Finally, allowing for transaction costs, it is worth noting that all the trading strategies

retained produce positive returns, except some based on the GARCH (1,1) benchmark

model for the USD/JPY volatility. RNN models appear as the best single modelling

approach for short-term volatility trading. Somewhat surprisingly, model combination,

the overall best performing approach in terms of forecasting accuracy, fails to improve

the RNN-based volatility trading results.

4.7 CONCLUDING REMARKS AND FURTHER WORK

The rationale for this chapter was to develop a nonlinear nonparametric approach to

forecast FX volatility, identify mispriced options and subsequently develop a trading

strategy based upon this modelling procedure.

Using daily data from December 1993 through April 1999, we examined the use of

NNR and RNN regression models for forecasting and subsequently trading currency

volatility, with an application to the GBP/USD and USD/JPY exchange rates.

These models were then tested out-of-sample over the period April 1999“May 2000,

not only in terms of forecasting accuracy, but also in terms of trading performance. In

150 Applied Quantitative Methods for Trading and Investment

order to do so, we applied a realistic volatility trading strategy using FX option straddles

once mispriced options had been identi¬ed.

Allowing for transaction costs, most of the trading strategies retained produced positive

returns. RNN models appeared as the best single modelling approach in a short-term

trading context.

Model combination, despite its superior performance in terms of forecasting accuracy,

failed to produce superior trading strategies. Admittedly, other combination procedures

such as decision trees, neural networks, as in Donaldson and Kamstra (1996) or Shadbolt

and Taylor (2002), and unanimity or majority voting schemes as applied by Albanis and

Batchelor (2001) should be investigated.

Further work is also needed to compare the results from NNR and RNN models with

those from more “re¬ned” parametric or semiparametric approaches than our GARCH

(1,1) benchmark model, such as Donaldson and Kamstra (1997), So et al. (1999), Bollen

et al. (2000), Flˆ res and Roche (2000) and Beine and Laurent (2001).

o

Finally, applying dynamic risk management, the trading strategy retained could also

be re¬ned to integrate more realistic trading assumptions than those of a ¬xed holding

period of either 5 or 21 trading days.

However, despite the limitations of this chapter, we were clearly able to develop rea-

sonably accurate FX volatility forecasts, identify mispriced options and subsequently

simulate a pro¬table trading strategy. In the circumstances, the unambiguous implication

from our results is that, for the period and currencies considered, the currency option

market was inef¬cient and/or the pricing formulae applied by market participants were

inadequate.

ACKNOWLEDGEMENTS

We are grateful to Professor Ken Holden and Professor John Thompson of Liverpool

Business School and to an anonymous referee for helpful comments on an earlier version

of this chapter. The usual disclaimer applies.

APPENDIX A

Table A4.1 Summary statistics “ GBP/USD realised and implied

1-month volatility (31/12/1993“09/04/1999)

Sample observations : 1 to 1329

Variable(s) : Historical vol. Implied vol.

Maximum : 15.3644 15.0000

Minimum : 2.9448 3.2500

Mean : 6.8560 8.2575

Std. Dev. : 1.9928 1.6869

Skewness : 0.69788 0.38498

Kurtosis “ 3 : 0.81838 1.1778

Coeff. of variation : 0.29067 0.20429

Forecasting and Trading Currency Volatility 151

Table A4.2 Correlation matrix of realised and

implied volatility (GBP/USD)

Realised vol. Implied vol.

Realised vol. 1.0000 0.79174

Implied vol. 0.79174 1.0000

Implied Vol. Realised Vol.

16 16

14 14

12 12

10 10

8 8

.

6 6

4 4

2 2

0 0

05/01/94

05/04/94

05/07/94

05/10/94

05/01/95

05/04/95

05/07/95

05/10/95

05/01/96

05/04/96

05/07/96

05/10/96

05/01/97

05/04/97

05/07/97

05/10/97

05/01/98

05/04/98

05/07/98

05/10/98

05/01/99

05/04/99

Figure A4.1 GBP/USD realised and implied volatility (%) from 31/12/1993 to 09/04/1999

Table A4.3 Summary statistics “ USD/JPY realised and implied

1-month volatility (31/12/1993“09/04/1999)

Sample observations : 1 to 1329

Variable(s) : Historical vol. Implied vol.

Maximum : 33.0446 35.0000

Minimum : 4.5446 6.1500

Mean : 11.6584 12.2492

Std. Dev. : 5.2638 3.6212

Skewness : 1.4675 0.91397

Kurtosis “ 3 : 2.6061 2.1571

Coeff. of variation : 0.45150 0.29563

Table A4.4 Correlation matrix of realised and

implied volatility (USD/JPY)

Realised vol. Implied vol.

Realised vol. 1.0000 0.80851

Implied vol. 0.80851 1.0000

152 Applied Quantitative Methods for Trading and Investment

Implied Vol. Realised Vol.

40 40

35 35

30 30

25 25

20 20

15 15

10 10

5

5

0

0

05/01/94

05/04/94

05/07/94

05/10/94

05/01/95

05/04/95

05/07/95

05/10/95

05/01/96

05/04/96

05/07/96

05/10/96

05/01/97

05/04/97

05/07/97

05/10/97

05/01/98

05/04/98

05/07/98

05/10/98

05/01/99

05/04/99

Figure A4.2 USD/JPY realised and implied volatility (%) from 31/12/1993 to 09/04/1999

APPENDIX B

B.4.1 GBP/USD GARCH (1,1) assuming a t-distribution and Wald test

GBP/USD GARCH (1,1) assuming a t-distribution

converged after 30 iterations

Dependent variable is DLUSD

1327 observations used for estimation from 3 to 1329

Regressor Coef¬cient Standard error T-Ratio[Prob]

ONE 0.0058756 0.010593 0.55466[0.579]

DLUSD(’1) 0.024310 0.027389 0.88760[0.375]

’0.0011330

R-Squared R-Squared

0.0011320

F -stat. F (3,1323)

S.E. of regression 0.45031 0.49977[0.682]

Mean of dependent 0.0037569 S.D. of dependent variable 0.45006

variable

’749.6257

Residual sum of 268.2795 Equation log-likelihood

squares

’754.6257 ’767.6024

Akaike info. Schwarz Bayesian criterion

criterion

DW-statistic 1.9739

Forecasting and Trading Currency Volatility 153

Parameters of the Conditional Heteroskedastic Model

Explaining H-SQ, the Conditional Variance of the Error Term

Coef¬cient Asymptotic standard error

Constant 0.0021625 0.0015222

E-SQ(’1) 0.032119 0.010135

H-SQ(’1) 0.95864 0.013969

D.F. of t-dist. 5.1209 0.72992

H-SQ stands for the conditional variance of the error term.

E-SQ stands for the square of the error term.

