If it helps, this class of problems has a non-existence proof tied to them. Mean-variance finance models have an assumption that all parameters are known built into the proofs. There is an existing proof that shows these models, if true, can never form an estimator for the parameters.
Consider the intertemporal budget constraint from the CAPM. It is commonly written in static models as $\tilde{w}=R\bar{w}+\epsilon.$ Let us assume that $R$ is unknown as I am sure you don't know it and neither do I. If you knew the valuation and so forth, then you wouldn't have needed to ask the question.
So future wealth equals current wealth ttimes a reward for investing plus a shock. You invested to make money so $R>1$. This is a static model of a general case where $w_{t+1}=R{w}_t+\varepsilon_{t+1},$ where $\varepsilon$ is drawn from any density centered on zero with finite variance greater than zero.
Mann and Wald showed that the maximum likelihood estimator, and the MVUE, for this AR(1) process is ordinary least squares subject to the restrictions on the mean and variance in the shock term. This meets all requirements for a Frequentist estimator for $R\in\mathbb{R}$. Mann and Wald showed the sampling distribution for $\hat{R}-R$ where $|R|<1$ is the normal distribution. White in 1958 showed that the sampling distribution for $|R|>1$ is the Cauchy distribution. Since least squares provides a version of the sample mean, you need to find the population mean of the Cauchy distribution for convergence, yet it has no population mean. Consequently, any such estimation has zero power with an infinite sample size.
With all of that said, there is a Bayesian solution to this class of problems, but it doesn't use a mean or a variance. I presented a distribution-free and parameter-free Bayesian solution at the Southwestern Finance Association Conference last week to price options. I also provided a parametric form. It takes advantage of the fact that while the distributions lack a sufficient statistic, predictions don't have that problem.
Save yourself time, ignore WACC. Do use the marginal cost of capital as that is a very real thing. Ignore WACC. Even if the math were valid, it has been shown that you can alway stochastically dominate the solution implying it cannot actually be a solution, even if the math is correct.
If you want to check the intuition to White's result above, consider $w_0=0$ and $\varepsilon_1=1$, where $R>1$. The shock would go to infinity as time went to infinity. Shocks go to zero when the normal distribution is the sampling distribution, implying learning.
Sorry, I am not at a place to provide a citation, but I believe Mann and Wald are either in 1941 or 1943 and White is 1958. Rao generalized White's result to all AR(n) processes, but I do not remember the date.
Proof of models like the CAPM, Black-Scholes, those built using Ito calculus are vacuous even if the assumptions are true in the strictest sense.