Predictive Software

Churn Prediction

Churn Prediction is a SaaS solution that predicts which of your users will quit before they do so.

How it works

Using its API, it receives user events, can predict on a pool of active users and communicate which of those users will quit by next week. A prediction will be issued every day by a model powered by a deep learning ensemble. The model learns from past behavior of other users in order to predict on current users.

To start using it, first contact us to discuss how to best tailor the service to your needs. After signing with us, you'll need to feed user events from your application to our API. The API will publish predictions daily, which you can then use to incentivise your users to stay.

Depending on the specifics of your application, it might take some time until the data collected yields a strong enough learner. During this time, the service is free and the API will provide a breakdown of the model's performance.


day  how many played in the last 10 days predicted to stay predicted to quit predicted to stay and actually stayed predicted to stay and actually quit predicted to stay and actually quit and could've intervened predicted to quit and actually stayed predicted to quit and actually quit predicted to quit and actually quit and could've intervened accuracy 
2018-02-15 17897 6007 11890 4643 1364 575 903 10987 1040 0.8733
2018-02-14 17687 6030 11657 3860 21701032 423 11234 1339 0.8533
2018-02-13 18220 6139 12081 4305 1834 911 578 11503 1318 0.8676
2018-02-12 18654 6130 12524 4455 1675 840 663 11861 1321 0.8746
2018-02-11 18978 6174 12804 4560 1614 790 697 12107 1291 0.8782
2018-02-10 18352 6174 12178 4624 1550 745 726 11452 1182 0.8759
2018-02-09 17855 6201 11654 4694 1507 725 747 10907 1146 0.8737
2018-02-08 17774 6152 11622 4729 1423 603 752 10870 1110 0.8776
2018-02-07 17807 6180 11627 4827 1353 550 747 10880 1020 0.8820
2018-02-06 18803 6310 12492 4937 1373 547 777 11715 1023 0.8856
2018-02-05 18972 6803 12169 5048 1755 645 726 11443 929 0.8692
2018-02-04 18655 6313 12342 4896 1417 633 779 11563 1022 0.8822
2018-02-03 18075 6203 11872 4876 1327 590 835 11037 999 0.8803
2018-02-02 17383 6111 11272 4826 1285 586 882 10390 1004 0.8753
2018-02-01 16810 6041 10769 4833 1208 549 868 9901 1033 0.8765

Optionally, we can provide a service that learns from the incentives given (provided that you're sending that feedback to the API), what worked and what didn't, and recommend what incentive to use for each user specifically so that you maximize your revenue and retention and further automate your application.

Learning duration and accuracy

The average accuracy achieved on real data so far was 85%. In order to reach it, the data required was on average about 3 months at 10,000 DAU.

Pricing

This service is billed monthly, with the months until accuracy exceeds 70% being free. Afterwards, the cost will be set depending on these factors:

*- A predictor is quantifiable event type that can be aggregated on a daily basis. Example predictors: hard currency spent, number of missions completed, number of main page requests etc.

Application Programming Interface


/api/v1/send_events

This is a POST call that accepts the following params:

application_id - the id assigned for your application
events - JSON-encoded array of events, e.g.

            [{"user_id":54321,
              "date":"2017-09-06",
              "predictor":"hard_currency_spent",
              "value":25},
             {"user_id":54321,
              "date":"2017-09-06",
              "predictor":"daily_missions",
              "value":2}
        

If you need to add new predictors, make sure you communicate it with us, as only accepted predictors will work.


/api/v1/predictions/<application_id>/<day>/<anticipation>

This is a GET call that returns a JSON-encoded list of users that are predicted to quit sometime between the given day and anticipation days later.


/api/v1/performance/<application_id>/<day>

This is a GET call that returns the HTML formatted table computing the performance of that day's prediction.

Contact

I'm Ludovic-Valentin Brasso, the owner of Predictive Software, a company focused on providing machine learning solutions. I have a background in Web Development (skilled in Python, SQL, Couchbase, Node.js, and PHP) and have made the switch two years ago to Machine Learning. Strong entrepreneurship professional with a Bachelor's degree in Automated Control and Computer Science from Polytechnic University of Bucharest. You can write to me at .