From records to decisions
Almost every business produces records without trying: sales receipts, website visits, delivery times, support tickets, inventory counts. Data analytics is the practice of examining such records to answer questions and support decisions. The “science” part is mostly discipline: asking a clear question, checking the data, choosing a suitable method and being honest about uncertainty.
The value is not in the volume of data. A small, accurate table that answers a real question can be worth more than a huge pile that nobody trusts.
The usual steps
- Define the question. “Why did sales fall?” is vague. “Did the number of orders from returning customers fall in the last quarter?” is something that can be checked.
- Collect and combine data. Records often sit in separate systems with different formats. Rented storage and computing from cloud computing has made it easier to gather them in one place.
- Clean it. Duplicates, missing values, inconsistent spellings and obvious errors are common. This stage frequently takes a large share of the effort.
- Explore and analyze. Summaries, charts and statistical methods reveal patterns.
- Communicate. Findings are useful only if decision makers understand them and know how confident to be.
- Act and review. After a decision, check whether the outcome matched expectations and learn from the gap.
Four common types of analysis
Descriptive: what happened? Totals, averages, trends and breakdowns, such as monthly sales by region. Dashboards usually live here.
Diagnostic: why did it happen? Analysts compare groups and look for factors that moved together with the change, such as whether a drop coincided with a price rise or a delivery delay.
Predictive: what might happen next? Forecasts use past patterns, from simple trend lines to machine learning models. Predictions are estimates, not certainties, and they weaken when circumstances change. Related ideas appear in predictive maintenance, where equipment data is used to estimate when a part may fail.
Prescriptive: what should we do? This combines predictions with rules or optimization to suggest actions, such as how much stock to order, always subject to human judgment.
A small hypothetical example
Imagine a hypothetical cafe with 400 customers in a week. Of these, 120 bought a pastry with their coffee. The share buying a pastry is 120 divided by 400, which is 0.30, or 30 percent. Suppose that on days with a morning offer the share was 36 percent, and without the offer 27 percent. The gap is 9 percentage points. That looks promising, but a careful analyst would ask follow-up questions. Were offer days also weekdays with commuter traffic? Were there enough customers on each type of day for the difference not to be chance? Did the offer cut profit per sale? The arithmetic is simple; the interpretation is where the skill lies. These numbers are invented purely for illustration.
Typical methods
- Summary statistics such as averages and medians. The median is often better when a few very large values pull the average up.
- Segmentation, grouping customers or products by shared traits to see how they differ.
- Trend and seasonality analysis, separating long-run direction from regular cycles such as holidays.
- Experiments, such as comparing two versions of a page or process for similar groups, which can give stronger evidence than observation alone.
- Regression and classification models, which estimate how outcomes relate to several factors at once or sort cases into categories.
- Visualization, because a clear chart often reveals what a table hides, though a poorly drawn chart can mislead just as easily.
Common pitfalls
Correlation is not causation. Two measures moving together does not show that one causes the other. A third factor, such as the season, may drive both.
Bad or biased data. If records are incomplete, mislabeled or collected from an unrepresentative group, conclusions will be wrong however sophisticated the method. The same issue underlies algorithmic bias: patterns in historical data can carry past unfairness forward.
Small samples and chance. Results from a handful of cases can swing widely. Seeing a pattern in noise is a very human habit.
Survivorship bias. Looking only at customers who stayed, or products that sold, ignores those that left or failed, which may hold the real lesson.
Overfitting. A model that matches past data too closely can perform poorly on new cases. Testing on data the model has not seen helps.
Cherry-picking and hunting for significance. Checking many comparisons and reporting only the interesting ones makes chance findings look real.
Measuring what is easy rather than what matters. Clicks are simple to count; customer satisfaction is harder. Targets can distort behavior once people aim at the number itself, a risk that appears in algorithmic management of workers.
Ignoring context. A number without knowledge of how the business works can be misread. Input from the people close to the process matters.
Responsible use
Analytics often involves information about people, so it raises data privacy duties. Collecting only what is needed, as in data minimization, storing it safely and being clear about how it is used reduce both risk and mistrust. Where analysis feeds decisions about individuals, such as pricing, hiring or credit, fairness and human review deserve extra attention.
Good analysis is modest about its limits. It narrows uncertainty rather than removing it, and it works best alongside experience and common sense. Claims that data will reveal a guaranteed answer should be treated skeptically.