Paper: arXiv 2508.14999

Authors: Maciej Wysocki, Paweł Sakowski

Abstract

This paper investigates an important problem of an appropriate variance-covariance matrix estimation in the Modern Portfolio Theory. We propose a novel framework for variancecovariance matrix estimation for purposes of the portfolio optimization, which is based on deep learning models. We employ the long short-term memory (LSTM) recurrent neural networks (RNN) along with two probabilistic deep learning models: DeepVAR and GPVAR to the task of one-day ahead multivariate forecasting. We then use these forecasts to optimize portfolios of stocks and cryptocurrencies. Our analysis presents results across different combinations of observation windows and rebalancing periods to compare performances of classical and deep learning variance-covariance estimation methods. The conclusions of the study are that although the strategies (portfolios) performance differed significantly between different combinations of parameters, generally the best results in terms of the information ratio and annualized returns are obtained using the LSTM-RNN models. Moreover, longer observation windows translate into better performance of the deep learning models indicating that these methods require longer windows to be able to efficiently capture the long-term dependencies of the variance-covariance matrix structure. Strategies with less frequent rebalancing typically perform better than these with the shortest rebalancing windows across all considered methods.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced math including stochastic processes (LSTM, VAR, GPVAR) and rigorous portfolio optimization frameworks, and demonstrates strong empirical rigor through extensive backtesting on stocks and cryptocurrencies with various parameters, code sharing for reproducibility, and detailed performance metrics like information ratio.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Optimize Portfolio via<br>Improved Covariance Estimation"] --> B["Data Inputs<br>Multi-Asset (Stocks & Crypto)"]
  B --> C{"Model Comparison<br>Observation Windows vs Rebalancing"}
  C --> D["Classical Methods<br>e.g., Sample Covariance"]
  C --> E["Deep Learning Methods<br>LSTM, DeepVAR, GPVAR"]
  D & E --> F["Portfolio Optimization<br>Mean-Variance Framework"]
  F --> G["Performance Metrics<br>Information Ratio &<br>Annualized Returns"]
  G --> H["Key Findings<br>LSTM Best Performance<br>Longer Windows Preferred<br>Less Frequent Rebalancing"]

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 novel deep‑learning‑based framework for estimating variance‑covariance matrices for portfolio optimization.

Source sentence
We propose a novel framework for variance‑covariance matrix estimation for purposes of the portfolio optimization, which is based on deep learning models.

Assumes 6

  • A fixed percentage commission of 50 basis points is applied to every asset when rebalancing. execution
    Source sentence
    we have decided to use such solution and consider all costs of changing position assuming a fixed percentage commission of 50 bps for every asset.
  • 365 trading days per year are used in all calculations. data
    Source sentence
    For every calculation we assumed 365 trading days per year.
  • Missing stock quotes on weekends/holidays are forward‑filled with the last available price. data
    Source sentence
    the missing stock quotes for these days were filled with the last available price
  • Median values of the Monte‑Carlo forecasts are used as point forecasts. model
    Source sentence
    In this study we decided to use median values as the forecasts, as it is a middle quantile and a commonly undertaken approach in probabilistic forecasting
Show 2 more
  • Each rebalancing date the portfolio is built from the 10 stocks and 10 cryptocurrencies with the highest market capitalization at that date. data
    Source sentence
    10 stocks and 10 cryptocurrencies with the highest market capitalization at a given date were selected and taken into the portfolio construction process
  • Longer observation windows improve deep‑learning model performance because they capture long‑term dependencies. model
    Source sentence
    longer observation windows translate into better performance of the deep learning models indicating that these methods require longer windows to be able

Data

  • Cryptocurrency daily close prices · Coinpaprika market-prices public
    2018-01-01 to 2021-10-31 · around 500 cryptocurrencies, top 10 by market cap each date
    Source sentence
    The dataset used in this study consists of daily close prices of cryptocurrencies
  • US stock daily close prices · Yahoo Finance market-prices public
    2018-01-01 to 2021-10-31 · around 1000 NYSE stocks, top 10 by market cap each date
    Source sentence
    The dataset used in this study consists of daily close prices of

Recognised sources: CoinMarketCap Yahoo Finance

Costs

Modeled Costs are part of the reported headline results.

