A Deep Learning Framework for Medium-Term Covariance Forecasting in Multi-Asset Portfolios

Technical Summary: A Deep Learning Framework for Medium-Term Covariance Forecasting in Multi-Asset Portfolios

Problem Statement
Accurate covariance matrix forecasting is fundamental to portfolio allocation, risk management, and asset pricing. While existing literature offers robust solutions for short-term horizons (often utilizing high-frequency data and GARCH-based models) and long-term horizons (relying on factor models or naive shrinkage), a significant gap remains in medium-term forecasting (multi-week to multi-month windows). This gap is particularly critical for institutional investors who rebalance portfolios over such periods. Furthermore, classical econometric models often struggle with the "curse of dimensionality" in large cross-sections and may fail to capture slow-moving structural changes or regime shifts (e.g., equity-bond correlation shifts) that define medium-term dynamics. Additionally, there is limited empirical evidence regarding whether the choice between raw returns and excess returns systematically impacts covariance forecasts over these horizons.

Methodology
The authors propose a novel Deep Learning (DL) framework, termed CAB, designed to capture complex spatio-temporal dependencies in multi-asset portfolios. The architecture integrates three distinct components:

1. 3D Convolutional Neural Networks (CNN): The model treats rolling covariance matrices as $(N \times N)$ grids stacked over time. A 3D CNN layer applies filters across temporal and spatial dimensions to identify local interaction patterns among subsets of assets, moving beyond linear assumptions.
2. Bidirectional Long Short-Term Memory (BiLSTM): Following the CNN, flattened sequences are processed by a BiLSTM. Unlike standard LSTMs that only process forward sequences, the BiLSTM captures context from both past and future data points within the lookback window, enhancing the model's ability to understand temporal dependencies.
3. Multi-Head Self-Attention: To refine temporal dependencies, a multi-head attention mechanism is applied to the BiLSTM outputs. This allows the model to dynamically weigh the importance of different time steps, focusing on the most relevant patterns for prediction.

Data Processing and Constraints:

• Input: The model uses sequences of scaled, rolling covariance matrices derived from daily returns (both raw and excess returns are tested).
• Output Constraints: To ensure the predicted matrix is a valid covariance matrix, the output undergoes specific post-processing:
  • Symmetrization: The output matrix is averaged with its transpose.
  • Positive Semi-Definiteness (PSD): An eigenvalue decomposition is performed, and any negative eigenvalues are set to zero before reconstruction.
  • Shrinkage: A final linear shrinkage step blends the DL output with the most recent historical covariance matrix to incorporate valuable information from recent market conditions.

Experimental Setup

• Data: Daily data from 14 major ETFs (covering equities, real estate, and fixed income) from January 2017 to December 2023. A larger robustness check utilized 260 ETFs.
• Benchmarks: The CAB model is compared against 10 benchmarks categorized as:
  • Naive: Lagged realized covariance (NA), Full sample (NAF), Exponential Weighted Moving Average (EWMA).
  • De-noising: Principal Component Analysis (PCA), Random Matrix Theory (RMT), Ledoit-Wolf shrinkage (LW, LW F).
  • GARCH-based: Constant Conditional Correlation (CCC), Dynamic Conditional Correlation (DCC), Nonlinear Shrinkage DCC (DCCNL).
• Evaluation Metrics: Euclidean distance ($L_E$) and Frobenius distance ($L_F$) between forecasted and realized covariance matrices.
• Market Regimes: The testing period (2021–2023) is divided into three regimes: Bull-1 (low yields, stable equities), Bear (rising rates, market drop), and Bull-2 (high rates, equity upswing).

Key Results

• Forecast Accuracy: The CAB model consistently outperformed all classical benchmarks across the entire testing period. It reduced Euclidean and Frobenius distance metrics by up to 20% relative to the best-performing classical models (e.g., DCC and EWMA).
• Robustness Across Regimes: CAB demonstrated superior performance in all three market regimes (Bull-1, Bear, Bull-2). While all models suffered decreased accuracy during the turbulent Bear period, CAB maintained the lowest error rates, particularly in the Bull-2 period characterized by dynamic cross-asset shifts.
• Scalability: In the high-dimensional setting (260 ETFs), CAB continued to outperform benchmarks, confirming its ability to handle the curse of dimensionality effectively.
• Return Specification: The model's performance advantage held true regardless of whether raw returns or excess returns were used as input. This suggests the framework's effectiveness does not hinge on the subtraction of the risk-free rate.
• Forecast Horizons: The model excelled in medium-term horizons (e.g., $F=20, 40, 90$ days). While GARCH models performed competitively at very short horizons, they underperformed as the horizon increased. CAB remained the top performer for medium-to-long horizons.

Economic Value in Portfolio Management
The authors evaluated the economic utility of the forecasts using a Global Minimum Variance (GMV) portfolio strategy with no short-selling constraints.

• Risk Reduction: Portfolios constructed using CAB forecasts achieved lower out-of-sample volatility compared to those using other models and the equal-weighted ($1/N$) benchmark.
• Turnover: Crucially, CAB achieved this risk reduction with moderate turnover. Unlike some dynamic models that required frequent rebalancing (high turnover) to achieve marginal variance reductions, CAB offered a stable balance between responsiveness to market changes and transaction cost efficiency.
• Regime Adaptation: In the Bear regime, CAB ranked near the top in daily and weekly horizons, balancing lower turnover against volatility.

Significance and Claims
The paper claims that advanced deep learning architectures, specifically those combining 3D convolutions, bidirectional LSTMs, and attention mechanisms, offer a practical solution for the under-explored challenge of medium-term covariance forecasting. The study highlights that:

1. Hybrid Approaches: Integrating data-driven DL techniques with traditional econometric concepts (like shrinkage) yields substantial gains in predictive accuracy, especially during market stress.
2. Practical Applicability: The framework provides tangible economic benefits for institutional investors by reducing portfolio volatility without incurring excessive trading costs.
3. Robustness: The model's performance is robust to different return specifications (raw vs. excess) and scales effectively to large asset universes.

The authors conclude that while their work bridges the gap between machine learning innovation and financial principles, future research should explore accommodating heavy-tailed distributions for better tail risk capture and explicitly incorporating transaction costs into the optimization process.