Wald test of restriction(s) imposed on parameters

Based on GARCH regression of DLUSD on:

ONE DLUSD(’1)

1327 observations used for estimation from 3 to 1329

Coef¬cients A1 to A2 are assigned to the above regressors respectively.

Coef¬cients B1 to B4 are assigned to ARCH parameters respectively.

List of restriction(s) for the Wald test:

b2 = 0; b3 = 0

CHSQ(2) = 16951.2[0.000]

Wald statistic

B.4.2 GBP/USD GARCH (1,1) assuming a normal distribution and Wald test

GBP/USD GARCH (1,1) assuming a normal distribution

converged after 35 iterations

Dependent variable is DLUSD

1327 observations used for estimation from 3 to 1329

Regressor Coef¬cient Standard error T-Ratio[Prob]

ONE 0.0046751 0.011789 0.39657[0.692]

DLUSD(’1) 0.047043 0.028546 1.6480[0.100]

’0.0010999

R-Squared R-Squared

0.0011651

F -stat. F (3,1323)

S.E. of regression 0.45030 0.51439[0.672]

Mean of dependent 0.0037569 S.D. of dependent variable 0.45006

variable

’796.3501

Residual sum of 268.2707 Equation log-likelihood

squares

’800.3501 ’810.7315

Akaike info. Schwarz Bayesian criterion

criterion

DW-statistic 2.0199

154 Applied Quantitative Methods for Trading and Investment

Parameters of the Conditional Heteroskedastic Model

Explaining H-SQ, the Conditional Variance of the Error Term

Coef¬cient Asymptotic standard error

Constant 0.0033874 0.0016061

E-SQ(’1) 0.028396 0.0074743

H-SQ(’1) 0.95513 0.012932

H-SQ stands for the conditional variance of the error term.

E-SQ stands for the square of the error term.

Wald test of restriction(s) imposed on parameters

Based on GARCH regression of DLUSD on:

ONE DLUSD(’1)

1327 observations used for estimation from 3 to 1329

Coef¬cients A1 to A2 are assigned to the above regressors respectively.

Coef¬cients B1 to B3 are assigned to ARCH parameters respectively.

List of restriction(s) for the Wald test:

b2 = 0; b3 = 0

CHSQ(2) = 16941.1[0.000]

Wald statistic

B.4.3 USD/JPY GARCH (1,1) assuming a t-distribution and Wald test

USD/JPY GARCH (1,1) assuming a t-distribution

converged after 26 iterations

Dependent variable is DLYUSD

1327 observations used for estimation from 3 to 1329

Regressor Coef¬cient Standard error T-Ratio[Prob]

ONE 0.037339 0.015902 2.3480[0.019]

DLYUSD(’1) 0.022399 0.026976 0.83032[0.407]

’0.0011874

R-Squared R-Squared

0.0010778

F -stat. F (3,1323)

S.E. of regression 0.79922 0.47580[0.699]

Mean of dependent 0.0056665 S.D. of dependent variable 0.79875

variable

’1385.3

Residual sum of 845.0744 Equation log-likelihood

squares

’1390.3 ’1403.3

Akaike info. Schwarz Bayesian criterion

criterion

DW-statistic 1.8999

Forecasting and Trading Currency Volatility 155

Parameters of the Conditional Heteroskedastic Model

Explaining H-SQ, the Conditional Variance of the Error Term

Coef¬cient Asymptotic standard error

Constant 0.0078293 0.0045717

E-SQ(’1) 0.068118 0.021094

H-SQ(’1) 0.92447 0.023505

D.F. of t-dist. 4.3764 0.54777

H-SQ stands for the conditional variance of the error term.

E-SQ stands for the square of the error term.

Wald test of restriction(s) imposed on parameters

Based on GARCH regression of DLYUSD on:

ONE DLYUSD(’1)

1327 observations used for estimation from 3 to 1329

Coef¬cients A1 to A2 are assigned to the above regressors respectively.

Coef¬cients B1 to B4 are assigned to ARCH parameters respectively.

List of restriction(s) for the Wald test:

b2 = 0; b3 = 0

CHSQ(2) = 11921.3[0.000]

Wald Statistic

APPENDIX C

GARCH Vol. Forecast Realised Vol.

40 40

35 35

30 30

25 25

20 20

.

15 15

10 10

5 5

0 0

05/01/94

05/05/94

05/09/94

05/01/95

05/05/95

05/09/95

05/01/96

05/05/96

05/09/96

05/01/97

05/05/97

05/09/97

05/01/98

05/05/98

05/09/98

05/01/99

05/05/99

05/09/99

05/01/00

05/05/00

Figure C4.1 USD/JPY GARCH (1,1) volatility forecast (%)

156 Applied Quantitative Methods for Trading and Investment

APPENDIX D

Table D4.1a GBP/USD NNR test results for the validation data set

NNR NNR NNR NNR NNR

(44-1-1) (44-5-1) (44-10-1) (44-10-5-1) (44-15-10-1)

Explained variance 1.4% 5.9% 8.1% 12.5% 15.6%

Average relative error 0.20 0.20 0.20 0.20 0.21

Average absolute error 1.37 1.39 1.40 1.41 1.44

Average direction error 33.3% 32.3% 31.5% 30.5% 31.5%

Table D4.1b GBP/USD RNN test results for the validation data set

RNN RNN RNN RNN RNN

(44-1-1) (44-5-1) (44-10-1) (44-10-5-1) (44-15-10-1)

Explained variance 13.3% 9.3% 5.5% 6.8% 12.0%

Average relative error 0.18 0.19 0.19 0.20 0.20

Average absolute error 1.25 1.29 1.33 1.38 1.42

Average direction error 30.1% 32.6% 32.6% 30.8% 32.3%

NNR/RNN (a-b-c) represents different neural network models, where:

a = number of input variables;

b = number of hidden nodes;

c = number of output nodes.

Realised Vol(t) = f [lVol(t ’ 21), Realised Vol(t ’ 21), |r|(t ’ 21, . . . , t ’ 41), DLGOLD(t ’ 21, . . . , t ’ 41)].

Table D4.2a USD/JPY NNR test results for the validation data set

NNR NNR NNR NNR NNR

(44-1-1) (44-5-1) (44-10-1) (44-10-5-1) (44-15-10-1)

Explained variance 5.1% 5.4% 5.4% 2.5% 2.9%

Average relative error 0.16 0.16 0.16 0.16 0.16

Average absolute error 1.88 1.87 1.86 1.85 1.84

Average direction error 30.1% 30.8% 30.8% 32.6% 32.3%

Table D4.2b USD/JPY RNN test results for the validation data set

RNN RNN RNN RNN RNN

(44-1-1) (44-5-1) (44-10-1) (44-10-5-1) (44-15-10-1)

Explained variance 8.4% 8.5% 8.0% 3.2% 3.0%

Average relative error 0.16 0.16 0.16 0.16 0.16

Average absolute error 1.86 1.85 1.84 1.85 1.85

Average direction error 30.1% 29.4% 29.4% 31.5% 30.1%

NNR/RNN (a-b-c) represents different neural network models, where:

a = number of input variables;

b = number of hidden nodes;

c = number of output nodes.

Realised Vol(t) = f [lVol(t ’ 21), Realised Vol(t ’ 21), |r|(t ’ 21, . . . , t ’ 41), DLOIL(t ’ 21, . . . , t ’ 41)].

0

5

10

15

20

25

30

35

0

2

4

6

8

10

12

14

16

18

05/01/94

05/01/94

05/05/94

05/05/94

05/09/94

05/09/94

05/01/95

05/01/95

05/05/95

05/05/95

Figure E4.2

Figure E4.1

05/09/95

05/09/95

05/01/96

05/01/96

05/05/96

05/05/96

05/09/96

Forecasting and Trading Currency Volatility

05/09/96

RNN Vol. Forecast

05/01/97 05/01/97

.

05/05/97

RNN Vol. Forecast

05/05/97

05/09/97 05/09/97

APPENDIX E

05/01/98 05/01/98

05/05/98 05/05/98

05/09/98

Realised Vol.

05/09/98

Realised Vol.

05/01/99 05/01/99

05/05/99 05/05/99

USD/JPY RNN (44-5-1) volatility forecast (%)

GBP/USD RNN (44-1-1) volatility forecast (%)

05/09/99 05/09/99

05/01/00

05/01/00

05/05/00

05/05/00

0

2

4

6

8

0

5

10

15

20

25

30

35

10

12

14

16

18

157

158

0

5

10

15

20

25

30

35

0

2

4

6

8

10

12

14

16

18

05/01/94 05/01/94

05/05/94 05/05/94

05/09/94 05/09/94

05/01/95 05/01/95

05/05/95 05/05/95

05/09/95

Figure F4.1

05/09/95

05/01/96 05/01/96

GR Comb.

GR Comb.

05/05/96 05/05/96

05/09/96 05/09/96

05/01/97 05/01/97

.

05/05/97 05/05/97

05/09/97 05/09/97

COM1 Comb.

APPENDIX F

COM1 Comb.

05/01/98 05/01/98

05/05/98 05/05/98

05/09/98 05/09/98

05/01/99 05/01/99

05/05/99 05/05/99

Realised Vol.

Figure F4.2 USD/JPY volatility forecast combinations (%)

GBP/USD volatility forecast combinations (%)

Realised Vol.

05/09/99 05/09/99

05/01/00 05/01/00

05/05/00 05/05/00

0

5

0

2

4

6

8

10

15

20

25

30

35

10

12

14

16

18

Applied Quantitative Methods for Trading and Investment

Forecasting and Trading Currency Volatility 159

APPENDIX G

Table G4.1 GBP/USD weekly volatility trading strategy

1 Threshold = 0.5 Trading days = 5

In-sample Out-of-sample

Sample

Observation (1-1329) (1330-1610)

Period (31/12/1993-09/04/1999) (12/04/1999-09/05/2000)

Models NNR(44-10-5-1) RNN(44-1-1) GARCH(1,1) COM1 GR NNR(44-10-5-1) RNN(44-1-1) GARCH(1,1) COM1 GR

’12.85% 8.30% 30.08%

P/L without gearing 162.15% 160.00% 262.41% 189.82% 313.55% 15.37% 13.30%

’2.39% 20.92% 53.64%

P/L with gearing 626.62% 852.70% 889.04% 597.16% 792.87% 36.46% 58.98%

Total trades 221 221 224 210 232 39 43 40 35 33

Profitable trades 66.97% 68.07% 81.70% 72.86% 91.38% 53.85% 60.47% 30.00% 45.71% 72.73%

Average gearing 2.86 4.20 2.71 2.54 2.19 1.72 2.58 1.82 1.62 1.63

2 Threshold = 1.0 Trading days = 5

In-sample Out-of-sample

Sample

Observation (1-1329) (1330-1610)

Period (31/12/1993-09/04/1999) (12/04/1999-09/05/2000)

Models NNR(44-10-5-1) RNN(44-1-1) GARCH(1,1) COM1 GR NNR(44-10-5-1) RNN(44-1-1) GARCH(1,1) COM1 GR

P/L without gearing 166.15% 147.13% 213.13% 167.03% 243.71% 20.09% 12.02% 7.87% 11.11% 18.57%

P/L with gearing 333.20% 420.39% 300.56% 366.44% 27.01% 20.87% 12.80% 14.27% 23.42%

434.96%

Total trades 166 166 147 138 21 27 17 13 14

153

Profitable trades 73.49% 72.00% 77.55% 94.93% 76.19% 55.56% 58.82% 76.92% 85.71%

84.31%

Average gearing 1.82 2.35 1.73 1.61 1.42 1.24 1.64 1.31 1.16 1.21

3 Threshold = 1.5 Trading days = 5

In-sample Out-of-sample

Sample

Observation (1-1329) (1330-1610)

Period (31/12/1993-09/04/1999) (12/04/1999-09/05/2000)

Models NNR(44-10-5-1) RNN(44-1-1) GARCH(1,1) COM1 GR NNR(44-10-5-1) RNN(44-1-1) GARCH(1,1) COM1 GR

P/L without gearing 128.90% 137.62% 176.30% 122.22% 129.73% 16.49% 7.06% 7.08% 7.08% 4.39%

P/L with gearing 198.76% 278.91% 268.10% 173.28% 158.83% 19.49% 10.52% 8.56% 7.61% 5.21%

Total trades 113 113 94 82 61 10 18 4 4 2

Profitable trades 74.34% 69.70% 93.62% 82.93% 95.08% 90.00% 50.00% 100.00% 100% 100.00%

Average gearing 1.49 1.81 1.45 1.38 1.21 1.12 1.27 1.19 1.07 1.14

4 Threshold = 2.0 Trading days = 5

In-sample Out-of-sample

Sample

Observation (1-1329) (1330-1610)

Period (31/12/1993-09/04/1999) (12/04/1999-09/05/2000)

Models NNR(44-10-5-1) RNN(44-1-1) GARCH(1,1) COM1 GR NNR(44-10-5-1) RNN(44-1-1) GARCH(1,1) COM1 GR

P/L without gearing 95.08% 115.09% 118.43% 62.47% 45.67% 3.26% 7.25% 4.67% 3.26% -

P/L with gearing 131.34% 194.33% 158.35% 79.39% 52.97% 3.99% 9.27% 5.65% 3.48% -

Total trades 68 68 56 36 18 1 10 2 1 0

Profitable trades 82.35% 72.44% 94.64% 88.89% 94.44% 100.00% 80.00% 100% 100% -

Average gearing 1.34 1.55 1.31 1.29 1.14 1.22 1.14 1.16 1.07 -

Note: Cumulative P/L figures are expressed in volatility points.

