Portfolio Construction#
Portfolio construction is the discipline of combining individual trading signals and assets into a coherent portfolio that optimizes risk-adjusted returns while respecting practical constraints. The field has evolved from Markowitz’s foundational mean-variance framework — still the theoretical baseline — through hierarchical risk parity, factor-based allocation, and deep reinforcement learning approaches. A persistent challenge in portfolio construction is that estimated expected returns and covariances are noisy, causing naive optimization to amplify estimation errors and produce unstable, concentrated portfolios.
This collection covers the full portfolio construction toolkit: Markowitz-based optimization with regularization and constraints, hierarchical risk parity (HRP) that avoids matrix inversion, clustering-based approaches for cardinality-constrained portfolios, dynamic Black-Litterman models, and the practical challenges of volatility targeting in multi-asset portfolios. Key themes include the trade-off between model complexity and estimation robustness, the role of correlation in position sizing, and how portfolio construction decisions can matter as much as the underlying signal quality.
Related topics include Risk Management for position sizing and drawdown control, Volatility Modeling for the covariance inputs to portfolio optimization, and Equity Factors for factor-based allocation frameworks.
Building Diversified Portfolios that Outperform Out-of-Sample#
The paper introducing Hierarchical Risk Parity (HRP), a portfolio allocation algorithm that uses hierarchical clustering of the asset correlation matrix to create more stable and diversified portfolios than classical mean-variance optimization. HRP avoids the matrix inversion that amplifies estimation errors in standard approaches, producing portfolios that consistently outperform out-of-sample.
By Marcos López de Prado.
Clustering in Cardinality-Constrained Portfolio Optimization#
This paper combines cardinality constraints with the classical Markowitz mean-variance model, using spectral clustering to group stocks by their return characteristics. The clustering step reduces dimensionality and achieves an optimal balance of risk and return while keeping the portfolio manageable in terms of number of holdings.
Optimal Allocation to Cryptocurrencies in Diversified Portfolios#
This paper applies four quantitative methods for optimal allocation to Bitcoin and Ether within alternative and balanced portfolios. Using roll-forward historical simulations, all four allocation methods produce a persistent positive allocation to Bitcoin and Ether with a median allocation of about 2.7%, emphasizing the diversification benefits of cryptocurrencies as an asset class.
Efficient Portfolio Estimation in Large Risky Asset Universes#
This paper develops a constrained sparse regression approach for large-dimensional portfolio optimization that produces stable, interpretable portfolio weights without requiring full covariance matrix estimation. The method handles universes of hundreds or thousands of assets while maintaining the convexity properties needed for tractable optimization.
Covariance Implied Risk Factors#
This paper develops a heteroskedastic PCA approach that extracts equity risk factors from the covariance structure of returns, allowing factor loadings to vary with market conditions. The method identifies more stable and predictive risk factors than standard PCA, providing better covariance matrix estimates for portfolio construction.
Asset Allocation: From Markowitz to Deep Reinforcement Learning#
This paper benchmarks nine portfolio allocation strategies ranging from classical Markowitz mean-variance optimization to deep reinforcement learning approaches, providing a comprehensive comparison of their risk-adjusted performance across different market environments.
Optimizing Portfolio Performance through Clustering and Sharpe Ratio-Based Optimization#
This paper combines K-means clustering of assets with Sharpe ratio-based optimization at the cluster level, creating a two-stage approach that first groups correlated assets and then optimizes across clusters. The method improves diversification and out-of-sample performance compared to single-stage optimization.
Dynamic Black-Litterman#
This paper extends the Black-Litterman framework to a continuous-time dynamic setting, allowing investor views to evolve and the model to adapt to changing market conditions. The dynamic formulation provides a principled way to incorporate time-varying beliefs about expected returns while maintaining the model’s property of producing well-diversified portfolios.
