SHAP Analysis: Complete Guide With Sales & Churn Examples

SHAP analysis is one of the most useful techniques for understanding why a machine learning model makes a particular prediction. Instead of simply telling you that a customer is likely to churn or that a sales opportunity is likely to convert, SHAP can show which features pushed the prediction higher or lower.

Why does this matter? According to recent enterprise AI benchmarks, SHAP remains one of the most widely adopted explainability frameworks, utilized by over 60% of organizations implementing model interpretability to satisfy regulatory and transparency requirements in high-stakes fields like finance and healthcare.

This is where SHAP becomes valuable. It connects predictive performance with human-readable explanations, helping data scientists, business analysts, sales teams, and decision-makers understand what a model is actually learning.

In this guide, you will learn SHAP analysis from the ground up, including what SHAP values mean, how to calculate them in Python, how to interpret SHAP plots, SHAP analysis vs feature importance, and how to apply SHAP to practical sales and customer churn problems.

---

What Is SHAP Analysis?

SHAP stands for SHapley Additive exPlanations. It is an explainability method based on Shapley values from cooperative game theory.

In simple terms, SHAP analysis explains how much each feature contributed to a machine learning prediction.

Imagine that you have a machine learning model predicting whether a customer will buy a product.

The model might consider:

• Customer age
• Previous purchases
• Website visits
• Email engagement
• Discount usage
• Customer income
• Number of sales calls
• Days since last purchase

The model may predict:

Customer purchase probability = 82%

But a business manager may immediately ask:

"Why 82%?"

SHAP helps answer that question.

For example:

Feature	Effect on prediction
Previous purchases	+18%
Website visits	+11%
Email engagement	+7%
Discount usage	+4%
Days since last purchase	-6%
Customer age	-2%

The SHAP explanation tells us that previous purchases and website activity pushed the prediction upward, while the long time since the customer's last purchase pushed it downward.

This makes a complex model easier to understand.

The SHAP project describes SHAP values as representing a feature's responsibility for a change in model output. It can also visualize these effects across an entire dataset using plots such as beeswarm plots.

---

Why Is SHAP Analysis Important?

Traditional machine learning answers:

"What will happen?"

SHAP analysis helps answer:

"Why does the model think it will happen?"

This distinction is extremely important in business applications.

Suppose a churn model says:

Customer A has an 89% probability of leaving.

That prediction alone is not enough for a customer-success manager.

The manager needs to know whether the prediction is driven by:

• High support-ticket volume
• Low product usage
• Price increases
• Reduced login frequency
• Contract expiration
• Poor customer satisfaction

SHAP can provide this additional context.

This makes SHAP particularly useful for:

1. Model interpretation
2. Debugging machine learning models
3. Detecting unexpected relationships
4. Understanding customer behavior
5. Sales prediction
6. Customer churn prediction
7. Fraud detection
8. Credit risk analysis
9. Marketing analytics
10. Model governance

For more background on explainability, see Kovendo's guide to Explainable AI (XAI) with examples, which covers SHAP alongside other explainability techniques.

---

How SHAP Values Work

The easiest way to understand SHAP is to think about contributions.

Suppose a model has a baseline prediction of 50%.

For one customer, the model predicts 80%.

SHAP explains the difference:

Baseline prediction = 50%

Then:

• Previous purchases: +15%
• Website activity: +10%
• Email engagement: +7%
• Discount usage: +3%
• Days since purchase: -5%

The contributions approximately add up to the model's output:

50 + 15 + 10 + 7 + 3 - 5 = 80%

The exact mathematical treatment depends on the model output and explainer configuration, but the basic idea is additive attribution.

Each feature receives a SHAP value.

Positive SHAP Value

A positive SHAP value means the feature pushes the prediction toward a higher model output.

For a churn model, this could mean:

Higher support-ticket frequency increases predicted churn.

Negative SHAP Value

A negative SHAP value means the feature pushes the prediction toward a lower model output.

For a churn model:

Long-term customer tenure may reduce predicted churn.

The interpretation depends on what your model is predicting and how its output is defined.

---

SHAP Machine Learning: Where Does It Fit?

SHAP machine learning is not a machine learning algorithm used to make predictions. Instead, SHAP is an explainability framework used to interpret machine learning models.