160 Applied Quantitative Methods for Trading and Investment

Table G4.2 USD/JPY weekly volatility trading strategy

1 Threshold = 0.5 Trading days = 5

In-sample Out-of-sample

Sample

Observation (1-1329) (1330-1610)

Period (31/12/1993-09/04/1999) (12/04/1999-09/05/2000)

Models NNR(44-1-1) RNN(44-5-1) GARCH(1,1) COM1 GR NNR(44-1-1) RNN(44-5-1) GARCH(1,1) COM1 GR

’10.17% ’11.66% ’82.07% 39.23% 69.31%

P/L without gearing 208.63% 77.06% 387.38% 63.95% 78.77%

’79.31% ’32.08% ’359.44% 111.86% 206.43%

P/L with gearing 1675.67% 517.50% 1095.73% 251.14% 372.70%

Total trades 251 251 242 232 229 50 50 53 48 50

Profitable trades 51.00% 50.60% 54.13% 55.17% 78.60% 76.00% 86.00% 18.87% 60.42% 78.00%

Average gearing 3.76 3.88 3.67 2.67 2.42 3.22 3.73 5.92 2.18 2.35

2 Threshold = 1.0 Trading days = 5

In-sample Out-of-sample

Sample

Observation (1-1329) (1330-1610)

Period (31/12/1993-09/04/1999) (12/04/1999-09/05/2000)

Models NNR(44-1-1) RNN(44-5-1) GARCH(1,1) COM1 GR NNR(44-1-1) RNN(44-5-1) GARCH(1,1) COM1 GR

’75.61%

P/L without gearing 18.73% 30.43% 234.66% 142.37% 308.50% 52.70% 63.01% 31.71% 76.88%

’162.97%

P/L with gearing 0.43% 102.50% 331.50% 578.10% 88.19% 135.49% 49.88% 124.63%

897.30%

Total trades 213 213 168 165 39 43 53 30 37

201

Profitable trades 52.11% 54.17% 58.93% 82.42% 76.92% 81.40% 24.53% 66.67% 94.59%

54.23%

Average gearing 2.16 2.16 2.15 1.68 1.65 1.77 1.98 2.96 1.52 1.56

3 Threshold = 1.5 Trading days = 5

In-sample Out-of-sample

Sample

Observation (1-1329) (1330-1610)

Period (31/12/1993-09/04/1999) (12/04/1999-09/05/2000)

Models NNR(44-1-1) RNN(44-5-1) GARCH(1,1) COM1 GR NNR(44-1-1) RNN(44-5-1) GARCH(1,1) COM1 GR

’65.02% 5.33% 52.59%

P/L without gearing 34.15% 54.46% 251.07% 104.67% 282.96% 64.00% 55.83%

’102.88% 11.99% 77.99%

P/L with gearing 60.28% 164.86% 643.47% 203.58% 399.48% 108.52% 81.09%

Total trades 157 157 144 98 104 33 35 50 13 22

Profitable trades 54.78% 56.21% 63.89% 60.20% 86.54% 81.82% 80.00% 24.00% 69% 100.00%

Average gearing 1.83 1.75 1.78 1.45 1.35 1.56 1.51 2.00 1.38 1.42

4 Threshold = 2.0 Trading days = 5

In-sample Out-of-sample

Sample

Observation (1-1329) (1330-1610)

Period (31/12/1993-09/04/1999) (12/04/1999-09/05/2000)

Models NNR(44-1-1) RNN(44-5-1) GARCH(1,1) COM1 GR NNR(44-1-1) RNN(44-5-1) GARCH(1,1) COM1 GR

’6.13% ’43.71% 9.23%

P/L without gearing 14.08% 247.22% 71.30% 154.88% 49.61% 53.90% 30.33%

’35.86% ’52.64% 14.94%

P/L with gearing 63.43% 514.30% 105.49% 202.86% 71.66% 86.20% 40.28%

Total trades 117 117 113 47 50 21 23 47 6 11

Profitable trades 51.28% 52.94% 70.80% 65.96% 86.00% 90.48% 86.96% 28% 83% 100%

Average gearing 1.58 1.66 1.56 1.28 1.25 1.46 1.52 1.74 1.55 1.30

REFERENCES

Albanis, G. T. and R. A. Batchelor (2001), 21 Nonlinear Ways to Beat the Market. In Developments

in Forecast Combination and Portfolio Choice, Dunis C, Moody J, Timmermann A (eds); John

Wiley: Chichester.

Baillie, R. T. and T. Bollerslev (1989), The Message in Daily Exchange Rates: A Conditional

Variance Tale. Journal of Business and Economic Statistics 7: 297“305.

Baillie, R. T. and T. Bollerslev (1990), Intra-day and Inter-market Volatility in Foreign Exchange

Rates. Review of Economic Studies 58: 565“585.

Baillie, R. T. and T. Bollerslev (1992), Prediction in Dynamic Models with Time-Dependent Con-

ditional Variances. Journal of Econometrics 52: 91“113.

Balkin, S. D. (1999), Stationarity Concerns When Forecasting Using Neural Networks, Presentation

at the INFORMS Conference, Philadelphia, PA.

Bartlmae, K. and F. A. Rauscher (2000), Measuring DAX Market Risk: A Neural Network Volatil-

ity Mixture Approach, Presentation at the FFM2000 Conference, London, 31 May“2 June.

Beine, M. and S. Laurent (2001), Structural Change and Long Memory in Volatility: New Evidence

from Daily Exchange Rates. In Developments in Forecast Combination and Portfolio Choice,

Dunis, C., Moody, J. and Timmermann, A. (eds); John Wiley: Chichester.

Black, F. and M. Scholes (1973), The Pricing of Options and Corporate Liabilities. Journal of

Political Economy 81: 637“654.

Forecasting and Trading Currency Volatility 161

Bollen, P. B., S. F. Gray and R. E. Whaley (2000), Regime-Switching in Foreign Exchange Rates:

Evidence from Currency Option Prices. Journal of Econometrics 94: 239“276.

Bollerslev, T. (1986), Generalized Autoregressive Conditional Heteroskedasticity. Journal of Econo-

metrics 31: 307“327.

Bollerslev, T. (1987), A Conditional Heteroskedastic Time Series Model for Speculative Prices and

Rates of Return. Review of Economics and Statistics 69: 542“547.

Burgess, A. N. and A. N. Refenes (1996), The Use of Error Feedback Terms in Neural Network

Modelling of Financial Time Series. In Forecasting Financial Markets, Dunis, C. (ed.); John

Wiley: Chichester.

Cheng, B. and D. M. Titterington (1994), Neural Networks: A Review from a Statistical Perspec-

tive. Statistical Science 9: 2“54.

