Capturing High Value Customers among Migo’s Newly Registered Users for Conducting Precision Marketing

Project Details


Fall 2019


Bill Bennett, Wen‐Teng Chang, Yu‐Ting Lin, Tzu‐Lun Lin





Migo is an online‐to‐offline video provider startup focuses in Philippines market. Customers can down‐ load video online within Migo platform and watch it offline. Due to data (wifi) is not prevalent in Philippines, Migo has collaborated with several stores (hotspot) to provide the wifi service for their customers to down‐ load videos. Migo’s strategy now is to expand its market share by providing lots of free trial opportunities to the newly‐registered customers. However, how to retain those customers may serve as a critical issue.

Since retaining newly registered users is a critical issue for Migo, not to mention retaining newly regis‐ tered who tend to create more value to Migo, therefore, we would like to provide Migo a business solution which is about capturing potential high value customers among all newly registered users. To speak pre‐ cisely, Migo’s marketing team could conduct “less discriminative” precision marketing measure to cap‐ ture 44% of potential high value customers by marketing to only 20% of total newly registered customers based on our analysis result. The definition of high value customers referring to those who will spend more than 75% of total newly registered customers’ spends collected from September to November, and that equals to 100 pesos.

The response variable is Next_spend_bin referring to a newly registered customer’s spending in next month. This is a binary variable which 1 represents those who are high value customers and 0 is the oppo‐ site. We have derived several features by ourselves which comes from 3 data sources, transaction, title and engagement datasets given by Migo’s data science team. The algorithm utilized includes Logistic Regression, Random Forest and Gradient Boosting. Lift is our performance criterion which aligns to the business goal being conducting precision marketing. Therefore, we compare different algorithms’ lifts and finally select Gradient Boosting as our preferable model to provide and the result is presented in the previous paragraph.

The next step of Migo on top of our prediction result may compare between the cost of conducting precision marketing to those potential high value newly registered users and the profits they would pos‐ sibly return to Migo; after then, we can set a break‐even point and find out the decent threshold. Let’s say, if the threshold is capturing 40% of potential high value customer by just advertising to 20% of total newly registered customers, as 44% is already above 40%, we could take action. Note that 44% is constrained by the size of the dataset we possessed, this percentage could be even higher if the dataset goes larger as the more data we have the stronger algorithm we may construct.

Application Area: