Economics 883: The Basic Diffusive Model, Jumps, Variance Measures, and Noise Corrections George Tauchen Economics 883FS Spring 2014
Main Points 1. The Continuous Time Model, Theory and Simulation 2. Observed Data, Plotting and Pitfalls 3. Empirical Variance Measures 4. Options and Implied Volatility 5. Microstructure Noise
1. The Continuous Time Model Dynamics of the log price (p) process: Continuous dp(t) = µ(t)dt + σ(t)dw(t). With jumps dp(t) = µ(t)dt + σ(t)dw(t) + dj(t), where J(t) J(s) = s<τ t κ(τ).
Basic Diffusion 80 78 76 74 72 70 68 66 64 62 60 0.5 1 1.5 2 2.5 x 10 4
Simulation Generated by Euler Scheme Generate Gaussian (normal) random variables: Z j N(0, 1), j = 1, 2,..., J M steps per day: M = 80 = 1/80 Standard deviation σ = 0.011 1.10 percent per day T days:
Euler Scheme (continued) Wiener increments dw j = Z j j = 1, 2,..., J Initialize p 0 = log(75) ($75 per share) Iterate: p j = p j 1 + µ + σ dw j, j = 1, 2,..., J Convert to levels: {P j } J j=1, P j = e p j j = 1, 2,,..., J(= M T ), and make a nice plot.
Euler Scheme: Time Varying Variance Wiener increments dw j = Z j j = 1, 2,..., J Initialize p 0 = log(75) ($75 per share) Iterate: p j = p j 1 + µ + σ j 1 dw j, j = 1, 2,..., J Convert to levels: {P j } J j=1, P j = e p j j = 1, 2,,..., J(= M T ), and make a nice plot.
Simulation of Jump Diffusion by an Euler Scheme Sample the number of jumps N Poisson, then draw N jumps κ k N(0, σ jmp ), k = 1,..., N, and scatter the jumps randomly.
Jump Diffusion With jumps dp(t) = µ(t)dt + σ(t)dw(t) + dj(t), 100 95 90 85 80 75 70 65 60 55 50 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
2. Observed Data: XOM September-October 2008 85 XOM: September October 2008 80 75 70 65 60 55 Sep 02, 2008 Sep 16, 2008 Oct 01, 2008 Oct 16, 2008 Oct 31, 2008
Within-day geometric returns X = p (log-price) r t,i n i X
XOM Returns, September-October, 2008 10 XOM, Return: September October 2008 8 6 4 2 0 2 4 6 8 10 Sep 02, 2008 Sep 16, 2008 Oct 01, 2008 Oct 16, 2008 Oct 31, 2008
Default Plots from Graphics are not Acceptable Poorly labeled axis, scales, and colors: 85 80 75 70 65 60 55 0 2000 4000 6000 8000 10000 12000 14000 16000 18000
Default gives very poor graphic: 6 Default no good 4 2 0 2 4 6 8 10 0 2000 4000 6000 8000 10000 12000 14000 16000 18000
Care with Data: e.g., Stock Splits Consider Apple (APPL): AAPL Price 300 250 200 150 100 50 Jan 02, 2002 Apr 06, 2004 Jul 05, 2006 Oct 03, 2008 Dec 30, 2010
Clearly, something is wrong, or needs attention: AAPL Returns 60 40 20 0 20 40 60 Jan 02, 2002 Apr 06, 2004 Jul 05, 2006 Oct 03, 2008 Dec 30, 2010
Culprit Data recording error or maybe a stock split? AAPL (Yahoo Finance, Charts, click max) Splits: Jun 16, 1987 [2:1], Jun 21, 2000 [2:1], Feb 28, 2005 [2:1] Need to adjust the data for any stock splits. The convention is to adjust backwards. Watch it, sometimes there can be a reverse stock split such as 1:3.
3. Variance Measures The Realized Variance For integer t RV t = n n i X 2 i=1 As 0, n, then from advanced probability theory: RV t t s=t 1 c s ds + t 1<s t X s RV t converges to the integrated variance plus sum of all withinday jumps squared.
Bipower Variation The Bipower variation of Barndorff-Nielsen and Shephard (2003) BV t = µ 1 1 n n 1 M n i X n i X j=2 As 0, M, and BV t t s=t 1 σ 2 (s) ds BV t converges to the integrated variance. It is jump robust. In the absence of jumps RV t BV t 0.
