on forecasting exchange rates using neural networks: p.h. franses and p.v. homelen, 1998, applied...

2
Research on Forecasting 139 Saddle River, NJ: Prentice-Hall. licative in nature. However, even though most Caulkins, J., Choen, J., Gorr, W., & Wei, J. (1996). measures of gang membership and illicit drug Predicting criminal recidivism: A comparison of neural use are very crude indicators of the actual network models with statistical procedures. Journal of dynamics involved (see our illustrative discus- Criminal Justice 24, 227–240. sion of gang membership), and then these Gottfredson, D. M., & Tonry, M. (1987). (Eds.) Prediction and classification: Criminal justice decision making. measures are processed through mathematical Chicago: University of Chicago Press. manipulations which do not reflect reality, these Harland, A. T. (Ed.) (1996). Choosing correctional options procedures still provide more accurate infor- that work. Thousand Oaks, CA: Sage. mation in many instances than any information Sherman, L. W., Gottfredson, D. C., MacKenzie, D., Eck, process of which I am aware. Therefore, with- J., Reuter, P., & Bushway, S. (1997). Preventing crime: out practical examples of types of information What works, what doesnt, what s promising. Washing- ton, DC: U. S. Department of Justice, Office of Justice processing mechanisms, I cannot respond mean- Programs. ingfully to the final comments by Dr. Caulkins. If he is pointing out that, for instance, admission PII: S0169-2070(00)00049-2 to a graduate program should be decided by information in addition to GRE scores or GMAT scores, or that the policies for admission should be made based on information which On Forecasting Exchange Rates Using Neural may be qualitative, I have no disagreement. It Networks, P.H. Franses and P.V. Homelen, 1998, was not our intent to say that predictive efficacy Applied Financial Economics, 8, 589–596 was the sole or even the best criteria for valuing information. Albeit we did not necessarily as- Artificial Neural Networks (ANN) have been sume that because existing data provided used for forecasting for a number of years. mediocre prediction, it is prima facie evidence Zhang, Patuwo and Hu (1998) provide a survey that the data are not very useful for any other of the state-of-the-art of ANNs for time series process, I have no problem in making that forecasting. Every study referenced therein gen- statement. If agencies cannot devise useful erates point forecasts as a nonlinear function of information processes for their central purpose, historical observations. it is my experience that they also do not know An interesting question that has received very useful ways of using information. little attention is whether or not neural networks can account for changes in variance, or more Brent B. Benda, Ph. D. specifically, account for GARCH structure. The School of Social Work goal of this paper is 1) to determine whether or University of Arkansas at Little Rock not neglected GARCH leads to spuriously suc- Little Rock, Arkansas 72204 cessful ANNs and 2) to confirm that neural USA networks can indeed account for nonlinearity in E-mail: [email protected] the level of a time series. The authors use simulation and four real exchange rate series to make their conclusions. Sign of the fitted return rather than actual level is used to assess fore- References casting accuracy. Motivated by the numerous studies finding Andrews, D. A., & Bonta, J. (1994). The psychology of that daily exchange rate series display GARCH criminal conduct. Cincinnati, OH: Anderson. properties, the authors simulate a GARCH Bartol, C. R., & Bartol, A. M. (1998). Delinquency and justice: A psychosocial approach. (2nd ed.) Upper series and generate forecasts from various neur-

Upload: sandy-balkin

Post on 16-Sep-2016

213 views

Category:

Documents


0 download

TRANSCRIPT

Research on Forecasting 139

Saddle River, NJ: Prentice-Hall.licative in nature. However, even though mostCaulkins, J., Choen, J., Gorr, W., & Wei, J. (1996).measures of gang membership and illicit drug

