Scribd
Upload
28 views

Empirical Distributions & Prediction of Returns

This document discusses empirical distributions and prediction of returns in financial markets. It covers random walk models, findings from empirical data, stable distributions, and fractals. The key points are: 1) Log returns are stationary but exhibit short-term memory; 2) The efficient market hypothesis states prices instantly reflect all information, but critics argue investors are not perfectly rational; 3) Random walk models show returns are difficult but not impossible to predict; 4) Empirical data shows distributions are non-Gaussian with fat tails and volatility clustering.

Uploaded by

z_k_j_v
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
28 views

Empirical Distributions & Prediction of Returns

This document discusses empirical distributions and prediction of returns in financial markets. It covers random walk models, findings from empirical data, stable distributions, and fractals. The key points are: 1) Log returns are stationary but exhibit short-term memory; 2) The efficient market hypothesis states prices instantly reflect all information, but critics argue investors are not perfectly rational; 3) Random walk models show returns are difficult but not impossible to predict; 4) Empirical data shows distributions are non-Gaussian with fat tails and volatility clustering.

Uploaded by

z_k_j_v
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
You are on page 1/ 14

4.

Empirical distributions & prediction of returns

4.1 Prices and returns

Price (P) ACF decays slowly. Log price: p = log(P)

Single-period (rate of) return: R(t) = [P(t) - P(t-1)]/P(t-1) = P(t) / P(t-1) 1

Log return: r(t) = log[R(t) + 1] = p(t) p(t-1)


Log returns stationary processes with (relatively) short memory.
Log returns can become negative in simulations (negative P has no sense).

Cumulative return: R(t, k) = P(t) / P(t-k) - 1

Cumulative log return: r(t, k) = r(t) + r(t-1) + + r(t-k)

1
4. Empirical distributions & prediction of returns

4.2 The efficient market hypothesis (EMH)


EMH: markets instantly incorporate all new information in the asset prices.
Hypothesis: explanation of empirical facts.
Random walk hypothesis (RWH): prices follow the random walk.
Bachelier (1900), Fama (1965), Malkiel (2003), Lo & MacKinlay (1999).

EMH (RWH + rational investors):


- Weak form: current price reflects all information on past prices.
Technical analysis impossible; fundamental analysis may work.
- Semi-strong form: prices reflect all publicly available information.
Fundamental analysis doesnt work either.
- Strong form: even private information instantly is incorporated in price.

2
4. Empirical distributions & prediction of returns

4.2 EMH (continued I)


EMH criticisms:
- Investors are not absolutely rational (greed & fear - animal spirits)
Akelrof & Shiller (2009)
- Grossman-Stiglitz paradox
- Lo & MacKinlay (1999): prices do not always follow random walk

Adaptive market hypothesis (Lo (2004)):


Market is an ecological system in which investors compete for scarce
resources. They have limited capabilities (bounded rationality).
Prices reach their new efficient values not instantly but over some time during
which investors adapt to new information by trial and error.

Pragmatic view (Malkiel):


OK, its not random walk but can anyone consistently beat the market?

3
4. Empirical distributions & prediction of returns

4.2 EMH (continued II)

Interviews with Fama & Shiller:

http://www.nytimes.com/2013/10/20/business/robert-shiller-a-skeptic-
and-a-nobel-winner.html

http://www.nytimes.com/2013/10/27/business/sharing-nobel-honors-
and-agreeing-to-disagree.html

http://www.nytimes.com/2013/10/27/business/eugene-fama-king-of-
predictable-markets.html

4
4. Empirical distributions & prediction of returns
4.3 Random walks (Not all random walks are equal...)
pt = pt-1 + + t , E[t] = 0; E[t 2] = 2; E[t s] = 0, if t s

Differential form (Brownian motion):


dpt = t dt + t dWt , Wt = N(0, 1)
E[pt | p0] = p0 + t
Var[pt | p0] = 2t
Sample => ARMA => unit root?
Also: augmented Dickey-Fuller (ADF) test

4.3.1 Random walk model RW1


RW1 when t is strict white noise: t = IID(0, 2).
Gaussian (normal) white noise: t = N(0, 2).
Any forecasting of RW1 is impossible.

5
4. Empirical distributions & prediction of returns

4.3.2 Random walk model RW2

In RW2, t has independent but not identical innovations, i.e. innovations


can be drawn from different distributions.

RW2 is based on the martingale theory. A process Xt is a martingale if


E[Xt | Xt-1, Xt-2,] = Xt-1 or E[Xt - Xt-1 | Xt-1, Xt-2,] = 0
The latter difference represents the outcome of fair game.

Gambling: St. Petersburg paradox, Gamblers ruin, etc.

Forecasting of expected value is impossible but higher moments (variance,


etc.) may be predictable.
6
4. Empirical distributions & prediction of returns

4.3.3 Random walk model RW3

In RW3, innovations remain uncorrelated, i.e. Cov(t, t-k) = 0). Yet higher
moments, e.g. variance, can be dependent, and hence forecastable:
Cov(t2, t-k2) 0.
RW3 is called white noise (but not the strict one). Hence RW3 exhibits
conditional heteroskedasticity, which is well supported with empirical data
on volatility.
RW3 permits only non-linear forecasting of expectations (neural nets,
genetic algorithms).

Empirical returns show weak correlations. Hence, even RW3 can be


technically rejected. As a result, stronger hypotheses on RW1 and RW2 can
be rejected, too.

Bottom line: financial markets are predictable to some degree.

7
8
4. Empirical distributions & prediction of returns

4.6 Empirical findings


- Distributions of daily returns are non-Gaussian
- Yet serial correlations are very small (hard to forecast)
- Squared (or absolute) returns are well correlated
- Volatility may be variable and clustered
- Distributions have high kurtosis and may be skewed
- Distributions have fat tails (power-law asymptotes)
Some reports (e.g. Gabaix (2003)) estimate the power-law index of 3:
equities in 1996 1998 on the grid from minutes to several days.
Several theories offered (review in Schmidt (2004)).
LeBaron (2001): power-law distributions can be generated by a mix of
normal distributions with different time scales.

9
4. Empirical distributions & prediction of returns
Distribution of returns for S&P 500 in 1997 2009

10
4. Empirical distributions & prediction of returns

4.7 Stable distributions (Levi flights)


Sum of two copies from the same distribution has the same distribution.

Fourier transform of probability distribution (characteristic function)


F(q) = f(x) eiqx dx

Levy distribution:
iq |q|[1 i tan(/2)], if 1
ln FL(q) = {
iq |q|[1 + 2i ln(|q|)/)], if = 1

= q/|q|, 0 < 2, -1 1, > 0.

mean; peakedness; skewness; spread.


11
4. Empirical distributions & prediction of returns

4.7 Stable distributions (continued)

Normal distribution ( = 2): fSN(x) = exp[-z2/2]

1
Cauchy distribution ( = 1, = 0): fC(x) =
[1 x 2 ]

Pareto distributions ( < 2) have power-law decay:


fP(|x|) ~ |x| -(1 + )
The problem: moments are divergent... Solution (?): truncation.

Mantegna & Stanley (2000); Bouchaud & Potters (2000); Schmidt (2004)

12
4. Empirical distributions & prediction of returns

4.8 Fractals in finance


Mandelbrot (1997); Peters (1996)
Self-similarity (repeating patterns on smaller scale).
Time series is not isotropic => self-affinity: X(ct) = cH X(t)
H Hurst exponent.

Fractional Brownian motion:


E[BH(t + T) - BH(t)] = 0; E[BH(t + T) - BH(t)]2 =T2H
If H=0.5, regular Brownian motion.

Correlation between E[BH(t) - BH (t - T)]/T and E[BH (t + T) - BH (t)]/T


C = 22H-1 1

13
4. Empirical distributions & prediction of returns
4.8 Fractals in finance (continued)
Persistent process: C > 0 => 1/2 < H < 1
if grew in the past, probably will grow in future
Anti-persistent process: C < 0 => H < 1/2
if grew in the past, probably will fall in future

Rescaled range (R/S) analysis:


Consider a data set xi (i = 1, ... N) with mean mN and variance N2.

Sk = ,1kN

R/S = [max(Sk) - min(Sk)]/N , 1 k N


R/S = (aN)H
Drawback: sensitivity to short-range memory
14

You might also like