Why Every ML Engineer Should Practice Explaining Their Model Aloud

1. WHAT IT PREDICTS (10 s)
   "It gives the probability that ____ will ____ within ____."

2. WHAT GOES IN (10 s)
   "It uses ____ groups of features: ____, ____ and ____."

3. HOW IT LEARNS (10 s)
   "It is a ____ model. I chose it over ____ because ____."

4. WHAT IT LEARNED (20 s)
   "The strongest drivers are ____ and ____. For example, ____."

5. WHERE NOT TO TRUST IT (10 s)
   "It is weaker for ____, so ____."

"It gives the probability that a subscriber cancels within
the next 30 days.

It uses four groups of features: recent usage, billing
history, support tickets and plan details.

It is a gradient boosting model. I chose it over logistic
regression because usage and plan type interact, and the
boosted model captured that better on validation data.

The strongest drivers are a drop in weekly usage and recent
failed payments. A customer whose usage halved and who had a
failed payment gets a much higher score.

It is weaker for customers in their first two months, so for
them the retention team uses simple rules instead."

from sklearn.datasets import load_breast_cancer
from sklearn.ensemble import RandomForestClassifier
from sklearn.inspection import permutation_importance
from sklearn.model_selection import train_test_split

data = load_breast_cancer(as_frame=True)
X_tr, X_va, y_tr, y_va = train_test_split(
    data.data, data.target, random_state=0, stratify=data.target)

model = RandomForestClassifier(n_estimators=300, random_state=0)
model.fit(X_tr, y_tr)

result = permutation_importance(model, X_va, y_va,
                                n_repeats=10, random_state=0)
order = result.importances_mean.argsort()[::-1][:5]
for i in order:
    print(f"{X_va.columns[i]:25s} "
          f"{result.importances_mean[i]:.4f} +/- {result.importances_std[i]:.4f}")

import numpy as np
import pandas as pd
from sklearn.linear_model import LogisticRegression

# Toy data: failed payments and weekly hours vs churn
df = pd.DataFrame({
    "failed_payments": [0, 0, 1, 2, 0, 3, 1, 0, 2, 4, 0, 1],
    "weekly_hours":    [10, 12, 6, 3, 9, 1, 7, 11, 4, 2, 8, 5],
    "churned":         [0, 0, 0, 1, 0, 1, 0, 0, 1, 1, 0, 1],
})
X, y = df[["failed_payments", "weekly_hours"]], df["churned"]
model = LogisticRegression().fit(X, y)

for name, coef in zip(X.columns, model.coef_[0]):
    print(f"{name}: coef={coef:.2f}, odds ratio={np.exp(coef):.2f}")
# Read aloud: 'each extra failed payment multiplies the odds
# of churn by about the odds ratio, holding hours fixed.'

import numpy as np
import pandas as pd
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler

df = pd.DataFrame({
    "weekly_hours":    [10, 12, 6, 3, 9, 1, 7, 11, 4, 2, 8, 5],
    "tenure_months":   [24, 30, 5, 3, 18, 2, 12, 36, 4, 1, 20, 6],
    "failed_payments": [0, 0, 1, 2, 0, 3, 1, 0, 2, 4, 0, 1],
    "churned":         [0, 0, 0, 1, 0, 1, 0, 0, 1, 1, 0, 1],
})
features = ["weekly_hours", "tenure_months", "failed_payments"]
scaler = StandardScaler().fit(df[features])
X = scaler.transform(df[features])
model = LogisticRegression().fit(X, df["churned"])

customer = pd.DataFrame([[2, 3, 2]], columns=features)
x = scaler.transform(customer)[0]
contrib = model.coef_[0] * x
log_odds = model.intercept_[0] + contrib.sum()

print("probability:", round(1 / (1 + np.exp(-log_odds)), 3))
for name, c in sorted(zip(features, contrib), key=lambda t: -abs(t[1])):
    print(f"{name:16s} {c:+.2f} log-odds")

"For this customer the model gives a high churn probability.

The biggest push up comes from very low weekly usage,
compared with the average customer. Failed payments push it
up further. Short tenure also adds to it.

I got this by breaking the logistic regression score into
per-feature contributions: coefficient times scaled value.

One caution: this customer is only three months old, and the
model is weaker for short histories, so I would treat this
as a prompt for outreach, not a certainty."

DAY 1  Record the 60-second explanation. No notes.
DAY 2  Listen back. Rewrite the weakest part. Record again.
DAY 3  Explain the top three drivers with your evidence.
DAY 4  Walk through one correct prediction aloud.
DAY 5  Walk through one WRONG prediction aloud.
DAY 6  Explain the limits and 'do not use for' cases.
DAY 7  Full run: 60 s summary + any two parts above,
       to a friend who asks follow-up questions.

When you play your recording, check:

[ ] I said what the model predicts, with a time window
[ ] I grouped features instead of listing columns
[ ] I gave a reason for the model choice
[ ] I named top drivers AND how I measured them
[ ] I used 'odds' correctly if I mentioned coefficients
[ ] I mentioned at least one limit
[ ] I stayed close to 60 seconds for the summary
[ ] No 'it's a black box' anywhere

Questions to practise after your summary:

- Which feature would hurt most if it disappeared?
- Two of your features are correlated. How does that affect
  importance?
- Why did the model get this case wrong?
- What would make you stop trusting this model?
- How would you explain a single prediction to a customer?
- What does an odds ratio of 1.0 mean?
- If you could add one new feature, what would it be?