You first train a model such as:

• XGBoost
• LightGBM
• CatBoost
• Random Forest
• Linear models
• Neural networks
• Other supported models

Then you use SHAP to explain its predictions.

For tree-based models, SHAP provides TreeExplainer, which uses Tree SHAP algorithms to explain tree models and ensembles. The current SHAP documentation lists support for models including XGBoost, LightGBM, CatBoost, PySpark and many tree-based scikit-learn models.

This distinction is important:

Machine learning makes the prediction; SHAP explains the prediction.

---

SHAP Analysis vs Feature Importance

One of the most common questions in explainable AI is SHAP analysis vs feature importance.

They are related, but they are not the same.

Traditional feature importance generally provides a ranking of features.

For example:

Feature	Feature Importance
Previous purchases	0.36
Website visits	0.25
Email engagement	0.18
Age	0.12
Discount usage	0.09

This tells us which features were important to the model overall.

However, it doesn't necessarily tell us:

Did this particular customer's high website activity increase or decrease the prediction?

SHAP can answer that question.

SHAP Analysis vs Feature Importance: Key Difference

Capability	Feature Importance	SHAP
Global feature ranking	Yes	Yes
Individual prediction explanation	Usually no	Yes
Direction of impact	Limited	Yes
Positive/negative contribution	Limited	Yes
Feature interactions	Limited	Can reveal them
Local explanations	No/limited	Yes
Business-friendly explanations	Moderate	High

SHAP's beeswarm plot, for example, displays individual observations as dots and uses their horizontal position to represent SHAP values. Feature value colors can then help reveal whether high or low feature values tend to push predictions in a particular direction.

Therefore, SHAP analysis vs feature importance should not be viewed as choosing one technique universally. Feature importance is useful for a quick global ranking, while SHAP provides a richer view of both global and individual model behavior.

---

Step-by-Step SHAP Analysis With a Simple Example

Let's build a simple customer churn example.

Suppose our dataset contains:

Customer	Tenure	Monthly Spend	Support Tickets	Login Frequency	Churn
A	24	80	1	25	No
B	5	120	8	5	Yes
C	18	70	2	20	No
D	3	150	10	3	Yes

The goal is to predict whether a customer will churn.

Step 1: Install the Libraries

Start by installing the required Python libraries:

pip install pandas scikit-learn xgboost shap

Then import them:

import pandas as pd

import shap

 

from xgboost import XGBClassifier

from sklearn.model_selection import train_test_split

---

Step 2: Create the Dataset

For a simple demonstration:

data = {

    "tenure": [24, 5, 18, 3, 36, 8, 12, 2],

    "monthly_spend": [80, 120, 70, 150, 65, 110, 95, 160],

    "support_tickets": [1, 8, 2, 10, 1, 7, 4, 12],

    "login_frequency": [25, 5, 20, 3, 30, 7, 15, 2],

    "churn": [0, 1, 0, 1, 0, 1, 0, 1]

}

 

df = pd.DataFrame(data)

Here:

• 0 means the customer stayed.
• 1 means the customer churned.

---

Step 3: Separate Features and Target

X = df[

    [

        "tenure",

        "monthly_spend",

        "support_tickets",

        "login_frequency"

    ]

]

 

y = df["churn"]

X contains the input features.

y contains the outcome we want to predict.

---

Step 4: Split the Dataset

X_train, X_test, y_train, y_test = train_test_split(

    X,

    y,

    test_size=0.25,

    random_state=42

)

The training data is used to learn patterns, while the test data is used to evaluate the model.

For a production project, you would normally use a much larger dataset.

---

Step 5: Train the Machine Learning Model

Now create an XGBoost classifier:

model = XGBClassifier(

    n_estimators=100,

    max_depth=3,

    learning_rate=0.1,

    random_state=42

)

 

model.fit(X_train, y_train)

The model now learns relationships between customer behavior and churn.

If you want a deeper understanding of XGBoost, Kovendo also has a practical guide to XGBoost forecasting on sales data.

---

Step 6: Create the SHAP Explainer

Now we move from prediction to explanation:

explainer = shap.TreeExplainer(model)

 

shap_values = explainer(X_test)

The SHAP explainer analyzes how each input feature contributed to the model output.