Chiras, D. P. and S. Manaster (1978), The Information Content of Option Prices and a Test of

Market Ef¬ciency. Journal of Financial Economics 6: 213“234.

Cybenko, G. (1989), Approximation by Superposition of a Sigmoidal Function. Mathematical Con-

trol, Signals and Systems 2: 303“314.

Deboeck, G. J. (1994), Trading on the Edge“Neural, Genetic and Fuzzy Systems for Chaotic Finan-

cial Markets; John Wiley: New York.

Donaldson, R. G. and M. Kamstra (1996), Forecast Combining with Neural Networks. Journal of

Forecasting 15: 49“61.

Donaldson, R. G. and M. Kamstra (1997), An Arti¬cial Neural Network“GARCH Model for Inter-

national Stock Return Volatility. Journal of Empirical Finance 4: 17“46.

Dunis, C. (1996), The Economic Value of Neural Network Systems for Exchange Rate Forecasting.

Neural Network World 1: 43“55.

Dunis, C. and M. Gavridis (1997), Volatility Trading Models: An Application to Daily Exchange

Rates. Working Papers in Financial Economics, BNP Global Markets Research 1: 1“4.

Dunis, C., J. Laws and S. Chauvin (2000), FX Volatility Forecasts: A Fusion-Optimisation App-

roach. Neural Network World 10: 187“202.

Dunis, C., J. Moody and A. Timmermann (eds) (2001a), Developments in Forecast Combination

and Portfolio Choice; John Wiley: Chichester.

Dunis, C., J. Laws and S. Chauvin (2001b), The Use of Market Data and Model Combination to

Improve Forecasting Accuracy. In Developments in Forecast Combination and Portfolio Choice,

Dunis, C., Moody, J. and Timmermann, A. (eds); John Wiley: Chichester.

Elman, J. L. (1990), Finding Structures in Time. Cognitive Sciences 14: 179“211.

Engle, R. F. and T. Bollerslev (1986), Modelling the Persistence of Conditional Variances. Econo-

metric Reviews 5: 1“50.

Engle, R. F., C. H. Hong, A. Kane and J. Noh (1993), Arbitrage Valuation of Variance Forecasts

with Simulated Options. In Advances in Futures and Options Research, Chance, D. M. and

Tripp, R. R. (eds); JIA Press: Greenwich, CT.

Fama, E. F. (1990), Term Structure Forecasts of Interest Rates, In¬‚ation and Real Returns. Journal

of Monetary Economics 25: 59“76.

Flˆ res, R. G. Jr and B. B. Roche (2000), Volatility Modelling in the Forex Market: An Empirical

o

Evaluation. In Advances in Quantitative Asset Management, Dunis, C. (ed.); Kluwer: Boston.

Franses, P. H. and P. Van Homelen (1998), On Forecasting Exchange Rates Using Neural Networks.

Applied Financial Economics 8: 589“596.

Garman, M. B. and S. W. Kohlhagen (1983), Foreign Currency Option Values. Journal of Interna-

tional Money and Finance 2: 231“237.

Gen¸ ay, R. (1996), Non-linear Prediction of Security Returns with Moving Average Rules. Journal

c

of Forecasting 15: 165“174.

Granger, C. W. and R. Ramanathan (1984), Improved Methods of Combining Forecasts. Journal

of Forecasting 3: 197“204.

Hassoun, M. H. (1995), Fundamentals of Arti¬cial Neural Networks; MIT Press: Cambridge, MA.

Hornik, K., M. Stinchcombe and H. White (1989), Multilayer Feedforward Networks Are Universal

Approximators. Neural Networks 2: 359“366.

Hsieh, D. A. (1989), Modeling Heteroscedasticity in Daily Foreign Exchange Rates. Journal of

Business and Economic Statistics 7: 307“317.

Hull, J. C. (1997), Options, Futures and Other Derivative Securities; Prentice Hall, Englewood

Cliffs, NJ.

162 Applied Quantitative Methods for Trading and Investment

Ivanova, D., K. Lahiri and F. Seitz (2000), Interest Rate Spreads as Predictors of German In¬‚ation

and Business Cycles. International Journal of Forecasting 16: 39“58.

Jorion, P. (1995), Predicting Volatility in the Foreign Exchange Market. Journal of Finance 50:

507“528.

Jorion, P. (1997), Value at Risk: The New Benchmark for Controlling Market Risk; McGraw-Hill:

New York.

Kouam, A., F. Badran and S. Thiria (1992), Approche M´ thodologique pour l™Etude de la Pr´ vision

e e

a l™Aide de R´ seaux de Neurones, Actes de Neuro-Nˆmes.

e ±

`

Kroner, K. F., K. P. Kneafsey and S. Claessens (1995), Forecasting Volatility in Commodity Mar-

kets. Journal of Forecasting 14: 77“95.

Kuan, C. M. and T. Liu (1995), Forecasting Exchange Rates Using Feedforward and Recurrent

Neural Networks. Journal of Applied Economics 10: 347“364.

Lamoureux, C. G. and W. D. Lastrapes (1993), Forecasting Stock-Return Variances: Toward an

Understanding of Stochastic Implied Volatilities. Review of Financial Studies 6: 293“326.

Latane, H. A. and R. J. Rendleman (1976), Standard Deviations of Stock Price Ratios Implied in

Option Prices. Journal of Finance 31: 369“381.

Laws, J. and A. Gidman (2000), Forecasting Stock Market Volatility and the Application of Volatil-

ity Trading Models. CIBEF Research Paper, Liverpool Business School.

Leung, M. T., H. Daouk and A.-S. Chen (2000), Forecasting Stock Indices: A Comparison of Clas-

si¬cation and Level Estimation Models. International Journal of Forecasting 16: 173“190.

Makridakis, S., S. C. Wheelwright and V. E. McGee (1983), Forecasting: Methods and Applica-

tions; John Wiley: New York.

Malz, A. M. (1996), Using Option Prices to Estimate Realignment Probabilities in the European

Monetary System. In Forecasting Financial Markets, Dunis, C. (ed.); John Wiley: Chichester.

Ntungo, C. and M. Boyd (1998), Commodity Futures Trading Performance Using Neural Network

Models Versus ARIMA Models. Journal of Futures Markets 8: 965“983.

Pindyck, R. S. and D. L. Rubinfeld (1998), Econometric Models and Economic Forecasts; McGraw-

Hill: New York.

Refenes, A. N. (1995), Neural Networks In the Capital Markets; John Wiley: Chichester.

¨

Rehkugler, H. and H. G. Zimmermann (eds) (1994), Neuronale Netze in der Okonomie “ Grund-

lagen und ¬nanzwirtschaftliche Anwendungen; Verlag Franz Vahlen: M¨ nchen.

u

Shadbolt, J. and J. G. Taylor (eds) (2002), Neural Networks and the Financial Markets “ Predicting,

Combining and Portfolio Optimisation; Springer-Verlag: London.