fixed percentage commission of 50 bps per asset applied to all position changes

Source sentence
assuming a fixed percentage commission of 50 bps for every asset

Finds 5

  • LSTM‑RNN models deliver the highest information ratios and annualized returns among all variance‑covariance estimation methods. performance
    Source sentence
    generally the best results in terms of the information ratio and annualized returns are obtained using the LSTM‑RNN models
  • Longer observation windows (e.g., 90‑day and 120‑day) improve the profitability and information ratios of deep‑learning‑based strategies. performance
    Source sentence
    The 90‑days window was the first setup when so many strategies provided positive returns and information rations
  • Strategies with less frequent rebalancing (e.g., 30‑, 60‑, 90‑day) outperform those with very short rebalancing periods. performance
    Source sentence
    Strategies with less frequent rebalancing typically perform better than these with the shortest rebalancing windows across all considered methods
  • Deep learning models require longer historical windows to capture short‑ and long‑term dependencies in the variance‑covariance matrix. statistical
    Source sentence
    longer observation windows translate into better performance of the deep learning models indicating that these methods require longer windows to be able
  • Classical variance‑covariance estimators produce unsatisfactory, often negative, returns and information ratios, especially with longer rebalancing periods. negative
    Source sentence
    All strategies based on the classical variance‑covariance matrix estimation methods provided rather similar results

Checked by others

Supported by

  • Regret-Driven Portfolios: LLM-Guided Smart Clustering for Optimal Allocation (2026)
    On the question: What rebalancing frequency leads to superior portfolio performance?
    Compare the findings

    This paper: Strategies with less frequent rebalancing (e.g., 30‑, 60‑, 90‑day) outperform those with very short rebalancing periods.

    That paper: 71% of configurations that outperformed SPY used a quarterly rebalancing schedule, indicating medium‑term rebalancing is beneficial.

    Both find medium‑term rebalancing optimal: A shows 30‑,60‑,90‑day periods outperform very short intervals; B shows 71% of out‑performing configs use quarterly rebalancing.

  • Semiparametric Dynamic Copula Models for Portfolio Optimization (2025)
    On the question: Which rebalancing frequency leads to better portfolio performance?
    Compare the findings

    This paper: Strategies with less frequent rebalancing (e.g., 30‑, 60‑, 90‑day) outperform those with very short rebalancing periods.

    That paper: Weekly rebalancing with the copula‑based method yields higher gains than daily rebalancing, especially in volatile periods.

    Both findings show that longer intervals (weekly or 30‑, 60‑, 90‑day) outperform daily or very short rebalancing.

Contradicted by

  • Causal Inference on Investment Constraints and Non-stationarity in Dynamic Portfolio Optimization through Reinforcement Learning (2023)
    On the question: What is the impact of rebalancing frequency on portfolio performance (e.g., Sharpe ratio)?
    Compare the findings

    This paper: Strategies with less frequent rebalancing (e.g., 30‑, 60‑, 90‑day) outperform those with very short rebalancing periods.

    That paper: Unconstrained (no rebalancing frequency limit) yields the highest average Sharpe ratio; performance declines as the rebalancing interval is lengthened (weekly, bi‑weekly, monthly).

    A finds performance declines as rebalancing moves from no limit to weekly/bi‑weekly/monthly, while B finds 30‑, 60‑, 90‑day intervals outperform very short periods.

  • Multi periods mean-DCVaR optimization: a Recursive Neural Network resolution (2026)
    On the question: How does the frequency or number of portfolio rebalancing dates affect performance?
    Compare the findings

    This paper: Strategies with less frequent rebalancing (e.g., 30‑, 60‑, 90‑day) outperform those with very short rebalancing periods.

    That paper: Performance improves as the number of rebalancing dates N increases.

    A finds less frequent (30‑, 60‑, 90‑day) rebalancing outperforms very short periods, while B reports performance improves as the number of rebalancing dates N increases.

  • Value-at-Risk-Based Portfolio Insurance: Performance Evaluation and Benchmarking Against CPPI in a Markov-Modulated Regime-Switching Market (2023)
    On the question: What is the optimal rebalancing frequency for portfolio insurance/optimization strategies?
    Compare the findings

    This paper: Strategies with less frequent rebalancing (e.g., 30‑, 60‑, 90‑day) outperform those with very short rebalancing periods.