Dynamic Mean-Variance Portfolio Allocation under Regime-Switching Jump-Diffusions with Absorbing Barriers#
This paper solves the dynamic mean-variance portfolio allocation problem under a regime-switching jump-diffusion model with absorbing barriers representing ruin. The optimal allocation policy depends on the current regime and adapts to regime transitions, providing a rigorous framework for regime-aware portfolio construction.
Why Mean-Variance Optimization Breaks Down#
Mean-Variance Optimization (MVO) is a central framework for portfolio construction, yet practitioners quickly encounter a paradox: the mathematically “optimal” portfolio built from estimated inputs is often unstable, highly leveraged, and disappoints out-of-sample. This is not a minor implementation detail—it is a structural consequence of combining a high-dimensional optimizer with noisy estimates of expected returns and covariances.
This article develops MVO from first principles and explains, in a mathematically explicit way, why raw MVO tends to maximize estimation error. It surveys the spectrum of practical fixes organized around two levers: improving or regularizing the inputs, and constraining or regularizing the optimizer. The unifying theme is that almost every successful fix works by injecting bias in exchange for a large reduction in variance of the resulting portfolio weights.
By VertoxQuant.
Fixing Position Sizing by Accounting for Inter-Asset Correlations#
A practical account of discovering and fixing a crude volatility targeting formula that was causing the portfolio to miss its risk targets. The author found that a simplistic position sizing approach was capping the portfolio too aggressively, resulting in realized volatility at roughly half of the intended target.
The article walks through the correlation fix that resolved the discrepancy, explaining how accounting for inter-asset correlations in the volatility targeting formula allowed the portfolio to properly hit its intended risk profile.
By Pedma.
How Binary Signals Create Inefficient Capital Allocation in Multi-Asset Frameworks#
An exploration of a bug encountered while building scalable trading infrastructure designed to handle multiple asset classes, exchanges, and strategy types from a centralized codebase. When testing a binary signal implementation, the system was producing a very low gross exposure profile with a mean of just 10.9% and a median of 5%.
The article explains how binary signals (on/off) can create inefficient capital allocation when combined with position sizing rules in a multi-asset framework. The author demonstrates how to rethink signal handling to improve exposure efficiency.
By Pedma.
Strategic Drag: Accepting Lower Returns for Better Risk-Adjusted Performance#
Research Article #69 from Trading Research Hub. Explores the concept of “strategic drag” in systematic trading, which refers to the intentional acceptance of reduced raw returns in exchange for improved risk-adjusted performance. The article examines how deliberate constraints on a trading strategy can paradoxically improve its overall utility.
By Pedma.
Using Rolling Correlation to Hit Volatility Targets in Crypto Trend Following#
A practical discussion on incorporating correlation estimates into position sizing for a long/short crypto trend following portfolio. The author was targeting 40% annualised volatility but consistently hitting only 20%, because the sizing formula assumed perfect correlation among positions.
The fix involved adding a rolling correlation estimate to the sizing formula. The author found R² values of 0.25–0.4 for correlation predictability in crypto — consistent with assumptions in Rob Carver’s work — and confirmed that using the rolling correlation estimate allowed the portfolio to consistently hit its volatility target.
By pedma7.
Portfolio Backtest and Trading Techniques#
This post explores why combining multiple independent trading strategies into a capital-aware portfolio is more effective than searching for a single “holy grail” strategy. The author argues that over-optimizing, constant tweaking, and adding complex rules create great backtests that fail in live markets.
The real key is building a team of robust strategies that work well together across different market conditions. By mixing strategies across different assets, timeframes, frequencies, and styles, and ensuring they have low correlation, traders can create a portfolio that compounds returns while dramatically reducing drawdowns, referencing Parrondo’s Paradox where combining “losing” strategies can create a winning approach.
By Build Alpha.
Understanding Covariance Matrices for Systematic Portfolio Allocation#
An educational article explaining the covariance matrix and its role in systematic portfolio allocation. The article breaks down this key quantitative concept in accessible terms, explaining why understanding asset correlations is essential for managing portfolio risk.