Simpson, P. K. (1990), Arti¬cial Neural Systems“Foundations, Paradigms, Applications, and Imple-

mentations; Pergamon Press: New York.

So, M. K. P., K. Lam and W. K. Li (1999), Forecasting Exchange Rate Volatility Using Autore-

gressive Random Variance Model. Applied Financial Economics 9: 583“591.

Swanson, N. R. and H. White (1995), A Model-Selection Approach to Assessing the Information in

the Term-Structure Using Linear Models and Neural Networks. Journal of Business and Economic

Statistics 13: 265“275.

Taylor, S. J. (1986), Modelling Financial Time Series; John Wiley: Chichester.

Tenti, P. (1996), Forecasting Foreign Exchange Rates Using Recurrent Neural Networks. Applied

Arti¬cial Intelligence 10: 567“581.

Theil, H. (1966), Applied Economic Forecasting; North-Holland: Amsterdam.

Trippi, R. and E. Turban (eds) (1993), Neural Networks in Finance and Investing “ Using Arti¬cial

Intelligence to Improve Real-World Performance; Probus: Chicago.

West, K. D. and D. Cho (1995), The Predictive Ability of Several Models of Exchange Rate Volatil-

ity. Journal of Econometrics 69: 367“391.

Xu, X. and S. J. Taylor (1996), Conditional Volatility and the Informational Ef¬ciency of the

PHLX Currency Options Market. In Forecasting Financial Markets, Dunis, C. (ed.); John Wiley:

Chichester.

Zhang, G., B. E. Patuwo and M. Y. Hu (1998), Forecasting with Arti¬cial Neural Networks: The

State of the Art. International Journal of Forecasting 14: 35“62.

Analysis of Financial Time Series. Ruey S. Tsay

Copyright ™ 2002 John Wiley & Sons, Inc.

ISBN: 0-471-41544-8

CHAPTER 5

High-Frequency Data Analysis

and Market Microstructure

High-frequency data are observations taken at ¬ne time intervals. In ¬nance, they

often mean observations taken daily or at a ¬ner time scale. These data have become

available primarily due to advances in data acquisition and processing techniques,

and they have attracted much attention because they are important in empirical

study of market microstructure. The ultimate high-frequency data in ¬nance are the

transaction-by-transaction or trade-by-trade data in security markets. Here time is

often measured in seconds. The Trades and Quotes (TAQ) database of the New York

Stock Exchange (NYSE) contains all equity transactions reported on the Consoli-

dated Tape from 1992 to present, which includes transactions on NYSE, AMEX,

NASDAQ, and the regional exchanges. The Berkeley Options Data Base provides

similar data for options transactions from August 1976 to December 1996. Trans-

actions data for many other securities and markets, both domestic and foreign,

are continuously collected and processed. Wood (2000) provides some historical

perspective of high-frequency ¬nancial study.

High-frequency ¬nancial data are important in studying a variety of issues related

to trading process and market microstructure. They can be used to compare the ef¬-

ciency of different trading systems in price discovery (e.g., the open out-cry system

of NYSE and the computer trading system of NASDAQ). They can also be used to

study the dynamics of bid and ask quotes of a particular stock (e.g., Hasbrouck, 1999;

Zhang, Russell, and Tsay, 2001b). In an order-driven stock market (e.g., the Taiwan

Stock Exchange), high-frequency data can be used to study the order dynamic and,

more interesting, to investigate the question “who provides the market liquidity.”

Cho, Russell, Tiao, and Tsay (2000) use intraday 5-minute returns of more than 340

stocks traded in the Taiwan Stock Exchange to study the impact of daily stock price

limits and ¬nd signi¬cant evidence of magnet effects toward the price ceiling.

However, high-frequency data have some unique characteristics that do not appear

in lower frequencies. Analysis of these data thus introduces new challenges to ¬nan-

cial economists and statisticians. In this chapter, we study these special characteris-

tics, consider methods for analyzing high-frequency data, and discuss implications

175

176 HIGH-FREQUENCY DATA

of the results obtained. In particular, we discuss nonsynchronous trading, bid-ask

spread, duration models, price movements that are in multiples of tick size, and

bivariate models for price changes and time durations between transactions associ-

ated with price changes. The models discussed are also applicable to other scienti¬c

areas such as telecommunications and environmental studies.

5.1 NONSYNCHRONOUS TRADING

We begin with nonsynchronous trading. Stock tradings such as those on the NYSE

do not occur in a synchronous manner; different stocks have different trading fre-

quencies, and even for a single stock the trading intensity varies from hour to hour

and from day to day. Yet we often analyze a return series in a ¬xed time interval such

as daily, weekly, or monthly. For daily series, price of a stock is its closing price,

which is the last transaction price of the stock in a trading day. The actual time of the

last transaction of the stock varies from day to day. As such we incorrectly assume

daily returns as an equally-spaced time series with a 24-hour interval. It turns out

that such an assumption can lead to erroneous conclusions about the predictability

of stock returns even if the true return series are serially independent.

For daily stock returns, nonsynchronous trading can introduce (a) lag-1 cross-

correlation between stock returns, (b) lag-1 serial correlation in a portfolio return,

and (c) in some situations negative serial correlations of the return series of a single

stock. Consider stocks A and B. Assume that the two stocks are independent and

stock A is traded more frequently than stock B. For special news affecting the market

that arrives near the closing hour on one day, stock A is more likely than B to show

the effect of the news on the same day simply because A is traded more frequently.

The effect of the news on B will eventually appear, but it may be delayed until the

following trading day. If this situation indeed happens, return of stock A appears

to lead that of stock B. Consequently, the return series may show a signi¬cant lag-

1 cross-correlation from A to B even though the two stocks are independent. For

a portfolio that holds stocks A and B, the prior cross-correlation would become a

signi¬cant lag-1 serial correlation.

In a more complicated manner, nonsynchronous trading can also induce erroneous

negative serial correlations for a single stock. There are several models available in

the literature to study this phenomenon; see Campbell, Lo, and MacKinlay (1997)

and the references therein. Here we adopt a simpli¬ed version of the model pro-

posed in Lo and MacKinlay (1990). Let rt be the continuously compounded return

of a security at the time index t. For simplicity, assume that {rt } is a sequence of

independent and identically distributed random variables with mean E(rt ) = µ and

variance Var(rt ) = σ 2 . For each time period, the probability that the security is not

traded is π, which is time-invariant and independent of rt . Let rto be the observed

return. When there is no trade at time index t, we have rto = 0 because there is

no information available. Yet when there is a trade at time index t, we de¬ne rto as

the cumulative return from the previous trade (i.e., rto = rt + rt’1 + · · · + rt’kt ,

where kt is the largest non-negative integer such that no trade occurred in the periods

177

NONSYNCHRONOUS TRADING

t ’ kt , t ’ kt + 1, . . . , t ’ 1). Mathematically, the relationship between rt and rto is

±

with probability π

0

r

t with probability (1 ’ π)2

r + r

t with probability (1 ’ π)2 π

t’1

r + r

t’1 + rt’2 with probability (1 ’ π)2 π 2

t

rto = . (5.1)

.

