Algoritmos genéticos y modelos multivariados recursivos en la predicción de índices bursátiles de América del Norte: IPC, TSE, Nasdaq y DJI

2004. Trimestre Económico. Vol. 71, Nº 284, Pp 789 - 810

Franco Parisi F, Antonino Parisi F, Edinson Cornejo S

Abstract:

Using weekly stock index prices, corresponding to the period between April 07 of 1998 and April 14 of 2003, we analyzed the efficiency of the dynamic multivaried models, from recursives genetic algorithms, to forecast the weekly sign variations of stock-exchange indices IPC, TSE, Nasdaq and DJI. The results were compared with those of a multivaried model AR(1) and a random model. The best models produced by the genetic algorithm threw a sign prediction percentage (SPP) of 59%, 60%, 59% and 59%, for indices IPC, Nasdaq, TSE and DJI, respectively. The forecast capacity was significant in each index, according to Pesaran & Timmerman’s directional accuracy test (1992). When analyzing the SPP of the models AR(1), were smaller, being significant only in the case of the Nasdaq. The random dynamic multivaried models presented the lowest SPP (except in index TSE), being significant only in the case of Nasdaq. In addition, the models constructed by the genetic algorithm generated the greater accumulated return, except in the case of the Nasdaq, where the highest yield was registered by the model AR(1). In the test of robustness through the analysis of 1 000 bootstrap series, in average, the SPP was of 50.88%, 52.58%, 49.07%, 52.93%, for indices DJI, IPC, Nasdaq and TSE. The multivaried models surpassed the return of a buy and hold strategy in 57%, 59% and 71%, DJI, IPC and TSE, respectively.

Palabras claves: algoritmos genéticos, modelo multivariado dinámico, funcionamiento recursivo, porcentaje de predicción de signo, prueba de acierto direccional. 

Acceder a la publicación completa