The article demonstrates how high correlations between portfolio holdings increase the risk that the entire portfolio moves adversely during market stress. It covers how to construct and interpret a covariance matrix, and how this information can be used to build more diversified portfolios.
By Pedma.
Building a Multi-Asset Trading Model from Scratch#
Part of a series on building a multi-strategy portfolio from scratch, this article covers the practical challenges of managing a multi-asset trading model. The author documents the process of developing and deploying an automated strategy portfolio.
The article addresses real-world implementation challenges including data management, signal aggregation across assets, and the operational complexity that grows as more instruments are added to a systematic trading portfolio.
By Pedma.
Volatility Drag, the Kelly Criterion, and Portfolio Position Sizing#
A post examining the relationship between volatility drag, the Kelly criterion, and optimal portfolio construction. The key insight is that the highest expected compound growth rate is achieved when portfolio volatility equals the Sharpe ratio — which corresponds to Kelly criterion sizing.
The post also addresses how variance drag diminishes for individual positions within a diversified portfolio. Because the risk penalty to idiosyncratic risk scales with the square of the portfolio weight, a 10% allocation reduces the importance of a component’s idiosyncratic variance by a factor of 100.
By Ptuomov.
Robust Optimization of Strategic and Tactical Asset Allocation for Multi-Asset Portfolios#
Artur Sepp, Ivan Ossa, and Mika Kastenholz present the ROSAA framework — a unified pipeline for strategic and tactical asset allocation of multi-asset portfolios. The paper addresses the practical challenges institutional investors face when constructing portfolios across heterogeneous asset classes with varying liquidity, rebalancing frequencies, and incomplete return histories. The framework integrates factor model covariance estimation, risk-budgeted strategic asset allocation (SAA), alpha signal computation, and tactical asset allocation (TAA) with tracking error constraints into a single coherent pipeline.
Our summary: The key methodological contribution is the Hierarchical Clustering Group LASSO (HCGL) factor covariance estimator, which builds sparse, structured covariance matrices for heterogeneous multi-asset universes by clustering assets and applying group LASSO regularization to factor loadings. For SAA, the framework uses constrained risk budgeting with group allocation limits, avoiding the instability of traditional mean-variance optimization. For TAA, an alpha-over-tracking-error optimization tilts the portfolio toward tactical signals while maintaining a risk budget relative to the strategic benchmark. The framework handles real-world complications including NaN-aware rolling backtesting for assets with missing data, drift-aware turnover accounting that measures trades against drifted holdings rather than stale baselines, and mixed-frequency rebalancing for illiquid positions.
Code: The optimalportfolios Python package on GitHub (ArturSepp/OptimalPortfolios) is the reference implementation. It includes solvers for minimum variance, constrained risk budgeting, maximum diversification, maximum Sharpe, alpha-over-tracking-error, and CARA utility under Gaussian mixture, with end-to-end backtesting examples.
Published in the Journal of Portfolio Management, 2026, Volume 52, Issue 4, Pages 86-120.
Mentioned by Artur Sepp on LinkedIn.
From Classical Optimization to Bayesian Integration: A Comprehensive Analysis of Systematic Portfolio Management#
This paper compares five portfolio construction approaches — mean-variance optimization, constrained optimization, Fama-French five-factor regression modeling, Monte Carlo simulation, and the Black-Litterman model — using ten U.S. stocks (TSLA, WMT, BAC, GS, LLY, MRK, GOOG, META, AAPL, XOM) over September 2023 to December 2025. The study examines how constraints, risk factors, simulated approximations, and market views affect portfolio allocation, performance, and stability.
