Research On Forecasting Method Of Intermittent Demand Of Vending Machine

The retail industry is one of the important industries that lead the development of the world's industries.In recent years,the rise of housing rents and labor costs and the popularity of electronic payments have led to a rapid increase in the cost of traditional retailing,while vending machines have developed rapidly.As an automated trading device that is mobile,easy to trade and sells goods without time and place constraints,vending machines can effectively reduce retail costs while meeting consumer demand.At present,it has become the focus of maj or retail companies around the world,and its market development prospects are very broad.In terms of sales,the price of vending machines is more stable than traditional retail stores,and is not affected by promotions and bundles;in terms of inventory,due to limited vending machine capacity,vending machines cannot provide the same retail stores.The same sufficient inventory,although the development of information technology makes the out-of-stock time and status can be informed in time,but because the replenishment of vending machines is less frequent than the traditional retail industry,it is still easy to cause sales losses,making it suitable for the inside of the vending machine.Product demand forecasting is significant in guiding replenishment;in terms of cost,food vending machines sell foods with shorter shelf life than regular vending machines,and companies need to reduce the demand for each product to reduce the loss of product expiration.This paper forecasts the demand for each item in each sales unit.The difficulty of this prediction problem is that the product demand for food vending machines is sporadic and volatility,resulting in the majority of daily demand sequences being intermittent.There are a lot of time points in the sequence with zero demand,using ordinary ARIMA time.The prediction bias of the sequence model is large,and the product demand of the vending machine is limited by the inventory.The prediction accuracy of the traditional method of dealing with the discontinuous time series,the Croston method and the Syntetos-Boylan approximation method(SBA method).Not ideal.According to the characteristics of time series of vending machine sales data,the research literatures in related fields at home and abroad are read,and the existing product demand forecasting methods and their defects are analyzed in detail,and on this basis,improvement and optimization are proposed to improve product demand.Predicting accuracy and applicability,forming new predictive models to guide production practices.In this paper,the traditional discontinuous time series prediction method is used as the benchmark model,and the prediction method of polymerization decomposition is proposed based on these benchmark models.First,the raw data is aggregated,the length of the observation period of the time series is changed,and a periodic demand is obtained,which is used as a time series with the periodic demand,or converted into an observation value at the sales unit level,and the demand of each sales unit is predicted.·Then,the aggregated time series is predicted in one step using the methods of ordinary ARIMA and Holt-winters.Secondly,based on the Croston method and the SBA method,this paper proposes a Croston and SBA method based on greedy thought,using the Croston and SBA methods based on greedy thought and the proportional decomposition method to decompose the aggregation result into the prediction unit of product level or daily observation period.In the above,as the prediction result of each product,the prediction accuracy of different aggregation decomposition combination methods is compared.Finally,combined with the integrated learning method,the model with better prediction effect is integrated and learned to form the final model.In order to obtain more accurate analysis results,we use the sales data of vending machines of a multinational retail company in recent years to predict their sales and use historical data to compare the models.Experiments show that after the improvement of the traditional time series prediction method,the prediction error is greatly reduced,the accuracy is improved,and the performance of the model is further improved.