In the telecom industry, customers are able to choose from multiple service providers and actively switch from one operator to another. In this highly competitive market, the telecommunications industry experiences an average of 15-25% annual churn rate. Given the fact that it costs 5-10 times more to acquire a new customer than to retain an existing one, customer retention has now become even more important than customer acquisition.
-
For many incumbent operators, retaining high profitable customers is the number one business goal.
-
To reduce customer churn, telecom companies need to predict which customers are at high risk of churn.
In this project, we will analyse customer-level data of a leading telecom firm, build predictive models to identify customers at high risk of churn and identify the main indicators of churn.
There are two main models of payment in the telecom industry - postpaid (customers pay a monthly/annual bill after using the services) and prepaid (customers pay/recharge with a certain amount in advance and then use the services).
In the postpaid model, when customers want to switch to another operator, they usually inform the existing operator to terminate the services, and we directly know that this is an instance of churn.
However, in the prepaid model, customers who want to switch to another network can simply stop using the services without any notice, and it is hard to know whether someone has actually churned or is simply not using the services temporarily (e.g. someone may be on a trip abroad for a month or two and then intend to resume using the services again).
Thus, churn prediction is usually more critical (and non-trivial) for prepaid customers, and the term ‘churn’ should be defined carefully. Also, prepaid is the most common model in India and southeast Asia, while postpaid is more common in Europe in North America.
This project is based on the Indian and Southeast Asian market.
There are various ways to define churn, such as:
Revenue-based churn: Customers who have not utilised any revenue-generating facilities such as mobile internet, outgoing calls, SMS etc. over a given period of time. One could also use aggregate metrics such as ‘customers who have generated less than INR 4 per month in total/average/median revenue’.
The main shortcoming of this definition is that there are customers who only receive calls/SMSes from their wage-earning counterparts, i.e. they don’t generate revenue but use the services. For example, many users in rural areas only receive calls from their wage-earning siblings in urban areas.
Usage-based churn: Customers who have not done any usage, either incoming or outgoing - in terms of calls, internet etc. over a period of time.
A potential shortcoming of this definition is that when the customer has stopped using the services for a while, it may be too late to take any corrective actions to retain them. For e.g., if we define churn based on a ‘two-months zero usage’ period, predicting churn could be useless since by that time the customer would have already switched to another operator.
In this project, we will use the usage-based definition to define churn.
In the Indian and the southeast Asian market, approximately 80% of revenue comes from the top 20% customers (called high-value customers). Thus, if we can reduce churn of the high-value customers, we will be able to reduce significant revenue leakage.
In this project, we will define high-value customers based on a certain metric (mentioned later below) and predict churn only on high-value customers.
The dataset contains customer-level information for a span of four consecutive months - June, July, August and September. The months are encoded as 6, 7, 8 and 9, respectively.
The business objective is to predict the churn in the last (i.e. the ninth) month using the data (features) from the first three months. To do this task well, understanding the typical customer behaviour during churn will be helpful.
Customers usually do not decide to switch to another competitor instantly, but rather over a period of time (this is especially applicable to high-value customers). In churn prediction, we assume that there are three phases of customer lifecycle :
-
The ‘good’ phase: In this phase, the customer is happy with the service and behaves as usual.
-
The ‘action’ phase: The customer experience starts to sore in this phase, for e.g. he/she gets a compelling offer from a competitor, faces unjust charges, becomes unhappy with service quality etc. In this phase, the customer usually shows different behaviour than the ‘good’ months. Also, it is crucial to identify high-churn-risk customers in this phase, since some corrective actions can be taken at this point (such as matching the competitor’s offer/improving the service quality etc.)
-
The ‘churn’ phase: In this phase, the customer is said to have churned. We define churn based on this phase. Also, it is important to note that at the time of prediction (i.e. the action months), this data is not available to us for prediction. Thus, after tagging churn as 1/0 based on this phase, we discard all data corresponding to this phase.
In this case, since we are working over a four-month window, the first two months are the ‘good’ phase, the third month is the ‘action’ phase, while the fourth month is the ‘churn’ phase.
The dataset can be downloaded from here.
Data dictionary is uploaded. The data dictionary contains meanings of abbreviations. Some frequent ones are loc (local), IC (incoming), OG (outgoing), T2T (telecom operator to telecom operator), T2O (telecom operator to another operator), RECH (recharge) etc.
The attributes containing 6, 7, 8, 9 as suffixes imply that those correspond to the months 6, 7, 8, 9 respectively.