The paper provides a structured walkthrough of each method’s mechanics and trade-offs. Unconstrained mean-variance optimization produces concentrated “corner portfolios” dominated by AAPL (29%), XOM (28%), and GOOG (22%). Adding a 15% maximum weight constraint redistributes allocations across more assets and, surprisingly, improves out-of-sample performance. Fama-French five-factor regressions (both OLS and robust) characterize the constrained portfolio as defensive large-cap value with a profitability tilt (R² = 0.715). The Monte Carlo section demonstrates that simulation can approximate optimization results but struggles with box constraints due to the curse of dimensionality. The Black-Litterman model, incorporating investor views and market equilibrium, produces the most diversified and stable allocations by anchoring weights to market capitalization and adjusting only where views are strong.
The study uses daily adjusted closing prices from Yahoo Finance. Estimation period is September 2023 to September 2025, with out-of-sample testing from October to December 2025. The optimization is implemented using the cvxpy library. No code repository is provided.
Key out-of-sample metrics (October–December 2025): the constrained GMV portfolio achieved a 16.77% cumulative return, Sharpe ratio of 5.36, and maximum drawdown of -3.18%, outperforming the unconstrained GMV portfolio (15.85% return, 5.05 Sharpe, -2.76% max drawdown). The constrained portfolio’s market beta was 0.81, with statistically significant loadings on value (HML = 0.20) and profitability (RMW = 0.12) factors, and no significant alpha. The estimated market risk aversion coefficient for the Black-Litterman model was δ = 6.79. Note that the reported Sharpe ratios are inflated by the extremely short 3-month out-of-sample window coinciding with a strong bull market; realistic long-term Sharpe ratios for passive large-cap portfolios are typically 0.5–1.0. The paper reads as a graduate-level pedagogical overview of standard portfolio theory rather than novel research, which is consistent with the authors’ framing as a beginner review.
Mentioned by Piotr Pomorski in this discussion.
Building Meta-Strategies with Quantpedia API#
Meta-strategies apply portfolio construction rules to universes of trading strategies rather than individual securities. Instead of allocating capital across stocks or ETFs, a meta-strategy allocates across strategy return streams, treating each strategy’s equity curve as a portfolio component. This post demonstrates the approach using the Quantpedia API to filter, retrieve, and allocate across a universe of 118 strategies (filtered for simple implementation, ETF-based instruments, and available QuantConnect source code).
The case study applies cross-sectional momentum to strategies: each month, all strategies are ranked by trailing 6-month cumulative returns, the top 10% are selected with equal weights, and the portfolio rebalances. The momentum meta-strategy is compared against equal-weight allocation and naive risk parity across the same universe. The post notes the core trade-off: broad equal-weight portfolios offer diversification but no responsiveness to performance shifts, while momentum overlays provide dynamism but concentrate exposure and increase drawdown risk during regime changes.
The underlying research question — whether published anomalies exhibit performance persistence — draws on a body of work collected in Momentum: Huang & Huang show recursive selection of the best-performing anomaly outperforms the market after costs; Wang, Yan & Zheng document time-series and cross-sectional momentum in anomaly returns; Ehsani & Linnainmaa demonstrate that stock momentum emanates from factor momentum; Geczy & Samonov confirm momentum across asset classes over 215 years; Zaremba & Shemer find momentum in factor premia across 24 countries (though largely explained by stock-level momentum); and Blitz documents the decade-long underperformance of several factors in 2010–2019, a cautionary case for factor-rotation strategies.
By Quantpedia.
A Levered ETF Anomaly Explained#
Between January 2022 and December 2023, the S&P 500 Index rose 0.076%, yet the ProShares Ultra S&P500 (SSO, 2x leverage) lost 11.09% and the ProShares UltraPro S&P500 (UPRO, 3x leverage) lost 28.24%. This paper by Bianchi and Goldberg decomposes this counterintuitive underperformance into two components: volatility drag from compounding and a timing effect from leverage deviations.
