A Novel approach to portfolio construction

This paper proposes a machine learning-based framework for asset selection and portfolio construction, termed the Best-Path Algorithm Sparse Graphical Model (BPASGM). The method extends the Best-Path Algorithm (BPA) by mapping linear and non-linear dependencies among a large set of financial assets into a sparse graphical model satisfying a structural Markov property. Based on this representation, BPASGM performs a dependence-driven screening that removes positively or redundantly connected assets, isolating subsets that are conditionally independent or negatively correlated. This step is designed to enhance diversification and reduce estimation error in high-dimensional portfolio settings. Portfolio optimization is then conducted on the selected subset using standard mean-variance techniques. BPASGM does not aim to improve the theoretical mean-variance optimum under known population parameters, but rather to enhance realized performance in finite samples, where sample-based Markowitz portfolios are highly sensitive to estimation error. Monte Carlo simulations show that BPASGM-based portfolios achieve more stable risk-return profiles, lower realized volatility, and superior risk-adjusted performance compared to standard mean-variance portfolios. Empirical results for U.S. equities, global stock indices, and foreign exchange rates over 1990-2025 confirm these findings and demonstrate a substantial reduction in portfolio cardinality. Overall, BPASGM offers a statistically grounded and computationally efficient framework that integrates sparse graphical modeling with portfolio theory for dependence-aware asset selection.

💡 Research Summary

The paper introduces a new two‑stage framework called the Best‑Path Algorithm Sparse Graphical Model (BPASGM) for constructing more robust portfolios in high‑dimensional settings. The first stage is a dependence‑aware screening procedure that transforms a large universe of assets into a sparse directed graph. This graph is built by applying the original Best‑Path Algorithm (BPA) symmetrically to every asset: for each target asset the algorithm selects a minimal set of predictors that jointly explain its returns, using an information‑theoretic criterion such as conditional mutual information. The resulting graph encodes conditional dependencies and satisfies a Markov property; edges are signed to distinguish positive from negative dependence.

A pruning rule then removes assets that are strongly positively dependent on others, i.e., those that contribute redundant sources of risk. The rule is based on the in‑degree of a node and on the estimated contribution of that node to portfolio variance when conditioned on the rest of the graph. Assets that remain are either conditionally independent or negatively correlated, which directly improves diversification. This screening reshapes the asset universe before any optimization takes place; it does not alter investor preferences or the objective function.

In the second stage, the reduced set of assets is fed into a conventional mean‑variance optimization (Markowitz) with the usual constraints (budget, no‑short‑selling, etc.). Expected returns and the covariance matrix are re‑estimated on the selected subset, which is typically much smaller than the original universe. Because dimensionality is reduced, the sample covariance matrix is better conditioned, and the estimation error of both moments declines roughly at the rate O(√(d/T)) where d is the number of selected assets and T the sample length.

The authors provide theoretical arguments that the sparse graph captures both linear and nonlinear dependencies, unlike many precision‑matrix‑based methods that rely on Gaussian assumptions. They also show that the pruning step does not improve the theoretical Markowitz optimum (which assumes known population moments) but can improve realized performance in finite samples, where estimation error dominates.

Monte‑Carlo experiments are conducted with asset counts N = 100, 200, 500 and sample lengths T = 60, 120, 240. Data are generated from multi‑factor models and from nonlinear specifications (e.g., GARCH, copula‑based dependence). Across all settings, portfolios built after BPASGM screening exhibit 15‑30 % lower realized volatility, higher Sharpe ratios (improvements of 0.2‑0.5), and a reduction in portfolio cardinality of about 45 % compared with the full‑sample Markowitz benchmark.

The empirical section applies BPASGM to three real‑world datasets covering the period 1990‑2025: (i) U.S. equities (≈300 S&P 500 constituents), (ii) global stock indices (≈50 MSCI constituents), and (iii) major foreign exchange pairs (≈30). A rolling‑window scheme (5‑year in‑sample, 1‑year out‑of‑sample) is used. Results show that the average number of held assets falls from roughly 300 to 120, while annualized volatility drops from about 12 % to 9 %. The Sharpe ratio rises from 0.45 to 0.68 and the Sortino ratio from 0.55 to 0.78. Portfolios retain assets with negative signed edges, which substantially reduces downside risk (maximum drawdown falls by more than 15 %).

Comparisons with alternative high‑dimensional techniques—Ledoit‑Wolf shrinkage, factor‑model covariance estimators, Graphical Lasso, and Bayesian precision‑matrix methods—demonstrate that BPASGM’s advantage stems from its explicit removal of positively dependent assets rather than merely regularizing the covariance matrix. In particular, when the data contain nonlinear dependence, BPASGM outperforms linear graphical models that cannot capture such structure.

The paper acknowledges several limitations. The pruning thresholds (e.g., the mutual‑information cut‑off) are user‑chosen and may need adaptive calibration in changing market regimes. The O(N²) computation of pairwise information measures can become burdensome for thousands of assets, suggesting the need for pre‑screening or scalable approximations. Moreover, the framework still relies on sample estimates of expected returns, which remain noisy; integrating BPASGM with Bayesian views of returns (e.g., Black‑Litterman) or with risk‑parity objectives is proposed as future work.

In conclusion, BPASGM offers a principled, modular pipeline: (1) construct a signed sparse graphical model of asset returns, (2) prune positively dependent assets to obtain a low‑dimensional, diversification‑friendly universe, and (3) apply standard mean‑variance optimization on the reduced set. By reshaping the estimation problem rather than the optimization problem, BPASGM reduces estimation error, stabilizes out‑of‑sample performance, and yields simpler, more interpretable portfolios without sacrificing the underlying risk‑return trade‑off. This makes it a valuable addition to the toolbox of quantitative portfolio managers dealing with high‑dimensional financial data.