Application Research On A Time Series Forecasting Method

Economic forecasting is the measurement of the future vision of the economic phenomena, depending on the accurate statistical and economic information, and using the scientific method guiding by the history, actuality and rules of the economic phenomena. Economic forecasting can be an effective means in areas from editing and guiding the state plans to improvement of the enterprise management. This paper's main effort is in the forecasting of the sales of an enterprise. In a word, the forecasting of sales means foretelling the future sales by analyzing the past sales information of this enterprise.By the nature of the method of forecasting, there could be qualitative economic forecast and quantitative economic forecast. The accuracy of qualitative economic forecast is determined by the experience, level of theory and businiess, and capability of getting clues and inferring of the forecaster. While the quantitative economic forecast measures the scale, level, speed and ratio relationship of the future development for the economic phenomena, using statistical methods and mathematical models, depending on accurate, timely, systematical, comprehensive statistical an economic information.In the economic statistic and forecast for society and enterprises, editing and anylysing the time series can reflect the process and characteristics of the changing of objective things over time, and it can help the research and exploration the trends and rules of the changing of phenomena, the scientific forecasting of the future status of social economics and enterprise management, and can be great significant as guidance and utility. By now, time series forecasting methods have included classical statistical methods, neural networks methods and machine learning methods etc. The requirement of the forecast result for this paper is that finding and designing an as simple as possible method of time series forecasting yet retaining some accuracy, which could be used to analyze and forecast the sales of an enterprise. So according to the practical situation, we choose to use the classical statistical methods but such new techniques as neural networks to perform the time series forecasting.As to the methods we choose, the main aspects of research on improving the accuracy of time series forecasting is twofold: (1) the selection of the forecasting model: select proper model depending on the nature of the time series.(2) the improvement of the forecasting model: constantly impove the forecasting model so as to the improved model can get better results than original model in spite of less of more accurate input time series.On the other hand, depending on the practical working experience, as to the time series forecasting, a common phenomenon is that the forecast results can be more accurate deriving from week data or month data than from workday data.So this paper proposed a method improving the accuracy of time series forecasting based on analyzing the traditional statistical methods of time series forecasting, aiming at the requirements of sales of enterprises. This method makes up for the shortcomings of traditional methods of time series forecasting, in the scenarios of the requirement that this paper needs to solve, and can provide better results of forecasting while retaining the simple and convenience of traditional methods.For the unstable time series ,the forecasting result of time series forecasting method is comparatively not accurate. In this paper, base of the analysis of shortcomings of current time series forecasting, a method of processing original time series and forecasting result is promoted to improve the forecast accuracy . That is, merge the original time series by reasonable formula to make the merged time series more smooth, and then do forecasting using the merged time series , then distribute the forecast result by the method given in this paper.On the merge and merged data must be the result of the practical significance of such merge method to have their reasonable. Because of the time series of different significance, the actual demand is also different, therefore, can not be a panacea standard apply to merging any type of time-series. This paper provides several manners (merge by week, merge by month, merge by goods, merge by location) is for the sales forecast.In this paper , the method of distributing the forecast result of merged data is that every merged data was distribute by a list of scale. Three method is given in this paper to get the distribute scale, and they are average method, simple average method and linear fit method . The ratio of the average method value is fixed, that is 1:1:…: 1. Simple average method and linear fit method is based on the past time series calculated. The calculation of ratio is more flexible ,suitable for estimating the ratio of each element in a period of any cyclical cycle data, not limited to any system.