The Study On Non-linear Economic Time Series Forecasting Based On The Policy Intervention And Error Correction

Economic forecast, one kind of statistical forecast, uses economic phenomena asprediction objects, directly or indirectly, provide information for micro or macromarket decision, management decision and policy making. Forecast which for thepurpose of provides objective basis for decision making or plan is not only inengineering field, even more for the social and economic fields. Economic time seriesprediction has vital significance for the government’s economic policy, enterprise orindividual investment activities, however, the internal rules of complex and massivedata processing makes the traditional prediction method fails to provide satisfactoryresults, therefore new prediction technology has been one of the important directionof time series research. The forecast of economic time series as the essential decisionsupport of management decision-making, market, policy and so on is becoming moreand more important. The performance of rapid social and economic changing in frontof people are more and more in the form of data, moreover, wider use of maturedatabase technology and computer has brought people accumulate data volume toincrease exponentially, therefore, how to seek the necessary information in thecomplicated data to support making predictions become one of the fundamental wayfor decision-makers.Actually, when forecasting, because of the different modeling mechanism,usually for the sameproblem there have different modeling methods. A singleforecasting model because of the limitation of prediction principle and operating rulesis hardly to effectively reflect the changing rule of the predictor variable, so thecombination forecast method become the inevitable developing trend. This papermainly studies two dilemmas “prediction error handling” and “the impact of policyevents”, analysis of the reasons, and puts forward the solution. The main content ofthis study are as follows:1. A double prediction method by means of synchronous prediction of theprediction error was proposed based on the existing support vector machine (SVM)method, and the predicted error was used to correct the preliminary predicted valuesin order to improve the prediction accuracy. Considering that the error sequence mayhave features of non-stationarity, nonlinearity system and insufficient information ofsystem dynamics, the empirical mode decomposition (EMD) method was used andembeds into the support vector machine method to predict the error series according to the preliminary training error and test error respectively. In order to get the finalprediction error, different parameters were chosen to forecast the error sequencedecomposed into several intrinsic mode functions (IMFs) components according to2. This paper put forward an improved method by combined EMD and STSAtogether direct at the defect of STSA method in financial time series analysis. Takingthe return data of6different stock indexes as the research sample and using EMDdecomposition method extracted a series of signal components which reflect differenttime scale information of the original sequence. By using STSA to each component,the reason that lead to the diverse changing pattern of the original sequence has beenfounded. Based on this, a method was put forward that using the single changingpattern component to estimate the original sequence, the condition and limited rangeare given as well. The experimental results show that the proposed approach hasunique superiority in extract the information of time series, and is of high precisionand practicability as well.3. An improved method based on multiple population parallel search andadaptive search step for a new type of intelligent evolutionary algorithm which isbeing named the cuckoo search algorithm was put forward in this study. Then theimproved method was introduced to optimize the support vector machine parameters.The results shows that the improved method can effectively improved theclassification performance of support vector machine. And the simulation resultsfound that: besides the high accuracy, the improved cuckoo search algorithm hasfaster convergence speed and stability compared the original cuckoo search algorithm,genetic algorithm and particle swarm optimization.4. Based on Hilbert-Huang transform method, this study decompose the realestate price index into several intrinsic mode functions which been classify into3categories and constitute3basic components of the original sequence by followingmeans of T test, Hilbert-Huang spectrum analysis and power spectrum analysis.Based on this, the142estate regulation and control policies which were promulgatedbetween May9,2002and February9,2011were tested. By simulated the forcecaused by those policies53policies affected the market price and54policies whichhave stronger force affected the major event price has been founded. Based on this,the effect mechanism that the policy events no time series was founded.5. A hybrid method which integrated the error prediction method and policyintervention model was proposed finally by synthetically using econometric model and intelligence computing. Firstly using Hilbert-Huang transform to constitute3basic components of the original sequence, different models were chosen according tothe different scale characteristics of each component and finally get the satisfyingforecasting results by integrating them together.