Introduction & Random Variables. John Dodson. September 3, 2008

Size: px

Start display at page:

Download "Introduction & Random Variables. John Dodson. September 3, 2008"

Ethan Conley
5 years ago
Views:

1 & September 3, 2008 Statistics

2 Statistics Statistics

3 Course Fall sequence modules portfolio optimization Bill Barr fixed-income markets Chris Bemis calibration & simulation introductions flashcards how to contact, memorable fact module syllabus office hours evaluations & grading module text required Meucci recommended DeGroot, Bertsimas Statistics

4 Module Goal My goal is to introduce modern concepts to describe uncertainty in the financial markets and how to convert investment manager expertise and investor preferences into optimal quantitative investment strategies. Out of Scope I will not be talking about the economics of financial markets, fundamental or technical investment analysis, or the origin of investment manager expertise. Statistics

5 Meucci s Program for Asset Allocation detect market invariance select the invariants estimate the market specify the distribution of the invariants decide on approach to dealing with heteroskedasticity model the market project the invariants to the horizon map the invariants back to market prices define investor optimality objective satisfaction constraints collect manager experience prior on the market vector parameters determine optimal allocation two-step approach Statistics

6 Definitions In practice, a random variable is a quantity that does not have a value known to us right now but will have a value at some point in the future. These are represented as measurable functions of a sample space and usually denoted by upper-case Roman letters. A particular value obtained by a random variable is often denoted by the corresponding lower-case letter. There must be a probability associated with every possible set of outcomes. Such sets are called events and might consist of intervals or points or any combination thereof. The corresponding probability is called the probability measure of the event. Statistics

7 This structure lends itself to a calculus interpretation, where the probability associated with a set is simply the integral of the probability density over that set. P {X (a, b)} = b a f X (x) dx If X ranges over the real numbers R, then f X ( ) must have certain properties. In particular, it must be a non-negative (generalized) function and lim f X (x) = 0 x f X (x) dx = 1 lim f X (x) = 0 x Statistics

8 In addition to the density function, there are at least three other characterizations of a random variable distribution function F X (x) = x f X (x ) dx quantile function Q X (p) = F 1 X (p) characteristic function φ X (t) = ei t x f X (x) dx This last is based on the Fourier transform, where i 2 = 1. While the density function is the most common, these four representations are all equivalent and we will be working with each. They can be distinguished by the nature of their arguments; resp. the value of an outcome, the upper range on a set of outcomes, a probability, and a frequency. Statistics

9 If we have a characterization of a random variable X, it is natural to ask if we can derive the characterization of a function of that variable Y = h(x ), or of a sample, {X 1, X 2,..., X n }, Y = h (X 1, X 2,...). This is in general difficult, but there are some notable easy cases. f, F, and Q for an increasing, invertible function f Y (y) = f ( X h 1 (y) ) h (h 1 (y)) ( F Y (y) = F X h 1 (y) ) Q Y (p) = h (Q X (p)) φ for the mean of a sample, Y n = 1 n n j=1 X j φ Yn (t) = [ ( t )] n φ X n Statistics

10 We will talk in detail about the topic of estimation later, but if you were asked to give a best guess about the value of random variable, the expected value would be a natural answer. The expected is value is a density-weighted average of all possible values EX = = = 1 0 x f X (x) dx 1 2 F X (x) dx Q X (p) dp = i φ (0) Statistics

11 While it may not be possible to evaluate the characterization of function of a random variable, it is generally possible to evaluate the expected value of a function of a random variable. Eh(X ) = h(x) f X (x) dx The probability measure of an event A is the expectation of the indicator function for the event P {X A} = E1 A (X ) In general, the expected value of a function of a random variable is not just value of the function at the expected value of the random variable. Eh(X ) h (EX ) Statistics Do not make this mistake!