The authors derive a closed-form approximation relating geometric return G to arithmetic return E and volatility V: G ~ (1+E)exp(-V^2/2) - 1. With annualized index arithmetic return of 1.92% and volatility of 19.40%, this formula alone predicts negative geometric returns for levered products. However, this “ideal constant leverage” estimate only explains roughly two-thirds of the actual shortfall. The remaining third arises from the covariance between the ETFs’ daily return ratios (realized leverage) and the index return. This covariance is consistently negative for SSO (-2.242) and UPRO (-5.643), meaning leverage tends to decrease when returns are positive and increase when returns are negative – a detrimental timing drag. The authors emphasize that levered ETFs are path-dependent volatility trades, not simple long-horizon leveraged exposure.
The study uses CRSP daily ETF return data (price returns, excluding dividends) for the S&P 500 Index, VOO, SSO, and UPRO over 501 trading days (January 3, 2022 to December 29, 2023). No code is provided.
Key metrics over the 501-day sample: S&P 500 cumulative return +0.076%, annualized volatility 19.40%, risk-adjusted return 0.099; SSO cumulative return -11.09%, annualized volatility 38.77%, risk-adjusted return 0.043; UPRO cumulative return -28.24%, annualized volatility 58.01%, risk-adjusted return 0.005. The ideal constant leverage model estimated SSO geometric return at -3.625% and UPRO at -10.580%, while the empirical estimate incorporating the covariance term yielded -5.694% and -15.226%, closely matching actual geometric returns of -5.707% and -15.288% respectively.
Mentioned by Ivan Blanco on LinkedIn. He emphasizes that “these are path-dependent volatility trades, not long-horizon leveraged exposure” and cautions against using leveraged ETFs as a simple way to increase beta exposure.
Yau’s Affine Normal Descent: Algorithmic Framework and Convergence Analysis#
This paper by Yi-Shuai Niu, Artan Sheshmani, and Fields Medal winner Shing-Tung Yau introduces Yau’s Affine Normal Descent (YAND), a geometric optimization framework in which search directions are defined by the equi-affine normal of level-set hypersurfaces. Unlike classical methods such as gradient descent or Newton’s method that rely on Euclidean or locally quadratic models, YAND derives its search directions intrinsically from the geometry of the objective function’s level sets. The resulting directions are invariant under volume-preserving affine transformations and naturally adapt to anisotropic curvature, making the method inherently robust to ill-conditioned coordinate scalings.
The paper establishes that for strictly convex quadratic objectives, affine-normal directions are collinear with Newton directions, implying one-step convergence under exact line search. For general smooth (possibly nonconvex) objectives, the authors characterize when affine-normal directions yield strict descent and develop a line-search-based YAND algorithm. They prove global convergence under standard smoothness assumptions, linear convergence under strong convexity and Polyak-Lojasiewicz conditions, and quadratic local convergence near nondegenerate minimizers. A key theoretical contribution is showing that YAND’s search directions are completely unaffected by arbitrarily ill-conditioned affine transformations of the variables – a property that gradient descent and quasi-Newton methods lack.
The method is directly relevant to portfolio optimization where objective functions (risk measures incorporating skewness, kurtosis, and tail risks) are highly anisotropic and where the number of assets creates severe scaling and conditioning challenges. YAND uses tensor-free computation of the affine normal direction (requiring only first and second derivatives, not full Hessian inversion) and determines optimal step sizes via polynomial line search rather than trial-and-error, addressing the computational scalability barrier in large-scale multi-asset allocation.
Numerical experiments are conducted in MATLAB on representative 2D problems. On well-conditioned quadratics, YAND-Exact converges in 1 iteration (matching Newton’s method). Under affine scaling with condition numbers up to 10^4 (Hessian condition number up to 10^8), YAND-Exact still converges in 1 step while gradient descent with fixed step size hits the 200-iteration cap for condition numbers >= 10. On the Rosenbrock function, YAND converges in 12-20 iterations depending on line-search variant, while gradient descent requires 200+ iterations. On smooth nonconvex problems including saddle regions and multi-well landscapes, YAND follows the curvature of level sets and converges stably. The implementation uses analytic derivatives via automatic differentiation with no external optimization libraries. No code is publicly available.
