5
$\begingroup$

I have just read this paper http://www.cfapubs.org/loi/doi/abs/10.2469/faj.v66.n5.3

In the paper they define financial turbulence formula as:

enter image description here

Could anyone help me calculate/understand this formula, maybe with a simple numeric example?


Numeric example you could improve:

Let's say we want to find the financial turbulence of Dow Jones Industrial at time t, and I have DJI monthly returns of last 24 months (monthly returns created by using month end prices):

DJIR = [0.03, 0.01, -0.04, ..., 0.015]

Thus:

DJIR[1] is 0.03 (it means 3% return)
DJIR[2] is 0.01
DJIR[3] is -0.04
DJIR[4] is -0.02
DJIR[5] is 0.05
...
DJIR[24] is 0.015

Now let's say I want to calculate the financial turbulence d at t = 5

$$d[5] = (DJIR[5] - \mu) * Covariance$$

So what's $\mu$ here at t = 5? How to calculate it? Is it the average of returns until this point, or is it a moving average of the last n returns?

And what's Covariante here at time t = 5, I know Covariance is calculated like this: $\frac{1}{N-1}\sum_{i=1}^{n}(x_i-\bar{x})(y_i-\bar{y})$

But who is $x_i$ and who is $y_i$ in the formula?


P.S. here is another article that explains the formula, but it's not clear to me.

$\endgroup$
2
$\begingroup$

You wrote: $$d[5] = (DJIR[5] - \mu) * Covariance$$ but you left out half of it (the inverse and the transposed vector on the right side). The correct formula is $$d[5] = (DJIR[5] - \mu)^2 / Var[DJIR]$$

The covariance "matrix" becomes the variance in a 1-dimensional case (in other words $x_i$ and $y_i$ are both equal to DJIR[i] in this case) and the "matrix inverse" of a number just becomes one over the number. [That is why you DIVIDE by variance, you don't multiply]. For mu you can use mean[DJIR] over a very long period of time (many years) but you could also set it to zero (which is commonly done and that is what I would recommend, it wont make much difference). Of course the variance is the variance up to this point.

But the formula is intended for more than 1 asset. You already have DJIA, add some other assets. Then you will have a true covariance matrix. The whole point is that there are a lot of markets and this formula lets you watch over all of them simultaneously in a clever way. With 1 market it is trivial.

$\endgroup$
  • $\begingroup$ Let's keep it in one dimension, so just one asset. If $\mu$ is commony set to zero, how do you calculate the $Var[DJIR]$ then? $\mu$ is needed to calculate also the $Var[DJIR] = \frac{1}{N-1}\sum_{i=1}^{n}(DJIR[i]-\mu)^2$. Moreover even by using the $Var$, shouldn't the function become: $$d[5] = (DJIR[5] - \mu) * Var[DJIR]$$ $\endgroup$ – Marco Demaio Aug 21 '15 at 12:05
  • $\begingroup$ (1) Make $mu$ equal to zero throughout. Another idea is use an intelligent value for $mu$ such as 0.0003 (3 basis point per day is about 8% a year which is reasonable for a stock market index). But use the same value everywhere. (2) As you will recall from linear algebra, the apostrophe means vector transpose and $v v'$ means the inner product of v with itself. If v is one dimensional that reduces to $v^2$, not $v$ as you have it. Also look at the charts you linked: the value of d is never negative, That is because they are taking the square! $\endgroup$ – noob2 Aug 21 '15 at 13:55
  • 1
    $\begingroup$ Look, it is very simple and can be described in words, with no formulas: Today's squared return is divided by the average squared return of the most recent T days. That's it. $\endgroup$ – noob2 Aug 21 '15 at 14:08

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.