. .

. .

k

i=0 rt’i with probability (1 ’ π)2 π k’1

. .

. .

. .

These probabilities are easy to understand. For example, rto = rt if and only if there

are trades at both t and t ’ 1, rto = rt + rt’1 if and only if there are trades at t and

t ’ 2, but no trade at t ’ 1, and rto = rt + rt’1 + rt’2 if and only if there are trades

at t and t ’ 3, but no trades at t ’ 1 and t ’ 2, and so on. As expected, the total

probability is 1 given by

1

π + (1 ’ π)2 [1 + π + π 2 + · · ·] = π + (1 ’ π)2 = π + 1 ’ π = 1.

1’π

We are ready to consider the moment equations of the observed return series {rto }.

First, the expectation of rto is

E(rto ) = (1 ’ π)2 E(rt ) + (1 ’ π)2 π E(rt + rt’1 ) + · · ·

= (1 ’ π)2 µ + (1 ’ π)2 π2µ + (1 ’ π)2 π 2 3µ + · · ·

= (1 ’ π)2 µ[1 + 2π + 3π 2 + 4π 3 + · · ·]

1

= (1 ’ π)2 µ = µ. (5.2)

(1 ’ π)2

In the prior derivation, we use the result 1 + 2π + 3π 2 + 4π 3 + · · · = 1

. Next,

(1’π)2

for the variance of rto , we use Var(rto ) = E[(rto )2 ] ’ [E(rto )]2 and

E(rto )2 = (1 ’ π)2 E[(rt )2 ] + (1 ’ π)2 π E[(rt + rt’1 )2 ] + · · ·

= (1 ’ π)2 [(σ 2 + µ2 ) + π(2σ 2 + 4µ2 ) + π 2 (3σ 2 + 9µ2 ) + ·] (5.3)

= (1 ’ π)2 {σ 2 [1 + 2π + 3π 2 + · · ·] + µ2 [1 + 4π + 9π 2 + · · ·]} (5.4)

2

= σ 2 + µ2 ’1 . (5.5)

1’π

In Eq. (5.3), we use

2 2

k k k

= Var +E = (k + 1)σ 2 + [(k + 1)µ]2

E rt’i rt’i rt’i

i=0 i=0 i=0

178 HIGH-FREQUENCY DATA

under the serial independence assumption of rt . Using techniques similar to that of

Eq. (5.2), we can show that the ¬rst term of Eq. (5.4) reduces to σ 2 . For the second

term of Eq. (5.4), we use the identity

2 1

1 + 4π + 9π 2 + 16π 3 + · · · = ’ ,

(1 ’ π)3 (1 ’ π)2

which can be obtained as follows: Let

H = 1 + 4π + 9π 2 + 16π 3 + · · · G = 1 + 3π + 5π 2 + 7π 3 + · · · .

and

Then (1 ’ π)H = G and

(1 ’ π)G = 1 + 2π + 2π 2 + 2π 3 + · · ·

2

= 2(1 + π + π 2 + · · ·) ’ 1 = ’ 1.

(1 ’ π)

Consequently, from Eqs. (5.2) and (5.5), we have

2πµ2

2

Var(rto ) =σ +µ ’1 ’µ =σ + .

2 2 2 2

(5.6)

1’π 1’π

Consider next the lag-1 autocovariance of {rto }. Here we use Cov(rto , rt’1 ) = o

E(rtort’1 ) ’ E(rt0 )E(rt’1 ) = E(rtort’1 ) ’ µ2 . The question then reduces to ¬nding

o o o

E(rtort’1 ). Notice that rto rt’1 is zero if there is no trade at t, no trade at t ’ 1, or no

o o

trade at both t and t ’ 1. Therefore, we have

±

with probability 2π ’ π 2

0

r r

t t’1 with probability (1 ’ π)3

r (r

t t’1 + rt’2 ) with probability (1 ’ π)3 π

r (r

t t’1 + rt’2 + rt’3 ) with probability (1 ’ π) π

32

rt rt’1 = .

oo

(5.7)

.

. .

. .

r (

t k

i=1 rt’i ) with probability (1 ’ π)3 π k’1

. .

. .

. .

Again the total probability is unity. To understand the prior result, notice that

rto rt’1 = rt rt’1 if and only if there are three consecutive trades at t ’ 2, t ’ 1,

o

and t. Using Eq. (5.7) and the fact that E(rt rt’ j ) = E(rt )E(rt’ j ) = µ2 for j > 0,

we have

E(rtort’1 ) = (1 ’ π)3 {E(rt rt’1 ) + π E[rt (rt’1 + rt’2 )]

o

3

+ π E rt + · · ·}

2

rt’i

i=1

= (1 ’ π)3 µ2 [1 + 2π + 3π 2 + · · ·] = (1 ’ π)µ2 .

179

BID-ASK SPREAD

The lag-1 autocovariance of {rto } is then

Cov(rto , rt’1 ) = ’πµ2 .

o

(5.8)

Provided that µ is not zero, the nonsynchronous trading induces a negative lag-1

autocorrelation in rto given by

’(1 ’ π)πµ2

ρ1 (rto ) = .

(1 ’ π)σ 2 + 2πµ2

In general, we can extend the prior result and show that

Cov(rto , rt’ j ) = ’µ2 π j , j ≥ 1.

o

The magnitude of the lag-1 ACF depends on the choices of µ, π, and σ and can

be substantial. Thus, when µ = 0, the nonsynchronous trading induces negative

autocorrelations in an observed security return series.

The previous discussion can be generalized to the return series of a portfolio that

consists of N securities; see Campbell, Lo, and MacKinlay (1997, Chapter 3). In

the time series literature, effects of nonsynchronous trading on the return of a single

security are equivalent to that of random temporal aggregation on a time series, with

the trading probability π governing the mechanism of aggregation.

5.2 BID-ASK SPREAD

In some stock exchanges (e.g., NYSE) market makers play an important role in facil-

itating trades. They provide market liquidity by standing ready to buy or sell when-

ever the public wishes to sell or buy. By market liquidity, we mean the ability to buy

or sell signi¬cant quantities of a security quickly, anonymously, and with little price

impact. In return for providing liquidity, market makers are granted monopoly rights

by the exchange to post different prices for purchases and sales of a security. They

buy at the bid price Pb and sell at a higher ask price Pa . (For the public, Pb is the

sale price and Pa is the purchase price.) The difference Pa ’ Pb is call the bid-ask

spread, which is the primary source of compensation for market makers. Typically,

the bid-ask spread is small”namely, one or two ticks.