12 A foundational result connects the expectation of a random variable with the sample mean. 1 lim n n n X j = EX j=1 almost surely We will demonstrate the weak version of this result using the characteristic function and then look at some numerical results. The key to the derivation is to recognize that ( φ Yn (t) = 1 + i t EX ( t ) ) n + o e i t EX n n Statistics

13 The LLN, combined with information technology, brought about a revolution in applied mathematics last century, introducing a completely novel way to evaluate integrals. integrals are expectations of functions of random variables h(x) dx = h(x) f X (x) f X (x) dx = E h(x ) f X (X ) expectations are means of random samples E h(x ) a.s. 1 = lim f X (X ) n n n j=1 h(x j ) f X (x j ) you can evaluate an arbitrarily complicated integral if you can identify an appropriate random variable generate a very large sample of random variates cheaply evaluate the integrand and density function Statistics

14 Statistics Affine equivariance guides us in defining measures for the location and dispersion of a random variable. In particular, we should expect Loc {a X + b} = a Loc {X } + b Disp {a X + b} = a Disp {X } Location The expected value Loc {X } def = EX is a natural candidate for a measure of location. Dispersion The standard deviation is a natural candidate for a measure of dispersion. Disp {X } def = E (X 2 ) (EX ) 2 Statistics

15 Statistics Other candidates for a measure of location include mode arg max f X median Q X ( 1 2 ) Alternate measures of dispersion include ) 1/2 modal dispersion ( 2 log f arg X max fx inter-quartile range Q X ( 3 4 ) Q X ( 1 4 ) absolute deviation E X EX Standard deviation as a measure of dispersion is justified by an important general result: Chebyshev (Qebyx v) inequality For any r.v. X with a finite standard deviation and k > 0, { } P X EX < k E (X 2 ) (EX ) 2 > 1 1 k 2 Statistics

16 Denote the expected value and standard deviation of a random variable X by µ and σ. The standardized transformation of X is Z = X µ σ Its characteristic function is ( φ Z (t) = e i µ t ) σ t φ X σ The moments of Z measure the skewness, kurtosis, etc. of X. The easiest way to evaluate these is to note that E (Z n ) = ( i) n φ (n) Z (0) N.B.: Moments do not always exist, and they are generally not adequate to characterize a random variable. Statistics

17 The most important distribution for X R is the normal or gaussian distribution. It takes two parameters, and is denoted X N ( µ, σ 2). µ is often called the mean. f X (x) = 1 2π e 1 x µ 2 ( σ ) 2 1 ( σ F X (x) = 1 2 erfc 1 2 x µ ) σ Q X (p) = µ 2 σ erfc 1 (2 p) φ X (t) = e i µ t 1 2 σ2 t 2 Statistics density

18 A foundational result connects the normal distribution with the. For any random variable X with finite expected value µ and standard deviation σ, n lim n 1 n n X j µ N ( 0, σ 2) in distribution j=1 That is, the deviation between the sample mean and the expected value is approximately normal, with standard deviation equal to the standard deviation of the random variable divided by the square root of the sample size. This result can be demonstrated by considering the characteristic function of the RHS above. Statistics

19 Recall, a random variable is a measurable function of the sample space with respect to a probability measure. It can be useful to work with the same random variable under an alternate probability measure. Radon-Nikodym theorem If P and P are equivalent measures, then there is a random variable Z 0 such that for any random variable X E X = E (Z X ) For example, say that X N (µ, σ 2 ) under P, but X N (µ, σ 2 ) under P. Then the Radon-Nikodym derivative is Z = e X σ ( µ σ µ σ ) ( µ σ ) ( µ ) 2 σ Statistics and this same Z would apply to any function of X.

20 The entropy of a distribution is a measure of how much information we do not know about the random variable. H X = E log f X (X ) A technique for increasing the entropy associated with a random variable is to introduce randomness into the characterization of the random variable. For example, X µ N ( µ, σ 2) µ N ( ν, τ 2) If we knew the mean, the r.v. would be normal with a given standard deviation. But we don t know the mean. By letting the mean be random, the entropy is increased. ( σ 2 + τ 2) 1 + log 2π + log σ 2 > 1 H X 1 + log 2π + log = H X µ Statistics

21 Common s Below is a taxonomy of common distributions, classified primarily by the topology of the support. finite Dirac Bernoulli binomial countable geometric Poisson compact uniform beta semi-compact exponential Lévy Gamma unbounded normal Cauchy stable transforms log-normal chi-squared inverse-gamma Pareto mixtures Student-t negative-binomial non-parametric empirical Statistics

MFM Practitioner Module: Risk & Asset Allocation. John Dodson. September 7, 2011

& & MFM Practitioner Module: Risk & Asset Allocation September 7, 2011 & Course Fall sequence modules portfolio optimization Blaise Morton fixed income spread trading Arkady Shemyakin portfolio credit