Mentioned by Luca Cai in this discussion. Cai highlights that YAND addresses two major problems in portfolio optimization: risk characterization beyond normal distributions (volatility, skewness, kurtosis, tail risks) and computational scalability when the number of assets is large, noting that the method shifts optimization to affine-normal geometry with tensor-free computation and polynomial step-size determination.
Bayesian Methods in Finance#
This handbook chapter by Eric Jacquier and Nicholas Polson (“Bayesian Methods in Finance,” in The Oxford Handbook of Bayesian Econometrics, 2011) is the anchor survey for Bayesian portfolio construction and the single best map of how Bayesian updating threads through asset allocation. It lays out the canonical lineage: (a) Klein & Bawa (1976) — optimal portfolio choice under parameter uncertainty requires optimizing expected utility over the Bayesian predictive density, with plug-in point estimates being provably suboptimal; (b) Jorion (1985/1986) — empirical-Bayes (Bayes-Stein) shrinkage of estimated means toward the global minimum-variance portfolio mean; and (c) Black & Litterman (1991) — building the prior from market-equilibrium-implied returns and combining it with investor views, a scheme the chapter notes is curiously absent a formal likelihood (a gap Zhou 2008/2009 later fills).
Our summary: read this first if you want to understand why raw sample means produce unstable, extreme mean-variance weights and how a Bayesian treatment fixes it — by carrying estimation uncertainty into the optimization rather than optimizing a fragile point estimate. It is a theoretical chapter, but it is the connective tissue for the four classic papers indexed below (Klein-Bawa, Jorion, Black-Litterman, Zhou), and it frames the entire predictive-density approach that underlies Bayesian position sizing. High-value full read for anyone building an allocation layer over multiple signals or strategies.
Data and code: a survey/handbook chapter; no data or code. Freely hosted as a PDF on the author’s site.
Key metrics: not applicable — this is a methodological survey, valuable for the framework and the literature map rather than any reported performance.
The Effect of Estimation Risk on Optimal Portfolio Choice#
This paper by Roger W. Klein and Vijay S. Bawa (1976, Journal of Financial Economics 3:215-231) is the seminal result on portfolio choice under parameter uncertainty and the theoretical root of Bayesian portfolio construction. Building on Zellner & Chetty (1965), the authors show that when the true return distribution parameters are unknown, the optimal strategy is to compute and then optimize expected utility with respect to the Bayesian predictive density — the parameter-uncertainty-integrated distribution of future returns — rather than plugging sample estimates into the utility as if they were the truth. Doing the latter (“estimation risk” ignored) leads to demonstrably suboptimal portfolios.
Our summary: this is the “why Bayesian at all” paper for allocation. The predictive-density argument is the formal justification for every downstream shrinkage and Bayesian-sizing method — including Bayesian Kelly — because it says uncertainty about your edge should change the allocation, not just be acknowledged. Worth knowing as the foundation even though the modern practitioner usually meets it through later shrinkage estimators (Jorion) or the Jacquier-Polson survey above.
Data and code: a theoretical paper; no public code. Behind the Elsevier/ScienceDirect paywall — we were unable to retrieve the full PDF (browser-based capture and Sci-Hub were both unavailable in this run); the abstract and citation are indexed here for completeness.
Key metrics: not applicable — a decision-theoretic result, not an empirical study.
Bayes-Stein Estimation for Portfolio Analysis#
This paper by Philippe Jorion (1986, Journal of Financial and Quantitative Analysis 21(3):279-292), with its international-diversification companion (Jorion 1985, Journal of Business 58(3):259-278), is the practical workhorse of Bayesian portfolio estimation: it applies James-Stein/empirical-Bayes shrinkage to the vector of expected returns, pulling each asset’s estimated mean toward the mean of the global minimum-variance portfolio. The shrinkage intensity is estimated from the data, so noisier, shorter samples are shrunk harder.
