optimizer
Mean-Variance Optimization
A method that picks portfolio weights to maximize expected return for a chosen risk level, balancing reward against volatility.
Mean-variance optimization chooses portfolio weights by trading off expected return (the mean) against risk (the variance of returns). Harry Markowitz introduced it in 1952, and it remains the foundation of quantitative portfolio construction. Its central insight is that each holding should be judged by how it changes the whole portfolio's risk and return, which depends on how it moves with the other holdings.
The classic objective maximizes expected return minus a penalty for variance: expected return minus half of lambda multiplied by variance. Expected return is the weighted sum of each asset's expected return; variance comes from the covariance matrix, which combines every asset's volatility with every pairwise correlation; lambda is the risk-aversion setting, and higher values favor safer portfolios. A common variant drops lambda and directly maximizes the Sharpe ratio, finding the single portfolio with the best excess return per unit of volatility. Constraints such as no short selling, position caps and sector caps are added on top.
The output is a set of weights on the efficient frontier for your inputs. The method is known for being sensitive to those inputs: a small change in expected returns can swing the weights sharply and pile capital into a few names, a problem often described as error maximization. Common fixes include shrinking the covariance estimate, blending expected returns with a neutral starting point (the Black-Litterman approach) and capping positions. Treat the result as the best answer to your assumptions, and test how much it moves when the assumptions change.
Example with illustrative numbers: Portfolio A expects 8% with 15% volatility (a variance of 0.0225) and Portfolio B expects 7% with 10% volatility (a variance of 0.01). With lambda at 4, A scores 0.08 minus 2 multiplied by 0.0225, or 0.035, and B scores 0.07 minus 0.02, or 0.05, so B wins. With lambda at 1, A scores about 0.069 and B 0.065, so the less risk-averse investor picks A.
On GNG, Mean-Variance is one of the Portfolio Optimizer's two recommended engines. For universes under 500 names it first maximizes the Sharpe ratio within your limits; larger universes, and runs where that solve fails, use the risk-penalty form with lambda fixed at 3, then minimum variance, and finally hierarchical risk parity with a notice. It uses a Ledoit-Wolf covariance estimate (a statistical factor model for very large universes), forward-looking expected returns and a small extra penalty that discourages extreme weights. The Risk aversion (lambda) setting in the Solver knobs drives the default Black-Litterman engine, which applies the same machinery after blending market-implied returns with your views or with GNG's own estimates.
The Efficient Frontier is the set of portfolios this method traces, and the Sharpe Ratio is the quantity its most common version maximizes.
Formula
Related terms
More optimizer terms
Get full access to GNG Research
Create a free account to access portfolio tracking, advanced tools, and more.