    That paper: Daily rebalancing yields the best Omega and Kappa measures across confidence levels.

    A finds daily rebalancing gives best Omega/Kappa; B finds 30‑, 60‑, 90‑day rebalancing outperforms very short periods.

Cited by 1 paper in the archive (1 by independent authors)

Code

https://github.com/Maciej-13/vcov_forecast — the authors' own repository

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

Cited References (46)

  1. Artificial intelligence in portfolio formation and forecast: Using different variance-covariance matrices (2021) — link
  2. Take Bitcoin into your portfolio: a novel ensemble portfolio optimization framework for broad commodity assets (2021) — link
  3. The Sharpe Ratio (2021) — link
  4. PyPortfolioOpt: portfolio optimization in Python (2021) — link
  5. Covariance matrix forecasting using support vector regression (2021) — link
  6. Investigating the relationship between volatilities of cryptocurrencies and other financial assets (2021) — link
  7. A deep Q-learning portfolio management framework for the cryptocurrency market (2020) — link
  8. Deep Learning for Portfolio Optimization (2020) — link
and 38 more
  1. Should stock investors include cryptocurrencies in their portfolios after all? Evidence from a conditional diversification benefits measure (2020) — link
  2. Deep learning in finance and banking: A literature review and classification (2020) — link
  3. Comparing high-dimensional conditional covariance matrices: Implications for portfolio selection (2020) — link
  4. Financial Time Series Forecasting with Deep Learning : A Systematic Literature Review: 2005-2019 (2019) — link
  5. High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes (2019) — link
  6. Cholesky–ANN models for predicting multivariate realized volatility (2019) — link
  7. GluonTS: Probabilistic Time Series Models in Python (2019) — link
  8. Cryptocurrency-portfolios in a mean-variance framework (2019) — link
  9. On mutual funds-of-ETFs asset allocation with rebalancing: sample covariance versus EWMA and GARCH (2018) — link
  10. Portfolio selection problems with Markowitz’s mean–variance framework: a review of literature (2018) — link
  11. DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks (2017) — link
  12. Can we Invest on the Basis of Equity Risk Premia and Risk Factors from Multi-Factor Models? (2016) — link
  13. A Test of Covariance-Matrix Forecasting Methods (2015) — link
  14. Construction, management, and performance of sparse Markowitz portfolios (2014) — link
  15. Constrained Mean-Variance Portfolio Optimization with Alternative Return Estimation (2013) — link
  16. Modelling and Forecasting Multivariate Realized Volatility (2011) — link
  17. Dominating estimators for minimum-variance portfolios (2010) — link
  18. Mean-Semivariance Optimization: A Heuristic Approach (2007) — link
  19. Improved estimation of the covariance matrix of stock returns with an application to portfolio selection (2003) — link
  20. MODELING AND FORECASTING REALIZED VOLATILITY (2003) — link
  21. Field Guide to Dynamical Recurrent Networks (2001) — link
  22. Python Reference Manual (2000)
  23. Long Short-Term Memory (1997) — link
  24. Multivariate Simultaneous Generalized ARCH (1995) — link
  25. Estimation for Markowitz Efficient Portfolios (1980) — link
  26. Portfolio Selection (1971) — link
  27. Portfolio Selection: Efficient Diversification of Investments (1959) — link
  28. The M4 Competition: 100,000 time series and 61 forecasting methods (2020) — link
  29. GluonTS: Probabilistic and Neural Time Series Modeling in Python (2020)
  30. Hierarchical Multi-Scale Gaussian Transformer for Stock Movement Prediction (2020)
  31. TensorFlow: Large-scale machine learning on heterogeneous systems (2015)
  32. BEKK Element-by-Element Estimation of a Volatility Matrix. A Portfolio Simulation (2014) — link
  33. Shrunk the Sample Covariance Matrix (2003)
  34. Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models (2002)
  35. Morgan Guaranty Trust Company of New York and Others v. Republic of Palau (1992) — link
  36. A Simplified Model for Portfolio Analysis (1963) — link
  37. International Journal of Uncertainty, Fuzziness and Knowledge-based Systems F the Vanishing Gradient Problem during Learning Recurrent Neural Nets and Problem Solutions
  38. A review of probabilistic forecasting and prediction with machine learning

0 of 46 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.