Predictive Marketing – What Can Actually Be Predicted?

A map of ML tasks in social media marketing: engagement, CTR, conversion, churn, LTV, sentiment – and the question of when ML really pays off and when a lean heuristic gets you there faster.

1. The Problem: Prediction ≠ Reporting

Marketing teams are surrounded by data – impressions, reach, click streams, engagement time series. The bulk of it ends up in dashboards. Dashboards are descriptive analytics: they show what was. Predictive analytics shifts the question into the future: What will happen if nothing intervenes? Only that question justifies building an ML model.

"If you can't measure it, you can't improve it." – Peter Drucker (often attributed)

The following table cleanly separates the two worlds:

2. The Map of Predictive Tasks

Six task families cover most of the marketing demand:

• Engagement prediction (likes, shares, comments) – binary classification high_engagement ∈ {0,1} or regression on a target metric.
• CTR prediction for ads – heavily imbalanced (positive rate often <3 %), calibrated probabilities are mandatory.
• Conversion / lead scoring – classification on a rare event, often with cost-sensitive thresholds.
• Churn prediction – "Who will cancel next month?" A classic binary setup.
• Customer lifetime value (LTV) – regression on long-term revenue; often framed as survival analysis.
• Sentiment / topic classification – NLP tasks on texts, reviews, comments.

3. When Does ML Pay Off – and When Is a Heuristic Enough?

ML is justified by three conditions, all of which must hold:

1. Labels exist in sufficient quantity. Rule of thumb: at least a few thousand examples per class.
2. The signal is stable over time. For viral, one-off phenomena the model learns noise.
3. A better decision delivers measurable value. If the business impact of a 1 % improvement is in the cent range, the effort is not worth it.

❌ Wrong question

"Can we use ML to figure out which posts will go viral?"

Viral posts are long-tail events. There are not enough examples, the signal is unstable, and classes are extremely imbalanced.

✅ Right question

"Can we predict which posts will land in the top 20 % of the engagement distribution?"

A well-defined binary label, enough examples, an evaluable metric (PR-AUC, lift).

4. Data Exploration as the First Step

Before training a model, an honest exploratory analysis pays off. The following snippet generates a synthetic social media dataset with realistic correlations and inspects the most important features.

import pandas as pd
import numpy as np

rng = np.random.default_rng(42)
n = 1000

# Bake in realistic correlations: posts with an image get more likes,
# posts during prime time (18-22h) too.
hour = rng.integers(0, 24, n)
has_image = rng.integers(0, 2, n)
text_len = rng.integers(20, 280, n)

prime_time = ((hour >= 18) & (hour <= 22)).astype(int)
base_likes = 5 + 8 * has_image + 6 * prime_time
likes = rng.poisson(base_likes)

df = pd.DataFrame({
    "hour": hour,
    "has_image": has_image,
    "text_len": text_len,
    "likes": likes,
    "high_engagement": (likes > np.quantile(likes, 0.8)).astype(int),
})

print("Target distribution:")
print(df["high_engagement"].value_counts(normalize=True).round(3))
print()
print("Mean likes by image usage:")
print(df.groupby("has_image")["likes"].mean().round(2))
print()
print("Mean likes by time-of-day bucket:")
df["time_bucket"] = pd.cut(df["hour"], bins=[0, 6, 12, 18, 24],
                           labels=["Night", "Morning", "Afternoon", "Evening"])
print(df.groupby("time_bucket", observed=True)["likes"].mean().round(2))

The output makes two things visible. First, high_engagement is at ~20 % by construction – that is the positive rate the later model has to deal with. Second, the two main drivers (has_image, prime time) are immediately recognisable in the group means. This kind of validation before training prevents a complex model from later revealing what a two-line heuristic could already have delivered.

5. Discussion: Prediction vs. Causality

One point is permanently underestimated in practice: prediction is not causality. A model that predicts with 90 % accuracy who will click says nothing about who clicks because of the ad. This gap is addressed explicitly in article 04 (uplift modelling).

A second, often overlooked point: data drift is the norm, not the exception, in social media. Platform algorithms change, user behaviour shifts seasonally, viral trends overwrite patterns. Every production model needs monitoring that detects drift – and a plan for regular retraining.

6. Wrap-up Quiz

7. Further Reading

• Next article: Engagement prediction with classical models (article 02 of this series).
• Cross reference: category recommendation-systems.
• Knowledge base: machinelearningwithpytorchandscikit-learn.md – chapter on classification and evaluation metrics.

Sources

• Machine Learning with PyTorch and Scikit-Learn (Sebastian Raschka et al.) – chapter 6 (Model Evaluation).
• Generative AI Foundations in Python – section on predictive vs. descriptive analytics.