For tree-based models, TreeExplainer is specifically designed to efficiently explain tree models and ensembles.

---

Step 7: Create a SHAP Summary Plot

The beeswarm plot is one of the most useful SHAP visualizations.

shap.plots.beeswarm(shap_values)

The resulting chart helps answer questions such as:

• Which features matter most?
• Do high feature values increase predictions?
• Do low feature values decrease predictions?
• How much variation exists across customers?

The SHAP documentation explains that the beeswarm plot orders features by their average absolute SHAP impact by default and shows the distribution of individual observations.

---

How to Read a SHAP Beeswarm Plot

Imagine the plot shows:

Support Tickets at the top.

You see many red dots on the right side.

This means:

High support-ticket values are generally pushing predictions toward higher churn.

Now imagine Login Frequency has blue dots mostly on the right.

This could mean:

Low login frequency is pushing churn predictions higher.

The important idea is that SHAP shows both:

Magnitude + Direction

Traditional feature importance usually focuses more heavily on magnitude.

SHAP gives you additional information about how a feature affects individual predictions.

---

Step 8: Explain One Customer

Global explanations tell us what happens across the dataset.

But business users often need a local explanation:

"Why is this particular customer likely to churn?"

You can examine an individual SHAP explanation:

shap.plots.waterfall(shap_values[0])

A waterfall plot breaks down one prediction.

For example:

Base churn probability: 30%

Then:

• Support tickets → +20%
• Low login frequency → +18%
• Short tenure → +12%
• High monthly spend → +3%
• Previous engagement → -5%

Final prediction:

78% churn probability

Now a customer-success manager has a much more actionable explanation.

---

SHAP Analysis for Sales

Sales is another excellent application of SHAP analysis.

Imagine you have a sales conversion model that predicts whether a lead will become a customer.

Your features could include:

• Company size
• Industry
• Number of website visits
• Email engagement
• Sales calls
• Previous purchases
• Demo attendance
• Contract value
• Lead age
• Marketing source

The model predicts:

Lead conversion probability = 76%

The sales manager wants to understand why.

A SHAP explanation could look like:

Feature	SHAP Impact
Demo attended	+0.21
Website visits	+0.14
Previous purchase	+0.11
Email engagement	+0.08
Lead age	-0.06
No recent sales call	-0.09

The salesperson immediately sees that the customer is highly engaged but the lack of a recent sales call is reducing the probability.

That creates an actionable recommendation:

Contact the lead and schedule another sales conversation.

This is much more useful than simply saying:

"The model predicts a 76% probability."

---

SHAP for Sales Forecasting

SHAP can also help analyze sales forecasting models.

Suppose your sales model predicts monthly revenue based on:

• Historical sales
• Product price
• Discount percentage
• Marketing spend
• Season
• Customer count
• Region
• Sales pipeline

The prediction might be:

Forecast revenue = $850,000

SHAP could reveal:

• Historical sales: +$150,000
• Customer count: +$90,000
• Marketing spend: +$50,000
• Discounting: -$20,000
• Seasonal effect: +$30,000

This gives management a much clearer understanding of what is driving the forecast.

For businesses already using machine learning for sales analytics, this is particularly useful because model explanations can become part of management dashboards.

Kovendo's Machine Learning and Data Analytics resources provide related material on machine learning, sales analysis, and data-driven decision-making.

---

SHAP Analysis for Customer Churn

Customer churn is one of the strongest business use cases for SHAP.

Suppose a subscription company trains a churn model.

The model predicts:

Customer 1024 has an 87% churn probability.

Without SHAP, the customer-success team knows the risk but not necessarily the reason.

With SHAP:

Feature	Impact
Support tickets	+0.24
Login frequency	+0.19
Days since last activity	+0.16
Contract age	+0.08
Customer tenure	-0.10

The explanation suggests that inactivity and support problems are the biggest warning signals.

The company could then create an automated retention workflow:

1. Identify high-risk customers.
2. Generate SHAP explanations.
3. Group customers by their main churn drivers.
4. Send the appropriate intervention.
5. Track the response.
6. Retrain the model using new outcomes.

