihood Inference and Comparison with ARCH Models (1998) — link
  5. Forecasting the volatility of stock price index: A hybrid model integrating LSTM with multiple GARCH-type models (2018) — link
  6. Long memory and nonlinearity in conditional variances: A smooth transition FIGARCH model (2011) — link
  7. LSTM in Algorithmic Investment Strategies on BTC and S&P500 Index (2022) — link
  8. Conditional Heteroskedasticity in Asset Returns: A New Approach (1991) — link
  9. A long short-term memory stochastic volatility model (2019) — link
  10. The Hybrid Forecast of S&P 500 Volatility ensembled from VIX, GARCH and LSTM models — → 2407.16780
  11. Financial Returns Modelled by the Product of Two Stochastic Processes - A Study of Daily Sugar Prices 1961-79 (1982)
  12. Individual Comparisons by Ranking Methods (1945) — link
  13. Modelling Financial Time Series (1986)
  14. Forecasting volatility in the New Zealand stock market (2002) — link
  15. Mean Absolute Directional Loss as a New Loss Function for Machine Learning Problems in Algorithmic Investment Strategies — → 2309.10546

2 of 23 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.