Explain One Prediction in Plain English

October 2026 · 7 min read

We have a fitted equation, verified calculations, and plots. The final step is to answer the reader’s actual question. We will use Maya throughout and choose the comparison before choosing the words.

Keep the facts visible

Return to the original fitted model, not the curved or interaction variants:

The reference predictions average 35. Maya’s observed future spending is not yet known. Every sentence below explains a prediction of this fitted model.

Question 1: why does the model predict 50 for Maya?

This is a question about the whole number. Start from the intercept and show the parts:

The model starts at 20 dollars. Your four visits add 20, and your Premium plan adds 10. Together, that gives a prediction of 50 dollars.

The explanation contains both inputs and all three terms. It does not require the reader to know what a coefficient or intercept is. If they want the arithmetic, show 20 + 5 × 4 + 10 × 1 = 50 directly below it.

Question 2: why is Maya above the reference average?

Now the comparison has changed. We need the SHAP calculation from Part 8, which starts at 35 rather than 20:

The four reference customers have an average prediction of 35 dollars. You made four visits compared with their average of two, which adds 10. Your Premium plan adds another 5 relative to that group. Your prediction is therefore 50 dollars.

The plan difference is 5 because half the reference group already has Premium. Saying “Premium adds 10” in this explanation would mix the raw equation’s reference with the reference population’s average.

This is an example of choosing reasons that answer the question. Visits is the larger reason for Maya’s 15-dollar difference from average. That is a more precise statement than calling visits the most important feature in every situation.

Question 3: why is Maya above Dev?

Both customers have Premium. Maya made four visits; Dev made three. Subtract their equations:

Your prediction is 5 dollars higher than Dev’s because you made one more visit. You both have Premium, so the plan does not explain the difference.

Premium contributes to both predictions, but cancels in this comparison. A feature can have a large contribution to an individual prediction and still explain none of the difference between two people.

For Maya versus Cara, both visits and plan differ: 5 × (4 − 3) + 10 × (1 − 0) = 15. The matching sentence is: “One extra visit adds 5, and Premium adds 10, so your prediction is 15 above Cara’s.”

Question 4: what would the model predict for Basic?

Keep Maya’s four visits and change only the plan:

With the same four visits and a Basic plan, this model would predict 40 dollars, which is 10 lower.

The phrase “this model would predict” keeps the meaning clear. We have changed an input to a fitted rule. We have not shown that changing plans would cause Maya’s real spending to fall by exactly 10 dollars.

Question 5: how does the model work overall?

This asks for the model’s general behavior rather than Maya’s specific result:

The model uses visits last month and the current plan to predict next month’s spending. Each additional visit raises its prediction by 5 dollars. At the same visit count, Premium raises the prediction by 10 dollars.

That sentence applies to our original model. For the interaction model, it must change: each extra visit adds 5 for Basic and 7 for Premium. Reusing the original sentence would hide the feature interaction we deliberately added.

Choose a useful amount of detail

Lead with the prediction and the most relevant reasons. For this tiny model, both inputs fit comfortably into a short explanation. In a larger model, a few main reasons may be enough for the first sentence; if a numeric breakdown omits smaller terms, collect them in a labeled remainder so the sum still reaches the prediction.

A generally influential feature is not automatically the best reason for a particular customer. A normally small term may be unusually large for that person. For a comparison between two customers, focus on the terms that differ. The individual-contribution plot in Part 7 helps identify these cases.

Use the feature’s measured meaning. “Four visits last month” is more informative than “high engagement,” and does not invent a motive such as loyalty or willingness to spend. A good explanation lets the reader follow the model’s reasoning without needing to accept an unsupported story.

Explain the result and its accuracy separately

We can account for every dollar in Maya’s prediction. That does not make the prediction certain. On the eight separate evaluation rows in Part 6, the original model had an average absolute error of 1.5 dollars.

This is a useful description of that small evaluation sample. It is not a guarantee that Maya’s error will be at most 1.5, and it is not a prediction interval. If a reader asks about accuracy, give the evaluation result in its own sentence rather than adding certainty to the explanation of the equation.

A complete short explanation

For Maya’s question about being above the reference average, the final version can be this simple:

Your predicted spending is 50 dollars. That is 15 above the average prediction for our four reference customers: your extra visits account for 10, and your Premium plan accounts for 5. The prediction uses last month’s visits and your current plan.

Every number comes from the visible table and the fitted equation. For a different question, change the comparison first, recalculate the differences, and then write the sentence. You can reproduce the entire series with the Python lab, or continue into the Shapley Values and SHAP series to explain more complex models.

Sources and worked examples

This chapter develops ideas from the lessons below in A Data Odyssey’s Explainable AI playlist. The customer data, calculations, figures, and interactive examples here are original teaching examples. The creator’s companion notebooks use different datasets.

• Explaining Machine Learning to a Non-technical Audience — Conor O’Sullivan, A Data Odyssey.