Minimum Variance Portfolios Explained

The global minimum-variance portfolio is the fully invested portfolio with the smallest estimated variance, requiring no expected-return forecast but depending critically on the covariance matrix. Its apparent simplicity hides the inverse-covariance estimation problem.

Optimization and closed form

For weights \(w\), covariance \(\Sigma\), and unit vector \(\mathbf{1}\):

minimize  w' Σ w
subject to 1' w = 1

w_GMV = Σ⁻¹1 / (1'Σ⁻¹1)

The closed form is useful for tests; in live systems, solve linear systems rather than explicitly inverting Σ, and include realistic constraints.

import numpy as np

def gmv_weights(cov):
    ones = np.ones(cov.shape[0])
    x = np.linalg.solve(cov, ones)
    return x / (ones @ x)

def annualized_vol(weights, cov, periods=252):
    return np.sqrt(periods * weights @ cov @ weights)

An unconstrained solution can be highly leveraged long-short. Long-only and maximum-weight constraints turn the problem into a quadratic program, usually improving implementability and reducing estimation-error amplification.

Why covariance quality dominates

GMV ignores expected returns, but it does not ignore inputs: every weight depends on all covariances. When the number of assets approaches the observation count, sample covariance eigenvalues are noisy and inverse covariance is unstable. Small changes to a correlation can cause large weight changes.

Failure	Symptom	Control
noisy covariance	unstable weights	shrinkage/factor model
changing regime	realized volatility miss	rolling windows, stress
concentration	one sector dominates	group constraints
turnover	excessive trading	turnover penalty
illiquidity	theoretical allocation	capacity and trade limits

Use a shrinkage estimator or factor covariance as described in covariance shrinkage. Estimate risk over multiple horizons and compare predicted with realized risk.

Practical formulation

A useful rebalance objective adds costs and a prior:

minimize w'Σw + λ ||w - w_prev||² + η ||w - w_benchmark||²
subject to 1'w=1, l≤w≤u, sector constraints

The quadratic turnover term is smooth and tractable; actual costs are often piecewise linear and should be evaluated after optimization. The benchmark term prevents accidental bets caused by sparse data or risk-model gaps.

Relation to risk parity

GMV minimizes total variance; it does not equalize risk contributions. It often allocates heavily to low-volatility, correlated assets, whereas risk parity targets balanced marginal risk. Neither guarantees diversification during a correlation spike. Report factor, sector, liquidity, and scenario exposures alongside volatility.

Evaluation

Use walk-forward estimation, point-in-time universes, realized transaction costs, and delisting-aware returns. Compare to equal weight, cap weight, and constrained risk parity; a backtest that only beats a poor benchmark has limited decision value. Analyze performance in stress periods because minimum variance can be implicitly short volatility or long duration.

Key takeaways

• GMV replaces uncertain expected returns with an equally consequential covariance estimate.
• Shrinkage, constraints, and turnover control are central, not optional enhancements.
• Evaluate realized risk, concentration, costs, and stress behavior—not only Sharpe ratio.

Estimation protocol

Choose the investment universe before observing subsequent returns, remove assets only using information available at the rebalance date, and include delisted instruments. Estimate covariance from returns synchronized to a common clock; asynchronous prices can mechanically depress correlations and induce false diversification. Compare multiple shrinkage intensities and factor structures through walk-forward tests rather than choosing the model that best fits one sample.

Measure realized volatility against the forecast at each rebalance and report the distribution of forecast errors. When estimated correlations spike, constraints may bind and change risk composition abruptly. Stress an all-correlation-to-one scenario, asset-specific liquidity shocks, and the loss of shorting capacity. A low ex-ante variance number is not evidence that the portfolio is robust.

Rebalance frequency is itself an optimization parameter. Frequent updates react faster to genuine risk changes but trade more on covariance noise; infrequent updates allow weights and exposures to drift. Select it using out-of-sample net results, then apply a predefined threshold so trivial target changes do not generate needless orders.

Record threshold breaches so future governance can distinguish disciplined inaction from operational delay. Apply the same review to forced rebalances caused by corporate actions.