I will do churn prediction and predictive modeling with machine learning
PhD Candidate, Survey Statistics, SPSS, R, SEM
About this Gig
Need to know which customers will leave, or which bookings will cancel, before it happens? I build churn prediction and predictive models in Python and R that a manager can act on.
What you get:
- A leakage check, so the model only uses data you have at the moment of prediction
- A simple baseline and a tuned gradient boosting model (XGBoost, LightGBM or CatBoost), tested on later months of data
- Probabilities that match real rates (calibrated), so 30% means about 30 in 100
- A cutoff set by your costs: what a missed churner costs vs a wasted offer
- Main drivers explained in plain words
- Scored customer list (CSV or Excel), clean code and a PDF report
- Standard and Premium: a FastAPI scoring endpoint and a monthly model health check
- Premium: churn timing (when each customer is likely to leave), profit tables, Dockerfile
Works for subscriptions, SaaS, telecom, e-commerce, insurance, hotels and rentals.
See the sample report in the gallery: 119,390 real hotel bookings, cancellation risk with an out-of-time test.
Before ordering, send me a message with your column names and the decision you want to make. I will tell you honestly if the data can support a model.
Programming language:
Python
•
R
•
SQL
Frameworks:
Scikit-learn
•
Panda
Tools:
Jupyter Notebook
•
Excel
•
Stata
My Portfolio
FAQ
What data do I need to send you?
One table with one row per customer or booking, the outcome (left or stayed, canceled or not) and the date it became known. 12 months of history is a good start. CSV, Excel or a SQL export all work. Names and emails can be removed first, the model does not need them.
How many rows do I need for a reliable model?
What matters is the number of churners or cancellations, not total rows. A few hundred events usually give a usable model. With fewer, I fit a simpler model and show how wide the uncertainty is, so you know how far to trust it.
How accurate will the model be?
Nobody can promise a number before seeing the data. You get results on later data the model never saw, with 95% intervals, next to your current rule of thumb, so you see the real gain in plain terms.
My churn rate is only 5%. Is that a problem?
No. I keep the real rate and do not create synthetic rows, because that distorts the probabilities. The model is judged on precision-recall and calibration, and the cutoff comes from your costs.
Can my team use the results without writing code?
Yes. You get an Excel or CSV file with a risk score and the top reasons for each customer, plus a one-page guide. In Premium you also get a script that scores new data with one command.
Which industries do you work with?
Subscriptions and SaaS, telecom, e-commerce, insurance, gyms, hotels and short-term rentals. The method stays the same: fix the moment of prediction, the outcome window and the cost of each type of mistake.
Is my data kept confidential?
Yes. Your files are used only for your order, never shared or published, and deleted after delivery if you ask. Please remove names, emails and phone numbers first; the model does not need them.
Python or R: which one will you use?
Whatever your team already uses. By default Python with scikit-learn and LightGBM, or R with tidymodels. Code is commented, with fixed seeds and package versions, so every number can be rerun.
Can you predict when a customer will leave?
Yes. Premium includes a churn timing (survival) model: for each customer it shows the risk of leaving month by month, or for bookings, by days before arrival. With Basic or Standard you can add it as an extra.
Can you connect the model to my app or website?
Yes. Standard and Premium include FastAPI code: your app sends a customer record and gets back a risk score. Your developer runs it on your own server; Premium adds a Dockerfile. I never need access to your servers or accounts.