Predicting criminal recidivism: A comparison of neuraluse are very crude indicators of the actualnetwork models with statistical procedures. Journal of

dynamics involved (see our illustrative discus- Criminal Justice 24, 227–240.sion of gang membership), and then these Gottfredson, D. M., & Tonry, M. (1987). (Eds.) Prediction

and classification: Criminal justice decision making.measures are processed through mathematicalChicago: University of Chicago Press.manipulations which do not reflect reality, these

Harland, A. T. (Ed.) (1996). Choosing correctional optionsprocedures still provide more accurate infor-that work. Thousand Oaks, CA: Sage.

mation in many instances than any information Sherman, L. W., Gottfredson, D. C., MacKenzie, D., Eck,process of which I am aware. Therefore, with- J., Reuter, P., & Bushway, S. (1997). Preventing crime:out practical examples of types of information What works, what doesn’t, what’s promising. Washing-

ton, DC: U. S. Department of Justice, Office of Justiceprocessing mechanisms, I cannot respond mean-Programs.ingfully to the final comments by Dr. Caulkins.

If he is pointing out that, for instance, admissionPII : S0169-2070( 00 )00049-2

to a graduate program should be decided byinformation in addition to GRE scores orGMAT scores, or that the policies for admissionshould be made based on information which On Forecasting Exchange Rates Using Neuralmay be qualitative, I have no disagreement. It Networks, P.H. Franses and P.V. Homelen, 1998,was not our intent to say that predictive efficacy Applied Financial Economics, 8, 589–596was the sole or even the best criteria for valuinginformation. Albeit we did not necessarily as- Artificial Neural Networks (ANN) have beensume that because existing data provided used for forecasting for a number of years.mediocre prediction, it is prima facie evidence Zhang, Patuwo and Hu (1998) provide a surveythat the data are not very useful for any other of the state-of-the-art of ANNs for time seriesprocess, I have no problem in making that forecasting. Every study referenced therein gen-statement. If agencies cannot devise useful erates point forecasts as a nonlinear function ofinformation processes for their central purpose, historical observations.it is my experience that they also do not know An interesting question that has received veryuseful ways of using information. little attention is whether or not neural networks

can account for changes in variance, or moreBrent B. Benda, Ph. D. specifically, account for GARCH structure. TheSchool of Social Work goal of this paper is 1) to determine whether or

University of Arkansas at Little Rock not neglected GARCH leads to spuriously suc-Little Rock, Arkansas 72204 cessful ANNs and 2) to confirm that neural

USA networks can indeed account for nonlinearity inE-mail: [email protected] the level of a time series. The authors use

simulation and four real exchange rate series tomake their conclusions. Sign of the fitted returnrather than actual level is used to assess fore-

References casting accuracy.Motivated by the numerous studies finding

Andrews, D. A., & Bonta, J. (1994). The psychology ofthat daily exchange rate series display GARCHcriminal conduct. Cincinnati, OH: Anderson.properties, the authors simulate a GARCHBartol, C. R., & Bartol, A. M. (1998). Delinquency and

justice: A psychosocial approach. (2nd ed.) Upper series and generate forecasts from various neur-

140 Research on Forecasting

al network architectures. They find that ANN Sandy Balkinmodels generally do not yield significantly more Ernst & Young LLPaccurate forecasts. A second set of experiments 1225 Connecticut Avenue, NWperformed on simulated bilinear series ended Washington, DC 20036with considerably better results. Empirical re- USAsults point toward the same conclusions, namelywhen the nonlinearity in a series is due toGARCH, no gain in forecasting accuracy is

Referencesobtained by using a neural network. However, ifthe nonlinearity is not due to GARCH, then a

Zhang, G., Patuwo, B., & Hu, M. (1998). Forecasting withneural network will indeed exploit the extraArtificial Neural Networks: The State of the Art.structure and produce more accurate forecasts.International Journal of Forecasting 14(1), 35–62.

While one should be careful about makinggeneralizations concerning a method based on a PII : S0169-2070( 00 )00047-9few examples, the work behind this paper isvery well done and leaves the reader with a hostof possible future directions for research.