- One thing we can potentially do for user churn is to build a model that predicts user churn. With a decent data set size we could split the data and actually create a simple decision tree or random forest that can make an 80-90% accurate prediction. - I think rather than the week and cohort wise churn it might be better to show a graph at what point of the journey are users most likely to churn. - With the model we can also identify characteristics that have a positive and negative correlation to churn i.e. which features lead to churn and which prevent churn. - Users most likely to churn is quite easy with the churn score. We just need to think through how to rank these users. This user ranking is also common to [[Product Feedback]]so might be worth doing separately. - Honestly rather than doing a separate by cohort analysis I think it makes sense to create a user scoring system for companies. It's something we did at WorkNetwork and we can leverage that experience. We can create a scoring algorithm for users based on past data that basically allows you to predict a LTV of a given user. This feeds well to our churn predictor, product feedback and in the long run can convert user it in lead qualification as well. - I think the value in this part of the product comes from creating some sort of rule based or AI model. With User Churn there's a lot of tools out there that aggregate data and allow you to build graphs and standard tables. I think other than Pecan though no one else is creating a model for you based on certain features. We could do the same thing as Pecan and allow users to select features and train a model. This is basically an AI product though and I'll definitely have to up skill a bit in terms of AI. ### Approach - Ask users to define churn and to give a description of what all they want us to use in our model. This can be a list of tables and also might include data sources like tickets etc. The second part we can also do manually or find algorithms for. - Have a few algorithms and train a few models. Divide the data into 90% training and 10% test and see how good the models predict. Return the right model to the user. - The models should also tell you which features have the highest weightage among all the described properties. - Graph 1 - Can show either what is there or we can show a graph with a histogram based on how long it takes for users to churn. - Graph 2 - Features of retained users. What are the features that good users have. This could both be part of their characteristics or actions they perform. - Graph 3 - Features of churned users. What are the features that bad users have. This could both be part of their characteristics or actions they perform. - Graph 4 - Users that are likely to churn. If we build a LTV model we can rank the users by LTV as well.