Welcome back to Building Probability Models! This is the fifth and final lesson of the course, and you have come a long way. You can now define valid probability models, build both uniform and non-uniform versions, and predict expected frequencies for any number of trials. Those skills let you say what a model expects to happen — but how do you know if the model is actually right?

That is the question we tackle here. In this lesson, you will learn to place model predictions next to real-world data, decide whether the two are close enough to trust the model, and pinpoint why they might disagree. By the end, you will have a complete four-step evaluation workflow — compute, compare, judge, diagnose — that turns a probability model into a genuine decision-making tool.

Before we start comparing numbers, it helps to set realistic expectations. Every chance process contains built-in randomness, so observed results will always scatter around the values a model predicts. If you flip a fair coin $100$ times, the model says to expect $50$ heads — but getting exactly $50$ would actually be somewhat unusual. Landing on $47$ or $53$ is perfectly normal.

The goal of comparing predictions to data is therefore never to demand an exact match. Instead, we ask: "Are the differences small enough to be explained by ordinary randomness, or are they large enough to suggest something is wrong with the model?" Developing good judgment around that question is the central skill of this lesson.

To answer that question, we first need both sets of numbers side by side. The setup involves two simple steps:

1. Compute expected frequencies for each outcome using $P(\text{outcome}) \times n$, where $n$ is the total number of observed trials.
2. Place them next to the observed frequencies collected from the actual data.

Let us try this with a quick scenario. A game shop sells three colors of dice — red, blue, and green — and believes customers choose equally among them. That gives a uniform model with $P(\text{each color}) = \frac{1}{3} \approx 0.333$. After recording sales, the shop lines up the numbers:

Look at the dice shop table above. The expected frequency for each color is $50$, and the observed values are $54$, $47$, and $49$. None match exactly, but every one is fairly close. A practical first check is to compute the difference for each outcome:

To sharpen your judgment, let us look at a case where the data clearly does not fit. A snack machine vendor assumes the three product slots are equally popular, assigning $P(\text{each slot}) = \frac{1}{3}$. After $n = 300$ purchases, the results look very different from what the model predicted:

When predictions and observations disagree, the natural follow-up is: why? There are four common explanations, and learning to distinguish among them is essential for evaluating any model.

• Natural random variation. Even a perfectly accurate model will not predict exact counts. Small differences are expected every time and do not indicate a problem.
• Too few trials. With a small sample, results can swing far from the model's predictions simply because the data has not had enough trials to stabilize. Collecting more data often resolves this kind of gap.
• Outdated or incorrect model assumptions. The model may have been built on old data or assumptions that no longer hold. In the snack machine example, perhaps Slot A was recently restocked with a trendy new item, breaking the "equally likely" assumption.
• Biased data collection. If data is gathered in a way that favors certain outcomes, the observed frequencies will not reflect the true process. For instance, recording snack purchases only during the morning shift might skew results if preferences change throughout the day.

Recognizing which source is most plausible in a given situation is what turns a numerical comparison into a meaningful insight.

Let us bring all four steps together in one walkthrough. A city transit agency models the punctuality of a bus route as follows:

This is a valid model: every probability is between $0$ and $1$, and $$. Over the next trips, the agency collects real data. Here is the full comparison:

In this final lesson of Building Probability Models, you learned a four-step workflow for evaluating a probability model against real data: compute expected frequencies, compare them to observed counts, judge whether the gaps are small enough to chalk up to randomness, and diagnose the most likely source when they are not. The four common sources of discrepancy — natural random variation, too few trials, flawed model assumptions, and biased data collection — give you a practical toolkit for explaining any mismatch. In more advanced statistics, formal tools such as hypothesis tests and confidence intervals can make these judgments more precise, but in this course we focus on clear comparisons between expected and observed counts.

Up next, the practice exercises will walk you through every stage of this process. You will calculate expected frequencies, judge whether data fits a model, match scenarios to their most plausible source of discrepancy, and write a complete evaluation of a real-world model — so jump in and put your new skills to work!