methodology
Posterior Sampling
Drawing many values from a posterior distribution, the beliefs left after combining a prior with data, to turn one forecast into a range of outcomes and probabilities.
In Bayesian statistics, posterior sampling means drawing many values from a posterior distribution: the distribution of plausible parameters or outcomes after a prior belief has been updated with data. Summarizing those draws turns a single best guess into a range, from which probabilities can be read directly. Forecasting tools apply the same idea when they simulate many possible futures for a portfolio.
A forecast of this kind has three parts. First, a model of returns: a distribution with parameters such as the expected return and volatility, ideally drawn from a posterior, or a resampling scheme that reuses the portfolio's own history. Second, a sampler that draws a path of daily returns long enough to cover the horizon (about 252 trading days per year), then many more like it. Third, a tally: for each target, count the share of paths that hit it. A block bootstrap draws chunks of consecutive historical days instead of single days, which keeps calm and turbulent stretches together and preserves the fat tails of real returns.
The output is a probability that holds only under the model's assumptions. Samplers built on the normal distribution tend to understate the odds of big losses, because real returns have fat tails and volatility clusters, while bootstrap samplers repeat whatever history you feed them, including an unusually good or bad stretch. More paths reduce sampling noise: with 5,000 paths, a 50% estimate carries a standard error of about 0.7 points, and with 600 paths about 2.0 points. Longer horizons widen the spread of outcomes.
Example with illustrative numbers: you ask for the chance of at least an 8% annual return over 5 years. A sampler builds 1,200 five-year paths, and 696 of them compound at 8% a year or better, so the estimate is 58%. The same paths might show a 31% chance of a drawdown deeper than 20% somewhere along the way, which is worth knowing before committing.
On GNG, the Portfolio Optimizer runs this kind of simulation after each optimization to estimate the probability of hitting your targets, a minimum annual return and a maximum drawdown tolerance (8% and 25% if you set none), at 1, 5 and 10 years by default, shown in the Probabilities section. Its Posterior simulation setting picks the method: the default Block bootstrap resamples 20-day blocks of the optimized portfolio's historical daily returns, using 5,000 paths at 1 year, 1,200 at 5 years and 600 at 10 years, while Normal (legacy) draws from a normal model and understates tail risk. Neither samples a Bayesian posterior over parameters; the Black-Litterman engine computes its posterior expected returns directly.
The Black-Litterman Model is where a Bayesian posterior enters portfolio construction, and Maximum Drawdown is one of the targets the simulated paths are checked against.
Related terms
More methodology terms
Get full access to GNG Research
Create a free account to access portfolio tracking, advanced tools, and more.