Sampling Frequency and Jump Detection Mike Schwert ECON201FS 2/27/08 This Week’s Approach • Last time: • Counted jump days at various sampling frequencies • Volatility signature plots for RV and BV • This week: • Semivariance calculations for GE price data • Volatility signature plots for RJ, RS- and RS+ • Counted common jump days between different sampling frequencies • Correlation matrices for ZQP-max statistics at different sampling frequencies Realized Semivariance • Introduced by Barndorff-Nielsen, Kinnebrock, and Shephard • Separates positive and negative components of realized variance M M RS t rj 1rj 0 RS t rj 1rj 0 2 j 1 2 j 1 Summary Statistics – GE Price Data, 5-minute sampling frequency Mean Std. Dev Min Max RVolannualized 0.2562 0.3117 0.0529 1.6693 RV 2.6053 x 10-4 3.8553 x 10-4 1.1104 x 10-5 0.0111 RS- 1.2866 x 10-4 1.8166 x 10-4 4.8495 x 10-6 0.0041 RS+ 1.3187 x 10-4 2.1816 x 10-4 4.3544 x 10-6 0.0069 Volatility Signature Plots • Introduced by Andersen, Bollerslev, Diebold, and Labys (1999). • Calculate RJ, RS-, RS+ at 1, 2, …, 30 minute sampling frequencies • Plot relationship between sampling frequency and mean RJ, mean RS • RJ and RS are higher for high-frequency samples because returns are distorted by microstructure noise such as bid-ask bounce • Must be wary of using too low of a sampling frequency, as sampling variation will affect volatility calculations Volatility Signature Plot – Relative Jump Volatility Signature Plot – Negative Semivariance Volatility Signature Plot – Positive Semivariance Contingency and Correlation Matrices • Calculated ZQP-max statistics and counted jump days for GE, ExxonMobil, AT&T, and S&P 500 at 5, 10, 15 and 20 minute sampling frequencies • Counted common jump days between each sampling frequency and organized in contingency matrices • Surprisingly few common jump days exist between sampling frequencies • Calculation of jump days seems to depend a great deal on sampling choices • ZQP-max statistics are relatively uncorrelated between sampling frequencies • GE data minute-by-minute for 1997 – 2007 (2670 days) • ExxonMobil data minute-by-minute for 1999 – 2008 (2026 days) • AT&T data minute-by-minute for 1997 – 2008 (2680 days) • S&P data every 5 minutes, 1985 – 2007 (5545 days, excluding short days) Contingency Tables GE freq S&P 500 5-min 10-min 15-min 20-min 5-min 84 5 1 1 10-min 5 60 4 15-min 1 4 20-min 1 5 freq 5-min 10-min 15-min 20-min 5-min 151 10 5 8 5 10-min 10 95 10 5 42 5 15-min 5 10 92 6 5 40 20-min 8 5 6 76 Exxon Mobil freq AT&T 5-min 10-min 15-min 20-min 5-min 48 6 1 1 10-min 6 34 5 15-min 1 5 20-min 1 4 freq 5-min 10-min 15-min 20-min 5-min 185 22 8 7 4 10-min 22 113 8 11 31 4 15-min 8 8 94 16 4 28 20-min 7 11 16 76 Correlation Matrices - Z-Statistics S&P 500 – N/A GE freq 5-min 10-min 15-min 20-min freq 5-min 10-min 15-min 20-min 5-min 1.000 .1030 .0150 .0431 5-min 1.000 NaN NaN NaN 10-min .1030 1.000 .2241 .0969 10-min NaN NaN NaN NaN 15-min .0150 .2241 1.000 .2809 15-min NaN NaN NaN NaN 20-min .0431 .0969 .2809 1.000 20-min NaN NaN NaN NaN Exxon Mobil AT&T freq 5-min 10-min 15-min 20-min freq 5-min 10-min 15-min 20-min 5-min 1.000 .0849 .0728 .0631 5-min 1.000 .1644 .0844 .0607 10-min .0849 1.000 .2274 .1273 10-min .1644 1.000 .2694 .1753 15-min .0728 .2274 1.000 .2996 15-min .0844 .2694 1.000 .3270 20-min .0631 .1273 .2996 1.000 20-min .0607 .1753 .3270 1.000 Possible Extensions • Try other jump test statistics to see if one outperforms the others in detecting jumps consistently between sampling frequencies • Regress z-statistics on changes in daily volume to see if days with high volume correspond to jump days, common jump days between samples • Other ways to formally analyze effects of sampling frequency on jump detection? • Any way to separate negative jumps from positive jumps? • Effect of sampling frequency on volatility forecasts?