Continuous and Jump Variation Model: dx t = µ t dt + c t dw t + J t, Split the quadratic variation into two pieces: RV t QV t = t s=t 1 σ 2 (s) ds + t 1<s t κ 2 s QV t = IV t + T JV t IV t = t s=t 1 σ 2 (s) ds, T JV t = t 1<s t κ 2 s
Decompose the Realized Variance RV t = CV t + JV t CV t = RV t I(z t c) + BV t I(z t > c) JV t = (RV t BV t ) I(z t > c)
Realized Variance (Annualized) XOM 2007-2010 Annualized Realized Volatility 200 150 100 50 0 Jan 03, 2007 Oct 19, 2007 Aug 11, 2008 Jun 01, 2009 Mar 17, 2010 Dec 30, 2010
Bipower Variation (Annualized) XOM 2007-2010 Annualized Bipower Variation 200 150 100 50 0 Jan 03, 2007 Oct 19, 2007 Aug 11, 2008 Jun 01, 2009 Mar 17, 2010 Dec 30, 2010
Relative Contribution of Jumps XOM 2007-2010 A poorly scaled graphic: 0.5 XOM: Relative Contribution of Jumps 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 Jan 03, 2007 Oct 19, 2007 Aug 11, 2008 Jun 01, 2009 Mar 17, 2010 Dec 30, 2010
Relative Contribution of Jumps XOM 2007-2010 A reasonable well-scaled graphic: XOM: Relative Contribution of Jumps 0.8 0.6 0.4 0.2 0 0.2 Jan 03, 2007 Oct 19, 2007 Aug 11, 2008 Jun 01, 2009 Mar 17, 2010 Dec 30, 2010
Truncated Variance (TV) as Estimator of IV t Threshold Estimator(Mancini, 2009) Truncated Variance: T V t = n n i X 2 I [ n i X cutoff t ] i=1 where I[ ] is the 0-1 indicator function (1 = true, 0 = false).
Value of cutoff for Threshold Variation What is the cutoff? Something like 4 (or 2, 3) standard deviations would be about right, but what standard deviation? Papers by Ait-Sahalia, Todorov, and others provide some guidance. A reasonable choice would be cutoff t = 3 BV t 1 ( n ) ω, 0 < ω < 1 2 where BV t 1 is the bipower variation of the preceding day.
4. Options and Implied Volatility XOM Call Options Expire at close Friday, April 20, 2012 Close of Trading, Friday, January 13, 2012 S 0 = $84.88 Strike Call Price at Close 65.00 $19.30 70.00 14.87 75.00 10.30 80.00 6.45 85.00 3.20 S 0 = $84.88 T = 98/365 = 0.2685
Call-80 $6.45 Implied 19.38 6.45 Valuation Under the Model Option Market Price σ Black-Scholes Value Call Call-80 $6.45 Historical 25 7.27 Call-80 10 5.34 Call-80 30 8.05 Put Put-80 $2.13 Historical 25 2.19 Put-80 10 0.24 Put-80 30 2.95 Option Market Price σ Black-Scholes Value
Option Market Price σ Black-Scholes Value Call Call-90 $1.12 Historical 25 2.49 Call-90 Implied 16.28 1.12
5. Market Microstructure Noise A better term would be trading frictions, or small deviations from the appropriate price. From the present value model, i.e. the correct stock price is: the Gordon Growth model, p t = Πe t r a,t g t
Market Price cannot Possibly Track the Correct Price Perfectly New information continuously flows to the market about the three fundamentals, Π e t, r a,t, g t, and no market can keep the actual price equal to p t. Thus, we write p t = p t + ɛ t
Implications Note the observed geometric return over the interval t t is: r,t = p t p t + ɛ t ɛ t r,t = r,t + ɛ t ɛ t Var (r,t ) = σ 2 r + 2σ2 ɛ As 0 the noise dominates: lim Var (r,t ) 0 = 2σ 2 ɛ
Coarse Sampling 5-Minute returns, or k-minute returns r t,j,k = 100 [p(t 1 + jk/m) p(t 1 + (j 1)k/m)], j = 1, 2,..., end of day Volatility signature plots are helpful.
ALL 1997-2002 (Allstate Corporation) 45 Annualized Volatility Signature 40 35 30 25 10 20 30 40 50 60 Minutes
XOM 2007-2010: 35 Annualized Volatility Signature 30 25 20 15 10 20 30 40 50 60 Minutes
Sub-Sampling RV (0) = r 2 9:35,9:40 + r2 9:40,9:45 + + r2 3:54,3:59 RV (1) = r 2 9:36,9:41 + r2 9:41,9:46 + + r2 3:53,3:58. RV (4) = r 2 9:39,9:44 + r2 9:44,9:49 + + r2 3:50,3:55 RV SS = 1 5 ( RV (0) + RV (1) + RV (2) + RV (3) + RV (4))
Pre-Averaging See papers by Mark Podolskij and many co-authors. (Tentative) Form local averages of the prices to wash out the noise: