In this letter, time series forecasting using optimal neural networks with optimal performance of generalization and stability is studied. After analyzing the reasons for overfitting and instability of neural networks, in order to find the optimal NNs (neural networks) architecture, we consider minimizing three objective index: average training AICc (ATRAICc), average testing AICc (ATEAICc) and testing error variance AICc (VAAICc) based on Akaike information criterion theory. Then we built a multi-objective optimization model and proved the existence and uniqueness theorem of optimal solution. After determining the searching interval, a multi-objective optimization algorithm for optimal neural network architecture based on AICc (ONNAICc) with optimal generalization and stability is constructed to solve above model. Some experiments with simple sample datasets A and B from simple function, complicated stock market time series datasets C and highly nonlinear tungsten price datasets D are taken to verify the validity of the model. And some performances of the presented ONNAICc algorithm are compared with the traditional prediction algorithms, such as AR, ARMA, ordinary BP, SVM, verified the superiority and accuracy of the proposed ONNAICc algorithm. The technical details, proofs and evaluations can be found in the support information.
[1]
N. Sugiura.
Further analysts of the data by akaike' s information criterion and the finite corrections
,
1978
.
[2]
Yun Zhang,et al.
Design of ensemble neural network using the Akaike information criterion
,
2008,
Eng. Appl. Artif. Intell..
[3]
Luiz Koodi Hotta,et al.
A note on influence diagnostics in AR(1) time series models
,
2012
.
[4]
Moon Ho Lee,et al.
A new constructive neural network method for noise processing and its application on stock market prediction
,
2014,
Appl. Soft Comput..
[5]
Alejandra Cabaña,et al.
Weak Convergence of Marked Empirical Processes for Focused Inference on AR(p) vs AR(p + 1) Stationary Time Series
,
2012
.