Most failed ML projects fail at framing, not modelling: the wrong target, a metric nobody cares about, or features that won't exist at prediction time. Turn the business goal into a precise prediction task first.

Business objective
What decision will the prediction change? “Reduce churn” → “which customers should get a retention offer this week?”
Prediction target
Define y exactly, including time: churned = no purchase in the 30 days after the prediction date.
Prediction point & features
Only use information available at the moment the prediction is made. Everything after it is leakage.
ML task
Binary classification, regression, ranking, clustering…
Metrics
An offline ML metric (PR-AUC, RMSE) that tracks an online business metric (retained revenue); set a success threshold up front.
Baseline
A trivial model or a simple rule. A model is only useful if it clearly beats it.
Constraints
Latency, cost, interpretability, fairness, privacy, how often labels arrive.

Common task types

Task	Predicts	Example	Typical metrics
Regression	A number	Ride price from distance and time	MAE, RMSE, R²
Binary classification	One of two classes	Is this transaction fraud?	Precision, recall, F1, PR-AUC, ROC-AUC
Multi-class classification	One of k classes	Which product category?	Accuracy, macro-F1, log-loss
Multi-label classification	Any subset of k labels	Tags for an article	Per-label F1, Hamming loss
Ranking	An ordering of items	Search results for a query	NDCG@k, MRR, MAP
Recommendation	Items a user will like	Movies to show on the home page	Recall@k, NDCG@k, online CTR / watch time
Forecasting	Future values of a time series	Next week's demand per store	MAE, RMSE, sMAPE/MASE
Clustering	Groups of similar rows (no labels)	Customer segments	Silhouette; usefulness to the business
Anomaly detection	Unusual rows	Faulty sensor readings	Precision@k, recall on known incidents

Always start with a baseline

On the breast-cancer dataset (63% benign), a “model” that always predicts the majority class already gets 63% accuracy. A real model has to beat that, not 50%.

Python

from sklearn.datasets import load_breast_cancer
from sklearn.dummy import DummyClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler

X, y = load_breast_cancer(return_X_y=True)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.25, stratify=y, random_state=0)

baseline = DummyClassifier(strategy="most_frequent").fit(X_tr, y_tr)
model = make_pipeline(StandardScaler(), LogisticRegression()).fit(X_tr, y_tr)
print(round(baseline.score(X_te, y_te), 3))   # 0.629 ← always predicts "benign"
print(round(model.score(X_te, y_te), 3))      # 0.958

Good baselines: the majority class or the mean (DummyClassifier, DummyRegressor), yesterday's value for a time series, a hand-written rule the business already uses, or a small logistic/linear regression.

Worked example: a spam filter

Decision	Choice	Why / caveat
Task	Binary classification: spam vs not spam	Output a probability so the threshold can be tuned
Input	Subject, body text, sender history, links, sending IP reputation	All available when the email arrives
Labels	“Report spam” / “Not spam” clicks plus a sample reviewed by annotators	User reports are biased and noisy: most spam is never reported and people also “report” newsletters they signed up for
Metric	Precision at a fixed recall, or recall at ≥ 99.9% precision	A legitimate email lost in the spam folder (false positive) is far worse than one spam in the inbox
Baseline	Current blocklist and keyword rules	The model must beat what already exists
Constraints	Milliseconds per email; spammers adapt	Retrain often and monitor drift

Framing traps

Proxy targets — optimising clicks when you care about satisfaction rewards clickbait.
Features from the future — days_since_last_login computed at training time over the whole history, but at prediction time only up to today.
Feedback loops — a model that decides who gets a loan only ever sees outcomes for approved applicants.
No action attached — a churn score nobody acts on has zero value, however high its AUC.

How would you frame a churn prediction problem?

Your model has 95% accuracy. Is it good?