For example:

High support-ticket customers

→ Assign customer-success manager.

Low engagement customers

→ Send product education campaign.

Contract-expiration customers

→ Start renewal campaign.

This turns explainability into an operational tool rather than merely a visualization.

---

Global SHAP vs Local SHAP

There are two important perspectives.

Global SHAP Analysis

Global analysis asks:

"What features generally drive the model?"

For example:

1. Support tickets
2. Login frequency
3. Contract length
4. Monthly spend

This is useful for:

• Model development
• Feature engineering
• Business strategy
• Executive dashboards
• Detecting unexpected patterns

Local SHAP Analysis

Local analysis asks:

"Why did the model make this prediction for this particular customer?"

This is useful for:

• Sales representatives
• Customer-success teams
• Credit decisions
• Fraud investigations
• Individual case reviews

The strongest implementation often uses both.

---

SHAP Interaction Values

Sometimes features do not work independently.

For example:

High monthly spending may not be a churn problem when engagement is high.

But:

High monthly spending + low engagement may create significant churn risk.

This is an interaction.

SHAP can investigate feature interactions using interaction values and related visualizations. The SHAP documentation also exposes interaction-related functionality for tree models.

This is particularly valuable in business models because many real-world relationships are conditional.

For example:

Discount × Customer Tenure

A discount might increase retention for new customers but have little effect on long-term customers.

A simple feature-ranking chart may not reveal this clearly.

SHAP can help investigate these patterns.

---

Common SHAP Plots You Should Know

1. Beeswarm Plot

Best for understanding global feature impact and direction.

shap.plots.beeswarm(shap_values)

2. Bar Plot

Useful for a simpler global ranking.

shap.plots.bar(shap_values)

The SHAP project documents the bar plot as a way to summarize mean absolute SHAP values across features.

3. Waterfall Plot

Best for explaining one prediction.

shap.plots.waterfall(shap_values[0])

4. Scatter/Dependence Plot

Useful for exploring the relationship between a feature's value and its SHAP impact.

shap.plots.scatter(

    shap_values[:, "support_tickets"]

)

This can help reveal whether the feature has a linear, nonlinear, or threshold-like effect.

---

SHAP Analysis vs Feature Importance: Which Should You Use?

The answer depends on the question.

If your question is:

"Which features are generally important?"

Traditional feature importance may be enough.

If your question is:

"How does this feature influence individual predictions?"

Use SHAP.

If your question is:

"Why did this customer receive a high churn score?"

Use SHAP.

If your question is:

"What are the top ten features driving my model?"

Both can be useful.

A practical machine learning workflow is therefore:

Model → Feature Importance → SHAP → Business Interpretation

Start with conventional feature importance for a quick overview, then use SHAP when deeper interpretation is required.

---

Important Limitations of SHAP

SHAP is powerful, but it should not be treated as a perfect explanation of causality.

SHAP Does Not Automatically Prove Causation

If SHAP shows:

High support-ticket volume increases churn prediction.

That does not automatically mean:

Support tickets cause churn.

Customers may submit more tickets because they are already dissatisfied.

Correlation and model attribution are not the same as causal inference.

Correlated Features Can Complicate Interpretation

Suppose you have:

• Annual income
• Monthly income

These variables are strongly related.

How their contribution is divided can depend on the assumptions and configuration used by the explainer.

The SHAP documentation explicitly notes that feature dependence assumptions matter when computing SHAP values.

SHAP Can Be Computationally Expensive

For very large datasets or complex models, calculating and visualizing explanations can require significant computational resources.

For tree models, specialized Tree SHAP methods make the process much more efficient.

Explanation Quality Depends on Model and Data Quality

SHAP explains the behavior of your model.

If your model is trained on:

• Biased data
• Incorrect labels
• Leakage
• Poor features
• Unrepresentative samples

SHAP will not magically fix those problems.

It may actually help expose them.

---

Best Practices for Using SHAP

1. Start With a Good Model

Do not use SHAP to compensate for a poorly performing model.

First evaluate:

• Accuracy
• Precision
• Recall
• F1 score
• ROC-AUC
• Calibration
• Business metrics

Then explain the model.

2. Use Representative Background Data

For explainers that require background data, choose data that represents the population you want to explain.

