Volatility Forecasting Using Support Vector Regression and a Hybrid Genetic Algorithm

Volatility forecasting is an important process required to measure variability in equity prices, risk management, and several other financial activities. Generalized autoregressive conditional heteroscedastic methods $$(\textit{GARCH})$$ ( GARCH ) have been used to forecast volatility with reasonable success due unreal assumptions about volatility underlying process. Recently, a supervised learning machine called support vector regression $$(SVR)$$ ( S V R ) has been employed to forecast financial volatility. Nevertheless, the quality and stability of the model obtained through $$SVR$$ S V R training process depend strongly on the selection of $$SVR$$ S V R parameters. Typically, these are tuned by a grid search method $$(SVR_{GS})$$ ( S V R G S ) ; however, this tuning procedure is prone to get trapped on local optima, requires a priori information, and it does not concurrently tune the kernels and its parameters. This paper presents a new method called $$SVR_{GBC}$$ S V R G B C for the financial volatility forecasting problem which selects simultaneously the proper kernel and its parameter values. $$SVR_{GBC}$$ S V R G B C is a hybrid genetic algorithm which uses several genetic operators to enhance the exploration of solutions space: it introduces a new genetic operator called Boltzmann selection, and the use of several random number generators. Experimental data correspond to two ASEAN and two latinoamerican market indexes. $$SVR_{GBC}$$ S V R G B C results are compared against $$\textit{GARCH}\left( 1,1\right) \hbox { and }SVR_{GS}$$ GARCH 1 , 1 and S V R G S method. It uses the mean absolute percentage error and directional accuracy functions for measuring quality results. Experimentation shows that, in general, $$SVR_{GBC}$$ S V R G B C overcomes quality of $$\textit{GARCH}\left( 1,1\right) \hbox { and }SVR_{GS}$$ GARCH 1 , 1 and S V R G S . Copyright Springer Science+Business Media New York 2015