Descriptive Statistics

This is the first step in using data to make decisions. Everything else in statistics—tests, models, even machine learning—starts with understanding what your data looks like. By the end of this page, you'll be able to summarize any list of numbers (like exam scores, product ratings, or sensor readings) with a few key figures: the typical value, how spread out the values are, and whether there are any unusual extremes. This is the foundation for later concepts like hypothesis testing, cross-validation, and even understanding how AI models evaluate their own performance. Without descriptive statistics, you're just staring at a wall of numbers.

Imagine you're handed a list of 10,000 numbers: the daily sales figures for a shop over about 27 years. You can't look at every single number and conclude anything. You need a way to compress all that information into a few digestible facts. That's descriptive statistics. It's the art of summarising a dataset's central tendency (where the 'middle' is), spread (how far from the middle things get), and shape (is it symmetric, or does it lean one way). It tells you the story of the data without reading every line. The mean (average) tells you the typical sale, but it doesn't tell you if some days are wildly different. The standard deviation tells you how much sales typically fluctuate. The median tells you the middle point, which is more robust if a few days were extraordinary. So you walk away with a mental picture: 'Sales average ₹50,000, typically swing by ₹10,000, and there's one weird day where someone bought a hundred refrigerators.' That's the whole point—turning chaos into insight.

Think of a restaurant owner trying to understand how many customers come in each day. They scrawl the day's count on a calendar. After a month, they have 30 numbers. To make sense of it, they could do a few things. First, they could find the average: 'About 80 customers a day.' That's the mean. But they might also notice that most days are around 60-70, and only on weekends it jumps to 150. The average of 80 is pulled up by those busy weekends. So they also look at the median—the middle count when they line up all 30 days in order. That's 65. 'Ah, the typical day is 65 customers,' they say. Then they want to know how much variation there is. Are there quiet days with 20 and packed days with 200? The standard deviation tells them, 'On average, a day's count is about 25 customers away from the mean.' That's the spread. Now, the owner also wants to know if there are any weird days—maybe the day they were closed for renovation (0 customers) or the day of a festival (300). These are outliers, and the median is more resilient to them than the mean. Where does the analogy break down? The owner is doing this on a small scale with simple arithmetic. A real dataset might have millions of rows, and we use software to do exactly the same calculations. But the intuition—the compression of a big pile of numbers into a few meaningful figures—is exactly what descriptive statistics is.

Descriptive statistics is the practice of summarising a dataset—its central tendency, spread, and shape—into a small set of numbers that describe the whole, without making inferences about a larger population.

This concept is the first rung on the ladder of statistical inference. You have no prior knowledge to build on, so we start from scratch. Once you master descriptive statistics, you'll move on to its parent concept: Statistical Inference. That's where you use a sample to say something about a whole population, with a measure of how confident you are. Descriptive statistics is the observation part; inference is the conclusion part. For example, you might use descriptive statistics on a sample of 100 users to find their average session length, then use inference to claim 'all users on average have a session length in this range.' Your library also lists neighboring topics like Sufficient Statistics and Cross-Validation Partitions. The idea of a sufficient statistic is that a well-chosen summary (like the mean) captures all the useful information in the data—that's a formal extension of what you'll learn here. Cross-validation, used to evaluate machine learning models, relies on partitioning a dataset and computing statistics (like accuracy) on each fold. So this basic toolkit is the precursor to everything else.