3. Analyze Global and Local Behavior

Do not rely only on a global SHAP ranking.

Inspect individual cases as well.

4. Look for Unexpected Patterns

If "customer age" suddenly becomes the strongest feature, investigate why.

SHAP can become a valuable model-debugging tool.

5. Don't Confuse Explanation With Causation

Always communicate that SHAP describes model attribution, not necessarily causal relationships.

6. Convert Explanations Into Actions

A good business implementation should answer:

"What should we do next?"

For example:

High churn + low engagement

→ Trigger onboarding campaign.

High churn + support issues

→ Escalate to customer success.

High sales probability + recent demo

→ Prioritize salesperson follow-up.

---

A Practical SHAP Workflow for Businesses

You can implement SHAP in a repeatable process:

Step 1: Collect Data

Gather historical customer, sales, product, or operational data.

Step 2: Prepare Features

Clean missing values, encode categories, remove leakage, and create meaningful features.

Step 3: Train a Model

Use an appropriate model such as XGBoost, LightGBM, CatBoost, Random Forest, or another suitable algorithm.

Step 4: Evaluate the Model

Check predictive performance using appropriate metrics.

Step 5: Create the SHAP Explainer

Use the explainer appropriate for your model.

Step 6: Generate SHAP Values

Calculate feature contributions for your validation or production data.

Step 7: Build Global Visualizations

Use beeswarm and bar plots to understand overall model behavior.

Step 8: Build Local Explanations

Use waterfall plots or other visualizations to understand individual predictions.

Step 9: Connect Explanations to Business Actions

Create rules, workflows, dashboards, or human review processes around the insights.

Step 10: Monitor Over Time

Feature importance and SHAP distributions can change as customer behavior, market conditions, and data distributions change.

---

How SHAP Can Improve Business Dashboards

Imagine an executive dashboard showing:

Revenue Forecast: $2.4M

Instead of displaying only the number, the dashboard could show:

Top positive drivers

• New customer growth
• Pipeline value
• Repeat purchases

Top negative drivers

• Reduced conversion rate
• Higher discounting
• Lower website engagement

Similarly, a churn dashboard could display:

Customers at high risk: 1,240

Then categorize them by dominant SHAP driver:

• 420 → Low engagement
• 310 → Support problems
• 280 → Contract expiration
• 230 → Reduced usage

Now SHAP becomes part of the decision-making system.

It moves machine learning from:

Prediction

to:

Prediction + Explanation + Action

---

SHAP and Explainable AI

SHAP is one part of the broader Explainable AI (XAI) ecosystem.

Other techniques include:

• LIME
• Permutation importance
• Partial dependence
• Individual conditional expectation
• Counterfactual explanations
• Model-specific interpretation techniques

SHAP is especially popular because it provides a unified framework for attributing model output to input features.

Kovendo's Ultimate Guide to Explainable AI provides a broader introduction to XAI, including LIME, SHAP, feature importance, and real-world applications.

---

FAQs

Which is better, SHAP or lime?

SHAP is generally better than LIME because it offers more consistent, theoretically grounded explanations and better global interpretability. LIME is simpler and faster for quick, local explanations.

Is SHAP better than feature importance?

SHAP provides richer explanations because it can show both global feature impact and individual prediction contributions, while traditional feature importance usually provides rankings.

 

Conclusion

SHAP analysis provides a practical way to understand machine learning predictions instead of treating models as unexplained black boxes. It can show which features influence predictions, whether their effects are positive or negative, and why individual customers or sales opportunities receive particular scores.

For sales, SHAP can identify the factors increasing conversion probability. For churn, it can reveal the behavioral signals driving customer risk. For data scientists, it can help debug models and discover unexpected relationships.

The most effective approach is not to replace traditional feature importance with SHAP completely. Instead, use feature importance for a quick global overview and SHAP analysis when you need detailed, actionable explanations.

Ultimately, the value of SHAP is not the visualization itself. The real value comes from turning model explanations into better decisions: prioritizing sales leads, retaining customers, improving products, detecting model problems, and building greater trust in machine learning systems.

As machine learning becomes increasingly embedded in business applications, the ability to explain predictions will become just as important as generating accurate predictions.