The existence of bid-ask spread, although small in magnitude, has several impor-

tant consequences in time series properties of asset returns. We brie¬‚y discuss the

bid-ask bounce”namely, the bid-ask spread introduces negative lag-1 serial corre-

lation in an asset return. Consider the simple model of Roll (1984). The observed

market price Pt of an asset is assumed to satisfy

S

Pt = Pt— + It , (5.9)

2

180 HIGH-FREQUENCY DATA

where S = Pa ’ Pb is the bid-ask spread, Pt— is the time-t fundamental value of the

asset in a frictionless market, and {It } is a sequence of independent binary random

variables with equal probabilities (i.e., It = 1 with probability 0.5 and = ’1 with

probability 0.5). The It can be interpreted as an order-type indicator, with 1 signify-

ing buyer-initiated transaction and ’1 seller-initiated transaction. Alternatively, the

model can be written as

+S/2 with probability 0.5,

Pt = Pt— +

’S/2 with probability 0.5.

If there is no change in Pt— , then the observed process of price changes is

S

Pt = (It ’ It’1 ) . (5.10)

2

Under the assumption of It in Eq. (5.9), E(It ) = 0 and Var(It ) = 1, and we have

E( Pt ) = 0 and

Var( Pt ) = S 2 /2 (5.11)

Cov( Pt , Pt’1 ) = ’S 2 /4 (5.12)

Cov( Pt , Pt’ j ) = 0, j > 1. (5.13)

Therefore, the autocorrelation function of Pt is

’0.5 if j = 1,

ρ j ( Pt ) = (5.14)

if j > 1.

0

The bid-ask spread thus introduces a negative lag-1 serial correlation in the series

of observed price changes. This is referred to as the bid-ask bounce in the ¬nance

literature. Intuitively, the bounce can be seen as follows. Assume that the fundamen-

tal price Pt— is equal to (Pa + Pb )/2. Then Pt assumes the value Pa or Pb . If the

previously observed price is Pa (the higher value), then the current observed price is

either unchanged or lower at Pb . Thus, Pt is either 0 or ’S. However, if the pre-

vious observed price is Pb (the lower value), then Pt is either 0 or S. The negative

lag-1 correlation in Pt becomes apparent. The bid-ask spread does not introduce

any serial correlation beyond lag 1, however.

A more realistic formulation is to assume that Pt— follows a random walk so that

—

Pt— = Pt— ’ Pt’1 = t , which forms a sequence of independent and identically

distributed random variables with mean zero and variance σ 2 . In addition, { t } is

independent of {It }. In this case, Var( Pt ) = σ 2 + S 2 /2, but Cov( Pt , Pt’ j )

remains unchanged. Therefore,

’S 2 /4

ρ1 ( Pt ) = ¤ 0.

S 2 /2 + σ 2

181

EMPIRICAL CHARACTERISTICS

The magnitude of lag-1 autocorrelation of Pt is reduced, but the negative effect

remains when S = Pa ’ Pb > 0. In ¬nance, it might be of interest to study the

components of the bid-ask spread. Interested readers are referred to Campbell, Lo,

and MacKinlay (1997) and the references therein.

The effect of bid-ask spread continues to exist in portfolio returns and in multivari-

ate ¬nancial time series. Consider the bivariate case. Denote the bivariate order-type

indicator by It = (I1t , I2t ) , where I1t is for the ¬rst security and I2t for the second

security. If I1t and I2t are contemporaneously correlated, then the bid-ask spreads

can introduce negative lag-1 cross-correlations.

5.3 EMPIRICAL CHARACTERISTICS OF TRANSACTIONS DATA

Let ti be the calendar time, measured in seconds from midnight, at which the i-th

transaction of an asset takes place. Associated with the transaction are several vari-

ables such as the transaction price, the transaction volume, the prevailing bid and ask

quotes, and so on. The collection of ti and the associated measurements are referred

to as the transactions data. These data have several important characteristics that do

not exist when the observations are aggregated over time. Some of the characteristics

are given next.

1. Unequally spaced time intervals: Transactions such as stock tradings on an

exchange do not occur at equally spaced time intervals. As such the observed

transaction prices of an asset do not form an equally spaced time series. The

time duration between trades becomes important and might contain useful

information about market microstructure (e.g., trading intensity).

2. Discrete-valued prices: The price change of an asset from one transaction to

the next only occurs in multiples of tick size. In the NYSE, the tick size was

one eighth of a dollar before June 24, 1997, and was one sixteenth of a dollar

before January 29, 2001. All NYSE and AMEX stocks started to trade in dec-

imals on January 29, 2001. Therefore, the price is a discrete-valued variable in

transactions data. In some markets, price change may also be subject to limit

constraints set by regulators.

3. Existence of a daily periodic or diurnal pattern: Under the normal trading con-

ditions, transaction activity can exhibit periodic pattern. For instance, in the

NYSE, transactions are heavier at the beginning and closing of the trading

hours and thinner during the lunch hours, resulting in a “U-shape” transac-

tion intensity. Consequently, time durations between transactions also exhibit

a daily cyclical pattern.

4. Multiple transactions within a single second: It is possible that multiple trans-

actions, even with different prices, occur at the same time. This is partly due

to the fact that time is measured in seconds that may be too long a time scale

in periods of heavy tradings.

182 HIGH-FREQUENCY DATA

Table 5.1. Frequencies of Price Change in Multiples of Tick Size for IBM Stock from

November 1, 1990 to January 31, 1991.

¤ ’3 ’2 ’1 ≥3

Number (tick) 0 1 2

Percentage 0.66 1.33 14.53 67.06 14.53 1.27 0.63

To demonstrate these characteristics, we consider ¬rst the IBM transactions data

from November 1, 1990 to January 31, 1991. These data are from the Trades, Orders

Reports, and Quotes (TORQ) dataset; see Hasbrouck (1992). There are 63 trading

days and 60,328 transactions. To simplify the discussion, we ignore the price changes

between trading days and focus on the transactions that occurred in the normal trad-

ing hours from 9:30 am to 4:00 pm Eastern Time. It is well known that overnight

stock returns differ substantially from intraday returns; see Stoll and Whaley (1990)

and the references therein. Table 5.1 gives the frequencies in percentages of price

change measured in the tick size of $1/8 = $0.125. From the table, we make the

following observations:

1. About two-thirds of the intraday transactions were without price change.

2. The price changed in one tick approximately 29% of the intraday transactions.

3. Only 2.6% of the transactions were associated with two-tick price changes.

4. Only about 1.3% of the transactions resulted in price changes of three ticks or