The following data preparation steps are crucial for this problem:
-
Derive new features This is one of the most important parts of data preparation since good features are often the differentiators between good and bad models. We will use our business understanding to derive features that we think could be important indicators of churn.
-
Filter high-value customers As mentioned above, we need to predict churn only for the high-value customers. Define high-value customers as follows: Those who have recharged with an amount more than or equal to X, where X is the 70th percentile of the average recharge amount in the first two months (the good phase).
-
Tag churners and remove attributes of the churn phase Now tag the churned customers (churn=1, else 0) based on the fourth month as follows: Those who have not made any calls (either incoming or outgoing) AND have not used mobile internet even once in the churn phase. The attributes we need to use to tag churners are:
- total_ic_mou_9
- total_og_mou_9
- vol_2g_mb_9
- vol_3g_mb_9
After tagging churners, we need to remove all the attributes corresponding to the churn phase (all attributes having ‘ _9’, etc. in their names).
Build models to predict churn. The predictive model that we are going to build will serve two purposes:
-
It will be used to predict whether a high-value customer will churn or not, in near future (i.e. churn phase). By knowing this, the company can take action steps such as providing special plans, discounts on recharge etc.
-
It will be used to identify important variables that are strong predictors of churn. These variables may also indicate why customers choose to switch to other networks.
In some cases, both of the above-stated goals can be achieved by a single machine learning model. But here, we have a large number of attributes, and thus we should try using a dimensionality reduction technique such as PCA and then build a predictive model. After PCA, we can use any classification model.
Also, since the rate of churn is typically low (about 5-10%, this is called class-imbalance) - we will try using techniques to handle class imbalance.
Below are the steps followed to build the model:
- Preprocess data (convert columns to appropriate formats, handle missing values, etc.)
- Conduct appropriate exploratory analysis to extract useful insights (whether directly useful for business or for eventual modelling/feature engineering).
- Feature Engineering : Derive new features.
- Use PCA to reduce the number of predctors.
- Train a variety of models, tune model hyperparameters, etc. (handle class imbalance using appropriate techniques).
- Evaluate the models using appropriate evaluation metrics. Note that it is more important to identify churners than the non-churners accurately - choose an appropriate evaluation metric which reflects this business goal.
- Finally, choose a model based on appropriate evaluation metrics.
The above model will only be able to achieve one of the two goals:
-
To predict customers who will churn. We can’t use the above model to identify the important features for churn. That’s because PCA usually creates components which are not easy to interpret.
-
Therefore, we will build another model with the main objective of identifying important predictor attributes which help the business understand indicators of churn. A good choice to identify important variables is a logistic regression model or a model from the tree family. In case of logistic regression, we will make sure to handle multi-collinearity.
After identifying important predictors, display them visually - we can use plots, summary tables etc. - whatever we think best conveys the importance of features.
The top variables chosen for the logistic regression model are listed below.
| Variables | Coefficients |
|---|---|
| loc_ic_mou_8 | -3.3287 |
| og_others_7 | -2.4711 |
| ic_others_8 | -1.5131 |
| isd_og_mou_8 | -1.3811 |
| decrease_vbc_action | -1.3293 |
| monthly_3g_8 | -1.0943 |
| std_ic_t2f_mou_8 | -0.9503 |
| monthly_2g_8 | -0.9279 |
| loc_ic_t2f_mou_8 | -0.7102 |
| roam_og_mou_8 | 0.7135 |
- The majority of the top variables, as can be seen, have negative coefficients. This indicates that the factors and the probability of churn have an inverse relationship.
E.g : -
The likelihood that a client will leave increases if their local inbound minutes of usage (loc ic mou 8) are lower in August than in any other month.
Recommendations
- Focus on clients who use fewer minutes for inbound local calls and outgoing ISD calls during the action phase (mostly in the month of August).
- Pay special attention to the clients who pay more in July than they do in August.
- In addition, clients that experience increasing value-based costs throughout the action phase are more likely to leave than other customers. So, making an offer to these customers may be a good idea.
- Customers have a greater chance to be churned if their monthly 3G recharge is higher in August.
- Customers who used fewer STD incoming minutes on fixed T lines from operators T in August are more likely to churn.
- Customers who use less 2G data each month in August are more likely to churn.
- Consumers who used less incoming minutes on fixed T lines from operators in August are more likely to leave.
- Variables in roam_og_mou_8 have positive coefficients (0.7135). That means for the clients, whose roaming outbound minutes of consumption is increasing are more likely to churn.