How to use historical data to improve marketing success rate


In the article "the whole process of using historical data to make business prediction" (hereinafter referred to as the previous article), we introduced how to use historical data to make business prediction. Different business needs will have their own particularity. This article will introduce how to use prediction technology to improve the success rate of marketing.

1. Prepare historical data

In the marketing scenario, the target to be predicted is the customer’s purchase behavior. It is necessary to collect some information fields that may affect the purchase behavior, such as the customer’s age, education background, work, income, family structure, habits, shopping preferences, product characteristics and promotion efforts. The more relevant information is collected, the better the prediction effect will be.

In addition, we can also make prediction by region and customer group according to business characteristics. For example, the house price in New York is totally different from that in central and southern cities. For another example, in car sales, male customers usually focus on performance, while female customers pay more attention to appearance. There are also differences in demand characteristics between high-end customers and medium and low-end customers. In many cases, the analysis and prediction of customer groups is better and more targeted than all direct prediction.

If the prediction is targeted to regions and customer groups, more sheets of corresponding wide tables should be prepared. For example, if there are three customer groups, three wide tables should be prepared.

2. Build the model

As mentioned in the previous article. If there are multiple customer groups, multiple models need to be built.

3. Predict customer purchase list

With the method mentioned in the previous article, we can achieve the prediction, and then according to the prediction probability results from high to low, we can find the customers with higher probability to carry out marketing activities. The marketing success rate of the top customers is higher.

4. Lift index

In the marketing scenario, in addition to the general AUC index to check the accuracy, there is also a very practical evaluation method called lift curve. Lift is a measure to evaluate the effectiveness of a prediction model. Its value is the ratio between the results obtained with and without the prediction model. As shown in the figure below.

The abscissa represents the number of the predicted probability in order from high to low, 10,20…represents the top 10%, 20%… of the samples, and the ordinate represents the corresponding lift value in the ranking stage. For example, the benchmark purchase rate of a product is 1.5%, which means that in traditional marketing without model, 1.5 people in 100 people will buy the product. Then, after establishing the model, it can be seen from the lift curve in the figure that the lift of the top 5% is 14.4, that is, there will be 1.5 * 14.4 = 21.5 people buying products in 100 people. That is to say, the success rate of marketing for the top 5% customers can be 14.4 times higher than that of traditional marketing. With the increase of the percentage of users on the abscissa, the lift value shows a decreasing trend, and the importance of corresponding customers is also decreasing. When it is reduced to a certain segment, it is not of great significance for marketing. For example, for the top 15% of customers in the figure, the lift value is greater than 1, which means that the success rate of marketing for the top 15% of customers is higher than that of randomly selected customers. According to the lift curve, we can decide how many percentage of customers in the top of the probability ranking are selected for marketing. The steeper the lift curve is, the better the ability of the model to select high-quality customers. As shown in the figure, the lift curve is a good model, which can help us find the target customers more effectively and find the potential customers at the lowest cost.

5. Multi product portfolio purchase list

If there is only one or a few products to sell, step 4 is complete.

If there are many kinds of products to be sold, such as a dozen or even hundreds of products, we can further improve the success rate and marketing value by exploring customers’ interests and recommending product combinations to them. For example, banks may have dozens of financial products to market, home appliance companies may have a variety of home appliances to sell, supermarkets or e-commerce companies may have a variety of products to sell, and insurance companies may have various types of insurance products to market…

The classic case of beer and diaper in history is to increase the sales of both diapers and beer by mining data rules and selling two seemingly unrelated product combinations. For another example, there are many kinds of financial products in banks, so we can combine several products with high purchase probability to sell by mining users’ purchase preferences.

The prediction of a multi-product portfolio purchase list is also very simple, with off-the-shelf functional modules in YModel. Specific operations are as follows:

(1) Modeling data set: prepare a wide table of multiple objectives. Make a wide table of historical information and all target variables needed to predict the products, as shown in the following figure. y1, y2, y3… represents historical data on whether or not each product was purchased, i.e. multiple targets.

(2) When configuring the target variable, change the single target variable to multiple target variables, as shown in the figure. YModel will automatically combines products based on user preferences.

Other operation steps are the same as single product purchase prediction. After the prediction is done, the result similar to the following figure will appear:

The first column on the left is the content of the product portfolio, and the second column is the probability that the user will buy the portfolio. Once the results are exported, the product portfolio purchase list is generated, and the top customers with higher probability can be marketed. It should be noted that there is no lift curve for the combination probability. The number of top customers will depend on the situation (usually this number will be more than the number of customers for a single product).