Algorithm Research For Time Series Prediction And Analysis In Retail Industry

With the continual rise of experiential shopping,the retail industry is welcoming its new life in various forms.There are many business models in the retail industry,and this study selects one of its most classic forms of expression,the shopping center.This paper takes the sales time series data of a shopping mall as the research object,analyzes various influencing factors that may affect the retail sales behavior,constructs appropriate feature engineering,and constructs a suitable time series analysis and prediction model.Two time series forecasting problems that are most concerned about shopping centers:sales and sales frequency,scientific analysis and forecasting.It mainly studies time series prediction algorithms and feature selection.First,in terms of feature selection,a correlation-complement-redundant feature selection model for retail time series problems is adopted.Introducing the complementarity between features can enhance the overall expressiveness of the feature subset.Experiments show that the features selected by the model have better prediction accuracy than the baseline feature extraction model in different prediction algorithms,and the relevant indicators are optimized by more than 3%.Secondly,in the retail time series forecasting,the EMD-based and deep learning prediction models for retail time series are adopted.The model uses the EMD decomposition algorithm to decompose the retail time series into sub-sequences with different information components,and mine the local characteristics of the sub-sequences to make targeted predictions.The sub-prediction model is constructed by using GRU and MLP models,and the sub-prediction results are combined to obtain the final prediction result.The RMSE indicator of the prediction model is optimized by more than 5%,and the MAE indicator is optimized by more than 1.5%.In addition,the model is optimized for efficiency,recombined by decomposed subsequences,and simplified into two subsequences,trend sequences and short-term fluctuation sequences.Experiments show that by modeling these two sets of sequences,the computational efficiency can be significantly improved at the cost of less than 1%accuracy.