The Stack Overflow podcast is back! Listen to an interview with our new CEO.

# Tag Info

## Hot answers tagged churn

8

I think your question could be further defined. The first distinction for churn models is between creating (1) a binary (or multi-class if there are multiple types of churn) model to estimate the probability of a customer churning within or by a certain future point (e.g. the next 3 months) (2) a survival type model creating an estimate of the risk of ...

4

Temporal abstraction is fancy pants jargon for changing the reference point of a time series variable. It is a type of variable transformation. A common example is that you have behavioral data that is indexed by calendar time, like number of minutes used in a given month on a cellular plan. This is typically how this sort of data is stored in a database. ...

4

It’s perfectly okay to use “days since last log-in” as one of the independent variables in the churn model. This independent variable does not, in and of itself, define churn. And it’s obvious that if a customer has not logged on for a long time (i.e., lagged) then she is more likely to churn in the near future. The important distinction here is that the “...

3

The best way to think about Bayes is conditioning. So when you have a question like this: What is the chance that client will leave if he has >0 refund? This question is answered with just data from the refund>0 row. $50/(50+133)\approx 27,3\%$ (Matching the question, I'm using comma as the radix point.) So it is the case that churning and refunds are ...

3

Re 1: If you predict well in the hold out sample then you're doing well (no time to worry about propriety ;-) But since you're asking... One way to look at the threshold is that when you set it to 0.1 you are implicitly specifying a loss function. That is, separating the question of what to do (e.g. approach a customer) from what to infer (e.g. that the ...

3

It depends entirely on whether historical churn in 2013 is a good predictor of churn in mid-2014, i.e. whether the training-set was predictive of test-set behavior. In general you should assume no. [*] (Obviously the actual individual customers churning are different. But do they churn for different reasons? duration? cost? product usage? etc. Did those ...

2

AFAIK, churn analysis is the domain specific lingo of survival analysis applied to customer relationship management. The literature and google-ability of "survival analysis" is far superior to "churn analysis." A popular way of modelling survival is the Cox proportional hazards model. This model assumes that the influence of an event on another event is ...

2

A Bayesian network is a graphical representation of an arbitrary probabilistic model. Logistic regression is one very specific kind of probabilistic model, and it can be represented by a Bayesian network. So you can't really say that one is better than another at prediction.

2

I believe that this is a question of a habit and not of mathematics. Most measure (e.g., precision, recall) are defined with respect to the positive class. Of course that you can use the related measures with respect to the negative class but by sticking with the norm you communication is less confusing. In the case of churn there is another benefit from ...

2

I have a small experience in churn prediction modeling and initially I had the same question in my mind. First of all, even if the labels are switched, the only thing that changes would be the interpretation of the model. Here is a quick R example: set.seed(1) x <- matrix(rnorm(100*2), 100, 2) y <- sample(0:1, 100, replace = TRUE) coef(glm(as.logical(...

2

This comes from the geometric distribution. If the probability of churning in any given year is $p$ (where $p$ is constant across years, and years are independent) then the expected number of years until you churn is $1/p$.

2

Consider the analogy to stock pricing. If a stock is at 10, and goes to 11, it goes up 10%. If it then goes up to 12, it only goes up 9.1%. Your monthly attrition is relative to what you have at the start of the month

2

If you use logistic regression instead of random forest, you will be able to associate every individual (churner or otherwise) to a probability of churn. With a separate asymptotic analysis or something similar you may be able to generate a critical threshold for this probability. This is one way to do it, but not the only nor the best (still I think it ...

2

Let's take a quick step back. A logistic regression takes the following form: $$\text{log}\frac{\pi}{1-\pi} = \beta_0 + \beta_1 x_1 + \beta_2x_2...$$ where $\pi = P(Y=1 | X)$, where $Y=1$ is churn and $Y=0$ is not churn. Next, we know that the odds in favor of $Y=1$ can be written as: \text{odds}(Y=1|X) = \frac{P(Y=1|X)}{P(Y=0|X)} = e^{\beta_0 + \...

