Return Distribution Models and Portfolio Optimization Priorities

Summary

The document questions whether more sophisticated stock-return distributions, including multifractal models, materially improve portfolio work beyond the familiar lognormal assumption. It notes that the usual lognormal rationale relies on independent, identically distributed price ratios with finite moments, assumptions that real returns may not satisfy. The response does not identify a preferred modern distribution model; instead, it redirects attention to portfolio optimization inputs.

It ranks expected returns and covariances, including how they are estimated or cleaned, as central concerns, with variances also important. A non-normal objective can incorporate additional features such as kurtosis, but doing so adds parameters and exposure constraints. The answer emphasizes that estimation error, differences between realized and estimated moments, and multiple nearby local solutions may dominate outcomes. Its practical lesson is that refining the assumed return distribution alone may not resolve poor portfolio performance. The discussion is concise and offers no empirical comparison of candidate models, so it serves as a caution about priorities rather than a model-selection guide.

Key ideas

• The lognormal approximation depends on independent, identically distributed price ratios with finite moments.
• Portfolio optimization is highly sensitive to estimated expected returns and covariances.
• Non-normal objectives can account for features such as kurtosis, at the cost of more inputs and constraints.
• Estimation error and competing local solutions may matter more than choosing a refined return distribution.
• The document gives no empirical ranking of modern stock-return models.

Tags

# Probability Distribution of Stock Returns


# Probability Distribution of Stock Returns












Is there a modern theory for the probability distribution of stock returns? It is relatively easy to deduce that under idealized conditions stock returns follow a log normal distribution. One arrives at this by considering the product of ratios of prices ("stock returns"), applying a natural logarithm to convert the product into a sum and then applying the Central Limit Theorem under the condition that the ratios are iid (independent and identically distributed) and have finite mean and variance.

The problem is of course that we cannot just assume that returns are iid or that they have finite variance. So I am seeking alternative theories that try to address these shortcomings.

I am aware of Mandelbrot's Multifractal Model of Asset Returns. Is this considered SOA in the field? Is there something else that is considered a better model or easier to work with?

## Answer by Arshdeep (score 2)

https://quant.stackexchange.com/a/78822

Order of importance of inputs in portfolio optimization are:

- Expected returns (how you clean them)

- Covariances (how you clean them)

- Variances

If you don't want normal distributions, you introduce more variables like kurtosis etc. explicitly into your objective. You then specify limits of exposure to kurtosis.

At this point, most of your risk is coming from the above 3 points as well as the having local solutions. Depending on asset universe, you will have multiple solutions close to each other.

Sorting out these is NOT inside a statistical paradigm anymore. It requires considerations in your portfolio.

In other words, you will see that the lack of performance of your portfolio is because realized means and variances are different from your estimation. What will you do by refining the stock distribution?

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.