Research On Model Selection Method For Statistical Time Series Prediction Algorithms

In recent years, time series prediction has been applied widely in various fields. Generally speaking, any prediction algorithm cannot solve prediction problems for all kinds of time series. That is, every algorithm can only solve prediction problems reasonably for data sets with certain features. So, selecting a reasonable model becomes the most significant step for obtaining reliable prediction information. This paper conducts study on optimal prediction model selection mechanism.Firstly, this paper establishes feature description system to analyze four kinds of data features including continuity of amplitude, long memory, seasonal and trend. Then, classification system which contains nine categories is proposed. After that, we conduct study on the features of statistical prediction algorithms and analyze theoretically the data features and categories each algorithm fits for. Then, establish the mapping relationship between models and time series using experimental results of public data sets. The above research provides foundation of qualitative model selection mechanism.Secondly, quantitative metrics system to evaluate the applicability of prediction models which contains six categories of metrics and can be customized is proposed. Then, the quantitative model selection mechanism is completed. And then, the model selection mechanism which contains qualitative and quantitative selection steps is proposed. The qualitative selection classifies the time series into certain category through analyzing its features and takes the models obtained from mapping system as candidate models. And the quantitative selection customizes the evaluation system in view of current prediction scenario and compares the applicability of the candidate models to select the optimal model. In addition to this, considering probable feature changes in practical applications, a decision-making advice feedback mechanism is proposed. This mechanism provides interactive advice according to the change scope of data feature to maintain accuracy and reliabilityof the result.Finally, a kind of time series prediction software is designed to meet the actual application demands. It uses a mixed programming method which embeds R process into Python. Python is used to eatablish the architecture and design the interface and R is used to achieve the function of data processing, analysis and the algorithms. The software has the ability to analyze data features, select optimal model, predict time series and evaluate the results and it is tested and verified using public data sets.Experimental results refer the proposed classification system, mapping relationship between models and time series categories, model selection mechanism have favourable applicability. That is, reasonable model selection and update can be realized.