Two Data-driven Solutions For Inventory Prediction

Inventory forecasting can help enterprises to make effective management decisions,such as staffing and promotion decision-making and so on.Precise prediction can lowers the cargo loss,improve the service quality,and enhance the enterprise competitiveness.Traditionally,there are two types of approaches to predict the unit to the stock and delivery.One is forecasting the demand based on the past experience.And the other is using the economic formula to figure out the required inventory related information.However,in current retailing market,trading activities have become more and more frequent and complex than ever.As a consequence,a huge amount of data regarding stocked goods(items)in an inventory will be generated every day.In addition,the correlated relations among items are tend to be more complex(e.g.some items’ sales amount may be affected by others),which further increases the difficulty of efficient inventory forecasting when adopting traditional methods.In this work,we treat inventory management as a data mining problem and propose two efficient and effective inventory forecasting methods.1.A two-step dynamic inventory forecasting model.Inventory data is susceptible to seasonal,trend and special events and other factors,resulting in greater volatility changes.It is difficult to obtain an accurate prediction using only a single time series analysis model.The dynamic prediction method proposed in this paper takes a variety of influencing factors into account in the forecasting process.Firstly,a variety of regression algorithms are used to establish the ensemble regression model to obtain the prediction base;then,we get the interpretable final forecasting results cooperating with various impact factors(e.g.seasonal,trend and special events)of inventory time series data.Our model is capable to capture the characteristics of stock-out time series from various sides,including the the trend of moving window in transaction time series(which are contained in the forecasting basis),as well as the impact from seasonal,trend and events(which are reflected in the dynamic forecasting process).Experimental results on real-world data reveal that the two-step dynamic forecasting model can acquires higher prediction accuracy and has a better interpretability to the analysis results.2.A joint prediction inventory forecasting model.According to the actual application of inventory management,the two types of time series of inventory are interdependent,and the existing methods often ignore the correlation between them and make independent prediction.We model the interdependence of time series data and integrate them into the process of time series prediction.The model can capture the dynamic relationships between multiple time series data set and predict their future values simultaneously.Specifically,in the domain of inventory management,we transform the requirement of inventory management into model constraints and perform time series prediction under the constraints.At the same time,a variety of time window width schemes are proposed,according to prediction accuracy,standard deviation and historical data The experiments are conducted on real world data set.And the results demonstrate the joint forecasting method can obtain the prediction result which satisfies the actual constraint in the case of guaranteeing certain prediction precision.