2
$\begingroup$

Question: Assume we have the following equation:

$$\widehat{\Theta}(\rho) = \frac{1}{(1-\rho)\Delta t} \ln\left(\frac{1}{T} \sum_{t=1}^T \left(\frac{1+r_t}{1+rf_t}\right)^{1-\rho} \right) \ \ \ \ \cdots \ (1)$$

where:

$T =$ total number of observations

$\Delta t =$ length of time between observations

$r_t =$ a variable with random values at each $t$

$rf_t =$ a variable with random values at each $t$

$\rho =$ a constant 'coefficient'

(Note: for those familiar with finance terminology, $r_t$ is an asset's return at time $t$ and $rf_t$ is the risk-free rate at time $t$.)

Computing $\widehat{\Theta}$ does not require any specific distribution for $r_t$ or $rf_t$. I have a large sample of data (i.e, $T = 162,401$) for $r_t$ and $rf_t$.

Now define $\hat{\rho}$ where $\hat{\rho}$ satisfies $\widehat{\Theta}(\hat{\rho}) = 0$. Solving for $\hat{\rho}$ is relatively straightforward using Newton–Raphson method or other linear optimization methods. My aim is to derive the asymptotic distribution for $\hat{\rho}$, that is, as $T \rightarrow \infty$:

  1. What can we say about the asymptotic distribution of $\hat{\rho}$?

  2. Can we derive a particular distribution to which it converges to?

  3. Can we derive it's asymptotic variance/mean?

  4. What other asymptotic properties can we derive?


Attempt: My attempt so far has been related to a 2nd order linear approximation to Eqn.$(1)$ rather than working directly with Eqn.$(1)$. Using Taylor series expansion and properties of the Generalized Mean, we can show that:

$$\widehat{\Theta}(\rho) \approx \frac{1}{\Delta t} \left[\overline{x} + \frac{1-\rho}{2} \left(s_x^*\right)^2\right] \ \ \ \ \ \cdots \ (2)$$

where:

$\overline{x} = \frac{1}{T} \sum_{t=1}^T x_t$ where $x_t = r_t - rf_t$

$\left(s_x^*\right)^2 = \left(\frac{T-1}{T}\right)s_x^2$ where $s_x^2 = \frac{1}{T-1}\sum_{t=1}^T (x_t - \overline{x})^2$

From Eqn.$(2)$, an approximation to $\hat{\rho}$ is given by:

$$\hat{\rho} \approx \frac{2\overline{x}}{\left(s_x^*\right)^2} +1 $$

Since there is no closed form solution to $\hat{\rho}$ from Eqn.$(1)$, how can we derive 'approximate' asymptotic properties of $\hat{\rho}$ using Eqn.$(2)$?

$\endgroup$
  • 1
    $\begingroup$ Have you tried doing MC simulations to see the how distribution looks like? $\endgroup$ – mpiktas Nov 19 '13 at 12:01
  • $\begingroup$ @mpiktas, I haven't actually, but simulating $\widehat{\rho}$ should be pretty straightforward. However, I was more interested if there is some kind of rigorous proof (or references to similar cases) to derive the asymptotic distribution mathematically. $\endgroup$ – TeTs Nov 19 '13 at 12:27
  • $\begingroup$ If the resulting distribution is normal, you can then just think how to apply CLT creatively, which is relatively easy, since there are a lot of similar results to learn from. If the result is not normal then the problem is more complicated. $\endgroup$ – mpiktas Nov 20 '13 at 7:48
2
$\begingroup$

$$\hat{\rho} \approx \frac{2\overline{x}}{\left(s_x^*\right)^2} +1$$ $$\Rightarrow \frac 12 (\hat{\rho} -1) \approx \frac{\overline{x}}{\left(s_x^*\right)^2} = [\left(s_x^*\right)^2]^{-1}\overline{x} = \Big(\frac{1}{T}\sum_{t=1}^T (x_t - \overline{x})^2\Big)^{-1}\cdot \Big(\frac{1}{T} \sum_{t=1}^T x_t\Big)$$

If we can assume that the process $\{X_t\}$ is weakly stationary, with mean $\mu_x$ and variance $\sigma^2_x$ then, by applying the relevant CLT we can quickly arrive at

$$\hat{\rho} \sim_{approx} N\left(1+ \frac {2\mu_x}{\sigma^2_x}, \frac 4{\sigma^2_xT}\right) $$

for "large samples".

ADDENDUM
To clarify some questions that emerged in the comments:
Assuming conditions for CLT holds, then define the variable $W=\bar x -\mu_x$ and obtain

$$\sqrt TW \rightarrow_d N(0, \sigma^2_x)$$

and then as an approximation, we have

$$W \sim_{approx} N(0, \sigma^2_x/T)$$

for large samples. Now substitute in the equation for $\hat \rho$ the expression $\bar x = W+ \mu_x$ to obtain

$$\hat{\rho} \approx \frac{2}{\left(s_x^*\right)^2}(W+ \mu_x) +1 = \frac{2}{\left(s_x^*\right)^2}W+ \Big(\frac{2}{\left(s_x^*\right)^2}\mu_x +1\Big)$$

which, given the approximate distributional result on $W$, gives you the approximate distribution for $\hat \rho$, for large samples, using of course Slutsky as pointed out by @mpiktas. And yes, these results are full of approximations.

$\endgroup$
  • $\begingroup$ Thanks, how exactly did you arrive at the final distribution? I know that $\overline{x} \stackrel{p}{\rightarrow} \mu_x$ and $\left(s_x^*\right)^2 \stackrel{asymp}{\sim} N(\sigma_x^2, \frac{\mu_4-(\sigma_x^2)^2}{n}) $ where $\mu_4 = E[(x_t-\mu_x)^4]$. $\endgroup$ – TeTs Nov 20 '13 at 5:16
  • $\begingroup$ Alecos used the fact that $\sqrt{T}\bar{x}\to N(0,\sigma_x^2)$ and $s_x^2\xrightarrow{P}\sigma^2$ in probability. Then you can employ Slutsky's lemma, if $X_n\to X$ in distribution and $Y_n\to a$ in probability with $a$ being the constant, then $X_nY_n\to aX$ in distribution. $\endgroup$ – mpiktas Nov 20 '13 at 7:51
  • $\begingroup$ @mpiktas, Actually I just want to clarify, why does $\sqrt{T}\overline{x} \rightarrow N(0, \sigma_x^2)$? I thought by the CLT, we have $\frac{\sqrt{T}(\overline{x} - \mu_x)}{\sigma_x} \rightarrow N(0,1)$, so $\sqrt{T}\overline{x} \rightarrow N(0, \sigma_x^2)$ only if we assume $\mu_x = 0$? Why can we make that assumption? Thanks. $\endgroup$ – TeTs Nov 21 '13 at 10:07
  • $\begingroup$ I wrote for clarity, not for correctness. Yes in general you need to subtract the mean to get assymptotic normality. $\endgroup$ – mpiktas Nov 21 '13 at 11:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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