Take the 2-minute tour ×
Quantitative Finance Stack Exchange is a question and answer site for finance professionals and academics. It's 100% free, no registration required.

(Apologies for any formatting mistakes)

Within the Black Scholes model, given that you are estimating the volatility from historical data - and all other parameters assumed exact - one usually substitutes the sample variance as a point estimate for the square of the volatility and evaluates the BScall using that point estimate.

However, why do we use a function of the point estimate instead of the expected value of the distribution of the estimate?

The sample variance follows a Chi-Squared distribution, so we now have a distribution of values of the Call Option based on the observed sample variance and degrees of freedom.

$$ D\sim BSCall \left( \frac{(n-1) \text{s}^2}{\chi_{n-1} ^2} \right) $$

The Expected Value of that distribution is rarely equal to the function of the point estimate.

Example, assume sample variance was .25 out of 52 weekly returns (so n=51 values used to estimate variance):

$$ S=100\\ K=95\\ r=0.10 \\ s^2=.25\\ T=0.25\\ $$

Yields the point estimate of

$$ BSCall(s^2)=13.6953 $$

But

$$ E[D]=13.8372 $$

with 95% confidence intervals of {12.2222, 15.9196}

Distribution of BS Call using Chi Squared

In fact

$$ P[D>BSCall(s^2)]=0.525 $$

Question is two fold:

  1. For using historical data, why do we use a function of the point estimate instead of the expected value of the distribution of the estimate?

  2. If using the point estimate, does the above imply there is a 52% chance the call option is actually undervalued?

Thank you

share|improve this question
    
If you use a biased estimate of the volatility, you will get a biased price for the option. –  experquisite Jan 12 at 17:59
    
Why should one use realized historical volatility for the BS model? In my experience this is hardly ever done. What one rather does is plugging in implied volatility derived from other sources if possible. Another question: what does your notation $BSCall(\frac{(n−1)s^2}{\chi^2_{n-1}})$ exactly mean? –  Richard Jan 14 at 12:26
    
My comment could help to answer your first question. The standard estimator of variance is also an MLE if you assume that the returns are normal (in which case the chi-squared distribution enters naturally). –  Richard Jan 14 at 12:28
    
To answer question 2: no, by no means analysis of historical variance has any necessary implication for option valuation. Option pricing is mainly about trading (delta hedging and so forth). At least theoretically trading will improve the value of your position. –  Richard Jan 14 at 12:32
    
@Richard The notation BSCall with one variable was just shorthand for having all other parameters defined exactly and it is just a function of volatility. The reason for using historical data was from reading "Derivative Markets" (McDonald) in which he used historical weekly data in an example to estimate the volatility, which was then put into the B-S formula. This got me thinking that the expected value of the B-S formula of the distribution of volatility should be used instead. –  sheppa28 Jan 15 at 1:58
add comment

2 Answers

up vote 3 down vote accepted

Two parts

  1. Real world vs risk neutral: Can we even estimate risk neutral volatility using historical data? There is a difference in distribution of the underlying stock price under the real world and risk neutral measures. Luckily, changing to the risk neutral measure does not affect volatility, only the drift. Thus, a real world measure of volatility will properly estimate the risk neutral volatility. In the BS framework, we assume that the stock price is an Ito drift diffusion process with constant coefficients. In equations; $$S_t = S_0 \exp\{(\mu - \sigma^2/2)t + \sigma W_t\} = S_0 \exp\{(r - \sigma^2/2)t + \sigma (W_t - \frac{\mu - r}{\sigma}t)\} \\ = S_0 \exp\{(r - \sigma^2/2)t + \sigma W_t^\star\} $$ see that volatility is the same when writing the equation for stock price in real world or risk neutral.
  2. Estimation: There is more good news, if you look into Statistical Inference by Casella and Berger you may find that given any function $f$, and any maximum likelihood estimator $\hat\Theta$ that the maximum likelihood estimator $\widehat{f(\Theta)}$ is exactly $f(\hat\Theta)$ this is sometmies referred to as "plug-in-principal of MLE". Thus when you are plugging in your MLE estimate $\hat\sigma^2$ into the BS formula in this case taking the place of $f$ you're still obtaining the MLE. Because it is the MLE, we also know that it attains the cramer rao lower bound on variance asymptotically! Still more good news, we can introduce (asymptotic) pricing confidence intervals into our analysis by simply using the delta method approximation of variance for an estimator!
share|improve this answer
add comment

Black-Scholes is just a model that tries to replicate what the market is doing. Unfortunately, any theoretical estimate of volatility (that is not the implied) that you come up with will be wrong.

In fact, you don't want to use historical volatilities at all.

The only correct volatility to use is the IMPLIED VOLATILITY. And the reason why it works is because it is designed to be the one that makes the model work. (sounds redundant, but this is why it's called implied... the one Implied by the model)

share|improve this answer
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.