Our summary: this is the most directly usable classic in the Bayesian-allocation set — a few lines of code that materially stabilize mean-variance weights. Because sample means are the dominant source of error in Markowitz optimization, shrinking them toward a sensible target (the min-variance mean) reduces the extreme, error-maximizing positions that plague naive optimizers and produces more stable, better out-of-sample portfolios. Practitioner-relevant; pairs naturally with covariance shrinkage (Ledoit-Wolf) for a fully regularized optimizer.
Data and code: empirical study on international equity indices; no public code. Behind the JSTOR/Cambridge paywall — full PDF not retrievable in this run (browser and Sci-Hub unavailable); citation indexed for completeness.
Key metrics: Bayes-Stein portfolios are shown to dominate raw sample-estimate portfolios out-of-sample, with more stable weights and improved risk-adjusted performance; the paper reports estimation-error and out-of-sample comparisons rather than a single headline Sharpe.
Asset Allocation: Combining Investor Views with Market Equilibrium (Black-Litterman)#
This paper by Fischer Black and Robert Litterman (1991, The Journal of Fixed Income 1(2):7-18) is the canonical Bayesian portfolio model used across the institutional industry. The market-cap-weighted portfolio is reverse-engineered into a set of equilibrium-implied expected returns (by inverting the mean-variance first-order condition), and these serve as the Bayesian prior. The investor’s own views — with explicitly stated confidences — act as the evidence, and the posterior blends prior and views in conjugate (Gaussian) form. The result avoids the wild, corner-solution weights that mean-variance optimization produces when fed raw return forecasts.
Our summary: this is the most influential applied instance of Bayesian updating in finance and the one a quant is most likely to deploy directly. The key idea worth internalizing is that the equilibrium prior anchors the optimizer, so your views only tilt the portfolio in proportion to your confidence — exactly the behaviour you want when blending uncertain alpha signals. The model’s one formal gap, flagged by Jacquier & Polson, is the absence of an explicit likelihood; Zhou (2009) closes it. Essential practitioner reading.
Data and code: a methodological/applied paper (originating as Goldman Sachs research); no public code, but numerous open-source implementations exist. Behind the JFI/Portfolio Management Research paywall — full PDF not retrievable in this run (browser and Sci-Hub unavailable); citation indexed for completeness.
Key metrics: the paper demonstrates well-behaved, intuitively reasonable allocations versus the extreme weights of naive mean-variance optimization rather than reporting a backtested Sharpe/return series.
Beyond Black-Litterman: Letting the Data Speak#
This paper by Guofu Zhou (2009, The Journal of Portfolio Management 36(1):36-45; widely circulated as the 2008 working paper “An Extension of the Black-Litterman Model: Letting the Data Speak”) closes the main theoretical gap in Black-Litterman. The original BL model combines an equilibrium prior with investor views but lacks a formal likelihood — it never lets the historical data update the posterior. Zhou adds a likelihood so the posterior expected returns reflect three sources at once: market equilibrium (prior), subjective views, and the sample evidence.
Our summary: this is the natural next read after Black-Litterman for anyone who wants their allocation to listen to the data rather than only to equilibrium plus hand-specified views. The “let the data speak” extension makes BL a fully coherent Bayesian updating scheme and is the bridge from the 1991 model back to the Klein-Bawa predictive-density foundation. Practitioner-relevant for systematic books where historical return information should temper both the equilibrium anchor and the trader’s views.
Data and code: methodological extension with empirical illustration; no public code. Behind the JPM/Portfolio Management Research paywall — full PDF not retrievable in this run (browser and Sci-Hub unavailable); a working-paper version circulates under the “Letting the Data Speak” title. Citation indexed for completeness.
Key metrics: the paper reports improved out-of-sample portfolio performance from incorporating the data likelihood versus standard Black-Litterman rather than a single headline metric.