1

Effect size applies to the effect you are measuring. As the examples you put in your question demonstrate, even within the field of churn, there are different effects that are interesting and the way to measure them will vary. As to small, medium and large, Cohen just came up with some guidelines to use when you don't know what to do; I believe he did this ...

1

No, your understanding doesn't seem to be correct. There are two parts to the paper: user clustering, and then fast response churn prediction. K-means clustering appears in the former but not the latter. The idea is to pass each user's activity embedding through K different LSTMs, each of which computes a different user behavior embedding. The user ...

1

If you have the end-date for all customers, then as far as I can see, you don't have censoring. Hence, you could treat the time a continuous variable (possibly consider a transformation, such as the log if it has a non-symmetric distribution), and use a linear regression model.

1

The questions is whether the data you collected can explain churn at all. The model can learn some weights / thresholds but it will not extract knowledge / rules that are outside your data. E.g. if the most influential factor is (change of) income and you don't have it in your data, your predictions will be far from perfect. Assuming positive answer - i.e. ...

1

Don't predict yes/no. Instead, predict a probability that someone will churn. Then base any actions (e.g., calling them, offering a discount, whatever) on that probability and the costs of correct/wrong actions. This earlier thread may be helpful.

1

First of all there needs to be clarification on your use of the phrase "survival analysis". In statistics there is a particular area associated with this term: https://en.wikipedia.org/wiki/Survival_analysis These survival methods (e.g., proportional hazards regression) address time to events and include concepts like censoring of observations and competing ...

1

Because your outcomes are one week churn events I do not think you data support analysis of 8 week conditional churn probabilities without some assumptions having a serious likelihood of being false. (I.e. successive churn probabilities are (1) independent, and (2) homogeneous.) tl;dr: You can't "predict the likelihood of churn for not only the next week (...

1

Look for Cox Regression applied in churn prediction: http://daynebatten.com/2015/02/customer-churn-cox-regression/

1

so breaking it down to 3 points of interest: RFM Approach Segmentation via clustering Changes in behavior RFM is a widely accepted way of customer segmentation. Since you have more data than traditional RFM scores, it might be worthwhile to fit a simple tree based model (GBM?) to your y variable of interest (lifetime value, last purchase etc.) and see ...

1

If there is not much difference between customers who stay active and those who churn then, pretty much by definition, no method will work well. You need to find some variable (or variables) which do distinguish the two sets.

1

I guess one way to model that problem is to assume that on normal days the waiting times between visits are distributed as some reasonably "compact" distribution, for example $\tau \sim Exp(1/2)$ however in festive season the waiting times are distributed as a mixture of original distribution and some much more extended distribution, for example N(20,5$^2$)....

1

Cox Proportional Hazard model is not really appropriate for churn modeling because the hazards are almost never proportional to each other in practice (they cross each other over time.) You should consider using Discrete-time Flexible Hazard model using logistic regression. Check out the summary of this method in the following paper: http://www.lexjansen....

1

I have recently done this kind of customer analytics and generally I would say it is wise to split this problem into two parts. But you do not necessarily need to apply survival analysis for churning since you can think it as a panel problem where initial customer set is observed and then churned customers are observed during fixed time period. Target ...

1

As Peter Flom notes, you can't use CART on unlabeled data. Nor any other supervised learning technique. All you can do is to cluster your data. Which will not help in predicting churn, if churn doesn't enter the model in some way. However, if you have transaction, deposit and withdrawal data, then you can label your data. Anyone with a zero balance and no ...

1

In a business-analytic context, this is often referred to as churn modeling or churn prediction. It employs a wide variety of statistical ideas and methodologies including survival analysis, Markov-chain and stochastic-process theory, hierarchical Bayes models, etc.

Only top voted, non community-wiki answers of a minimum length are eligible