A dynamic conditional approach to forecasting portfolio weights
Introduction
A key aspect of active portfolio management revolves around the projection of the weights that optimize portfolio holdings with respect to some representative measure of the investor’s preferences. Since Markowitz (1952), such optimal portfolio weights forecasts are generally derived from the forecasts of conditional moments of asset returns. More recently, the availability of realized measures from high-frequency data has allowed for model-based and model-free forecasting of conditional variance–covariance (var-cov) matrices and, by successive manipulation thereof, of optimal portfolio weights: see, e.g., Aït-Sahalia et al. (2010), Christensen et al. (2010), Barndorff-Nielsen et al. (2011), Zhang (2011), and Bibinger et al. (2014), among others.
Model-based approaches are inspired by the logic behind multivariate GARCH (MGARCH) models.1 Examples of these approaches are the fractionally integrated processes of Chiriac and Voev (2011), the vector autoregressions of Callot et al. (2017), and the specifications based on the Wishart distribution of Gourieroux et al. (2009), Golosnoy et al. (2012), Noureldin et al. (2012), and Jin and Maheu (2010), among others. Within this framework it is not uncommon to separately model conditional variances2 and correlation matrices to achieve a good balance between parameter parsimony and richness in the description of the second-order dynamics.
On the other hand, model-free approaches, also referred to as nonparametric, impose driftless random-walk dynamics on the conditional second moments, and thus eliminate the parameter estimation problem altogether. However, for large cross-sectional dimensions, the lag-1 realized var-cov matrices may result in extreme portfolio weights, poor portfolio performance out-of-sample (OOS), and even positive-semidefiniteness (psd). To mitigate this problem, various shrinkage approaches are available: the most direct is to impose constraints on the portfolio weights3 as in Jagannathan and Ma (2003), El Karoui (2010), Fan et al. (2012), and Gandy and Veraart (2013). Shrinkage of the realized var-cov matrix has been proposed by Fan et al. (2008), Fan et al. (2011), Tao et al. (2011), Ledoit and Wolf (2012), Tao et al. (2013), Fan et al. (2016), and Aït-Sahalia and Xiu (2017), to name a few. The ideas behind these approaches may also be traced back to the MGARCH literature and consist of imposing a factor structure on the returns and a sparse error var-cov matrix with blocks defined by some characteristics of the assets, such as sector, industry, etc.
In this paper, we introduce dynamic conditional weights (), an approach inspired by the autoregressive representation of the portfolio-variance optimization problem, originally expressed with time-varying coefficients and simplified to a vector ARMA process for the realized weights with time-independent parameters. The resulting specification provides conditional portfolio weights forecasts from a linear function of past conditional and realized (and, hence, observable) terms. With respect to model-based approaches, which model the elements of the conditional var-cov matrix, requires modeling of only portfolio weights. Furthermore, while the dynamic specifications of model-based approaches must enforce positive-definiteness on the ensuing predictions and forecasts, the constraint associated to the is that the resulting portfolio weights add to unity (a condition imposed with a simple rescaling when needed). With respect to model-free approaches, does not require the imposition of particular var-cov matrix structures, nor discretionary choices about the level of shrinkage.
Modeling weights directly has a twofold motivation. First, as mentioned, optimal weights are a centerpiece in portfolio management, so for any utility functional representing investors’ preferences, the framework may be applied to the corresponding time series of realized optimal portfolio weights, bypassing the need to specify var-cov dynamics.4 In this respect, the estimation of (especially for the equation-by-equation estimation with least squares, in the diagonal case) is as simple as the standard least-squares estimation of empirical asset pricing models in the financial literature. Second, within a more financial econometrics perspective, provides a parsimonious and challenging benchmark when forecasting conditional var-cov matrices is involved, given that the performance of different specifications is customarily evaluated in terms of the portfolio allocations via the ensuing optimal weights.
Focusing on the minimum-variance allocation, empirical results on a panel of 28 DJ-30 assets show that outperforms model-based and model-free approaches in terms of out-of-sample portfolio variance and turnover (De Miguel et al., 2009). Since transaction costs can significantly alter the outlook in the performance of the approaches, we introduce the break-even transaction cost as a more comprehensive measure of forecasting performance, to confirm the goodness of the allocation in terms of minimal portfolio variance and turnover. Since research on model-based and model-free approaches has proceeded on somewhat parallel tracks,5 a further contribution of this paper is a comparison across approaches, assessing the quality of the respective forecasts and portfolio allocations.
The paper is organized as follows. Section 2 introduces the optimal portfolio allocation problem. Section 3 introduces direct modeling of the portfolio weights. Model-based and model-free approaches are discussed in Section 4. Data, measures of performance, and results are presented in Section 5. Section 6 is devoted to the evolution of empirical forecasts, when aggregated by sector and averaged by month. Section 7 concludes.
Section snippets
Minimum variance portfolio
Following Fan et al. (2008), Fan et al. (2012), Behr et al. (2013), Fan et al. (2016), and Aït-Sahalia and Xiu (2017), among others, we focus on minimizing portfolio variance, which allows for a clean evaluation of the contribution of modeling and forecasting second moments to the optimal allocation. Furthermore, the minimum-variance portfolio has often been found to perform equally well as, if not better than, the mean–variance portfolio, even when measured in terms of Sharpe ratios: see
Dynamic conditional weights modeling
Our approach derives from a data transformation in which we use the realized covariance matrix to introduce the daily time series of realized optimal portfolio weights: Such weights are observable at time (given the observability of the realized ) and minimize the portfolio realized variance . The time series profile of may be graphically appraised for a few tickers (Apple, Boeing, Johnson and Johnson, and Merck) in Fig. 1: they display different ranges
Projecting covariances
To determine optimal weights forecasts, the traditional venue is to conditionally project the var-cov matrix and then reconstruct the weights according to Eq. (1). To that end, let be a realized measure of the var-cov matrix of assets at time , and let be its conditional expectation in . We can group the approaches into two categories: model-based and model-free approaches. Either approach, however, requires considering how elements dynamically evolve (while
Empirical application
The data used for portfolio selection pertain to of the constituents of the Dow Jones 30 Index. The sample has 11 years of high-frequency daily observations from 01/03/2005 to 12/31/2015 for a total of 2768 days. Two series, with tickers TRV and V, are not included in the study because they are not available for the full sample period.12 The tickers of the 28 included stocks are AAPL, AXP,
Evolution of optimal portfolio weights
In order to complete the empirical analysis, we address the substantive issue of how the one-step-ahead forecasts of optimal weights behave in the period considered, grouping the stocks by sector and considering how the portfolio composition has changed through the events affecting the economy in the six years considered as OOS periods. We take the forecasts of the with and we organize one-step-ahead results for individual stocks by taking their absolute value and rescaling them to
Conclusions
In this paper, we framed the issue of how to model a portfolio variance minimization problem by focusing on realized weights derived from realized var-cov matrices. We started from the autoregressive representation of such an optimization and showed that, with a simplification, represents the dynamic structure of realized weights as a VARMA model whose estimation and specification search follow standard procedures.
The main focus was on evaluating forecasting performance with respect to
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (57)
- et al.
Using principal component analysis to estimate a high dimensional factor model with high-frequency data
Journal of Econometrics
(2017) - et al.
A reduced form framework for modeling volatility of speculative prices based on realized variation measures
Journal of Econometrics
(2011) - et al.
Multivariate realised kernels: Consistent positive semi-definite estimators of the covariation of equity prices with noise and non-synchronous trading
Journal of Econometrics
(2011) - et al.
On portfolio optimization: Imposing the right contraints
Journal of Banking & Finance
(2013) - et al.
A generalized dynamic conditional correlation model for portfolio risk evaluation
Mathematics and Computers in Simulation
(2009) - et al.
Modeling and forecasting (un)reliable realized covariances for more reliable financial decisions
Journal of Econometrics
(2018) - et al.
Financial econometric analysis at ultra–high frequency: Data handling concerns
Computational Statistics & Data Analysis
(2006) - et al.
Pre-averaging estimators of the ex-post covariance matrix in noisy diffusion models with non-synchronous data
Journal of Econometrics
(2010) - et al.
High dimensional covariance matrix estimation using a factor model
Journal of Econometrics
(2008) - et al.
The conditional autoregressive Wishart model for multivariate stock market volatility
Journal of Econometrics
(2012)
The Wishart autoregressive process of multivariate stochastic volatility
Journal of Econometrics
Sequential conditional correlations: Inference and evaluation
Journal of Econometrics
Comparing asset pricing models: An investment perspective
The Journal of Financial Economics
Estimating covariation: Epps effect and microstructure noise
Journal of Econometrics
High-frequency covariance estimates with noisy and asynchronous financial data
Journal of the American Statistical Association
Multivariate orthogonal factor GARCH
Volatility and correlation forecasting
Multivariate GARCH models: a survey
Journal of Applied Econometrics
Estimating the quadratic covariation matrix from noisy observations: Local method of moments and efficiency
The Annals of Statistics
Flexible dynamic conditional correlation multivariate GARCH models for asset allocation
Applied Financial Economics Letters
Multiplicative error models
Modeling and forecasting large realized covariance matrices and portfolio choice
Journal of Applied Econometrics
Modelling and forecasting multivariate realized volatility
Journal of Applied Econometrics
Semiparametric vector MEM
Journal of Applied Econometrics
A simple approximate long-memory model of realized volatility
Journal of Financial Economic
HAR modeling for realized volatility forecasting
Optimal Versus Naïve Diversification: How Inefficient is the 1/N Strategy?
Review of Financial Studies
Parameter uncertainty in multiperiod portfolio optimization with transaction costs
Journal of Financial and Quantitative Analysis
Cited by (1)
Multivariate analysis of cryptocurrencies
2021, Econometrics