How did you score clients before switching to ML scoring?

Before implementing an ML model for client scoring, simple rules or logistic regression on basic features are usually used. This allows for a quick initial assessment of the client's risk or quality.

The process of transitioning to ML scoring includes several stages:

1. Data collection and preparation — historical data about clients and their behavior are gathered.
2. Development of a basic model — a simple model (e.g., logistic regression) is created for comparison with the ML model.
3. Training the ML model — a more complex model (e.g., gradient boosting) is built to demonstrate improved quality.
4. Validation and A/B testing — results of the basic scoring and ML scoring are compared on real data.
5. Gradual transition — initially, ML scoring is applied to a subset of clients, then scaled up.

An example of simple scoring in Python with logistic regression:

from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)
model = LogisticRegression()
model.fit(X_train, y_train)
predictions = model.predict_proba(X_test)[:, 1]

Thus, a basic scoring is built first, and then gradually replaced with an ML model with confirmed quality improvement.