INSTITUTO NACIONAL DE ASTROFÍSICA, ÓPTICA Y eLECTRÓNICA

Non-linear time series approximation and forecasting, based on neural systems and multi-resolution analysis

This project has to do with the use of artificial neural networks and signal processing algorithms, for the estimation of future values in time series, which are generated by highly non-linear systems. Our estimations take place considering only past values of a time series; this is useful when no exogenous variables are available.

Time series forecasting is an important research topic, due to the high number of applications requiring such task. However, this is a complex problem, particularly when the involved time series are chaotic or highly non-linear.

Forecasting implies the modeling of an approximation function, that is:

 

 

 

where X represents the time series. There are several types of prediction; recurrent prediction involves de use of previously calculated estimation for the estimation of new values:

 

Our research work has been working with research prediction for several years. We have built several prediction models, which are detailed in several publications.

Last modified on April 10, 2015