Research On Promotional Sales Forecast Based On PCA And Particle Swarm Optimized BP Neural Network

Sales promotion plays an important role in modern enterprises marketing strategies. With the growing trend of consumption diversity and products homogenizationretailers are investing more and more in sales promotion of terminal market in order to inspire peolpe's desire to purchase.Retailers' promotion activities increase the uncertainty of market demand to a certain degree and make it more difficult to forecast demand of promotional products both promptly and accurately.In the light of the relative short life cycle and seasonal feature of apparel products theory of artificial neural network (ANN) was introduced to the domain of products demand forecast and a hybrid forecasting model was established to predict monthly sales of a specific men's shirt.Firstly, this paper made a systematic study on the theory of promotional demand factors and eight different factors was selected including seasonality product life cycle, promotion price discount, festival days, advertising investment and so on and those different factors were also quantified in order to measure the influential degree on products sales in a quantitative manner.Then, principal components analysis was made using SPSS statistical analysis software to decrease dimensions of input datas and eliminate the multicollinearity of different factors.Due to the limitations of its own learning mechanisms standard back propagation neural network (BPNN) may get into a local optimal value and take a relatively longer time to train the net.In order to overcome the limitations of the algorithm a hybrid forecasting model was established which combined principal components analysis(PCA), BP neural network and particle swarm optimization algorithm (PSO), aiming at improving BP neural network both from the aspects of sample data quality and initial parameters of BP neural network——Firstly, principal components analysis was used to reduce the dimensions of promotional factors and took the extracted components as the new input of the net in order to simplify the structure of network and improve the generalization ability. Secondlyparticle swarm optimization algorithm was introduced to train the initial parameters of the BP network including initial weights and bias values, aiming to decrease networks training time and improve the forecasting accuracy.Then simulation analysis was made on the basis of the fully trained forecasting model.Finally, an example analysis was made based on monthly sales datas of a specific men's shirt of brand S and the performances of different forecast models were also compared. in order to verify the validation of this model The results showed that the proposed model simplified the architecture of BP network decreased the training time of the network and to a certain degree improved the forecast accuracy.This proposed model can be used as a feasible solution to multi-factors forecasting problems and has a significant theoretical and practical value.