The allocation optimization of promotion budget and traffic volume for an online flash-sales platform

Abstract

This paper discusses an allocation issue of promotion budget and traffic volume, faced by VIP.com that is the largest online flash-sales platform in China. Through flash sales mode, each day VIP.com provides consumers one hundred authorized brands of products on consignment, which last for a short period of time (frequently 3–5 days), called by VIP.com ‘Dangqi’. As a consignee, VIP.com has no pricing rights of products but can allocate the promotion budget and traffic volume to each brand. Due to different categories of products and different brands of products having different responses to the same promotion budget and different profit margins, it is very necessary for VIP.com to allocate promotion budget and traffic volume among all brands of products offered in each ‘Dangqi’ reasonably in order to improve the usage efficiency of limited resources. Based on the real historical data of VIP.com, we first find main elements that influence the sales revenues of different brands of products through machine learning. We then predict the sales of all brands of products in each ‘Dangqi’ by multiplicative regression model, which has better accuracy of forecasting than other forecast models and obtain the function relationship between the total sales of each brand and its main impact elements. Finally, considering VIP.com’s actual concerns, we develop allocation optimization models with objectives of maximizing VIP.com’s total sales and total sales profit, respectively. The results from VIP.com’s real data tests show that under the same resource investment the presented allocation optimization approach can yield a significant increase in VIP.com’s total sales and sales profit in each ‘Dangqi’.

Appendices

Appendix A: The comparison of different similarity measurement methods

1.1 A1: Euclidean distance

First, we give out the expression of Euclidean distance, which is:

$$ d\left( {X,Y \, } \right) = \sqrt {(x_{1} - y_{1} )^{2} \, + \, \cdots \, + \, (x_{i} - y_{i} )^{2} \, + \, \cdots \, + \, (x_{n} - y_{n} )^{2} } . $$

(7)

Use Euclidean distance as the measurement of the similarity of two brands, and according to the process explained in Sect. 3.2, we get the optimal number of brand groups under suit dress category is 4, the results of forecasting are shown in Table 5.

Table 5 The forecasting accuracy of different models for 4 BGs with Euclidean distance

1.2 A2: Minkowski distance

First, we give out the expression of Minkowski distance, which is:

$$ d\left( {X,Y} \right) = \left[ {\sum\limits_{i \in N} {\left| {x_{i} - y_{i} } \right|^{m} } } \right]^{{\frac{1}{m}}} . $$

(8)

Similarly, use Minkowski distance as the measurement of the similarity of two brands, and according to the process explained in Sect. 3.2, we get the optimal number of brand groups under suit dress category is 5, the results of forecasting are shown below (Table 6)

Table 6 The forecasting accuracy of different models for 5 BGs with Minkowski distance

Next table gives out the comparison of the results from three different similarity measurement methods.

As we can see from Table 7, using cosine value of two brands to present their similarity brings higher forecasting accuracy. On other hand, we also see that regardless of the similarity measurement methods we use, multiplicative regression model is always better than other regression models with respect to forecasting accuracy, which verifies the reasonability of forecasting model we use.

Table 7 The results of different similarity measurement methods

Appendix B: Error normality test

To verify the correctness of the results of stepwise regression and the reasonability of the final forecasting model, we carry out normality test for the forecasting error, take suit dress category as example (Fig. 17):

Fig. 17

Error distribution for different brand groups under suit dress category

We use K–S test to exam the error data, the results are (significance level: 0.05):

From Table 8, the p values of all brand groups are much bigger than the significance level, therefore, the errors are normally distributed in our test, which proves the correctness of our forecasting model.

Table 8 The results of K–S test

Appendix C: The results of 5 and 6 BGs

See Figs. 18, 19 and Tables 9, 10, 11, 12.

Fig. 18

The midpoints of 5 BGs

Fig. 19

The midpoints of 6 BGs

Table 9 Independent variables for 5 BGs

Table 10 Independent variables for 6 BGs

Table 11 The forecasting accuracy of different models for 5 BGs

Table 12 The forecasting accuracy of different models for 6 BGs