The analytics mindset: from question to insight

Accountants have always analyzed data. A ratio, a variance report, a reconciliation and an aging schedule are all analytics. What has changed is the volume of data, the tools available to work with it, and the expectations of the people who rely on accountants. A mid-sized company now records millions of transactions a year in an enterprise resource planning (ERP) system, and every one of them can be extracted, filtered, joined and summarized in minutes. Managers expect finance to explain not just what happened last month but why it happened and what is likely to happen next. Auditors are expected to look at whole populations rather than small samples where the data allow it. This course teaches the skills to meet those expectations: preparing data, querying it with SQL, analyzing it with spreadsheets and Python, visualizing it honestly, and applying analytics in audit, tax and advisory work.

This first lesson is about mindset rather than tools. The most common failure in analytics is not a coding error or a wrong formula. It is a well-executed analysis of the wrong question, on data nobody checked, presented in a way that does not lead to a decision. A good analyst starts with the decision someone needs to make, works out what evidence would help, and only then opens the data.

The running example

Throughout the course we follow one invented company, Cedarbrook Supply Inc., a wholesale distributor of plumbing, heating and electrical supplies headquartered in Fargo, North Dakota. Cedarbrook sells to about 3,000 contractor and institutional customers through five regional branches, which its reports call North, Central, South, East and West. It buys from about 900 vendors, employs 240 people and had revenue of roughly $86 million in its fiscal year ended December 31, 2025. It runs a mid-market cloud ERP system, reports under US GAAP, and is audited by an invented firm, Halden Rowe LLP. Its controller, Daniel Reyes, has asked a staff accountant with an interest in data, Maya Okafor, to build Cedarbrook's analytics capability. Every company, person and figure in the course is invented for teaching.

Four kinds of analytics

Analytics is often described in four levels, each answering a harder question than the one before. The levels are a convenient way to think about what an analysis is for, not a strict hierarchy: most real projects mix them.

Descriptive analytics is the bread and butter of accounting and should not be dismissed as simple. A correct, well-reconciled description of what happened is the foundation for everything else, and many costly errors come from descriptive work done carelessly: a filter that silently dropped a branch, a join that duplicated rows, a date field read as text. Diagnostic analytics is where accountants often add the most value, because explaining a variance requires knowing how the business works and how transactions are recorded. Predictive and prescriptive analytics borrow more from statistics and operations research. Lesson 10 introduces regression and forecasting; prescriptive methods are touched on only lightly in this course.

The analytics cycle

Several frameworks describe how an analytics project runs. CRISP-DM, the cross-industry standard process for data mining, dates from the late 1990s and is still widely used. Accounting textbooks often teach the IMPACT model: identify the question, master the data, perform the test plan, address and refine results, communicate insights, and track outcomes. This course uses a simple six-step cycle that matches both.

The cycle is iterative. Preparing data often reveals that the question was poorly framed, and the first analysis usually raises new questions. Experienced analysts expect to loop back. What they avoid is skipping steps: jumping straight from a vague request to a chart, or presenting a result without checking that the data were complete.

Step 1: asking a good question

Requests to finance usually arrive as vague concerns. "Margins look soft." "Can you look at our receivables?" "The auditors want analytics." The analyst's first job is to turn the concern into a question specific enough to answer with data. A useful test is to ask what decision the answer will support and what result would change that decision. If no plausible result would change anything, the analysis is not worth doing.

Good analytical questions share a few features. They are specific about the measure, the population and the period: not "are margins soft?" but "which branches and product categories account for the fall in gross margin percentage from fiscal 2024 to fiscal 2025?" They are answerable with data that exist or can be obtained at reasonable cost. They are tied to a decision or a risk. And they come with a sense of what a material answer looks like, so the analyst knows when to stop digging.

It also helps to state hypotheses before looking at the data. If Maya suspects that margins fell because of a price war in one region, rising freight costs, a shift toward lower-margin products, or a costing error, she can write those down and design tests to distinguish them. Writing hypotheses first guards against a well-known trap: looking at enough slices of data until something interesting appears and then telling a story about it. With dozens of branches, categories and months, some slices will look unusual by chance alone.

A first look at the data

Maya starts with a descriptive summary. She takes revenue and cost of goods sold by region for both years, already reconciled to the general ledger, and computes the margin percentage overall and by region. The program below uses pandas, the Python library for tabular data that we use throughout the course. You do not need to follow every line yet; lessons 3 and 4 build these skills step by step. Focus on the logic: compute the overall change, then break it down to see where it comes from.

# Cedarbrook: why did gross margin fall? (invented data)
import pandas as pd
sales = pd.DataFrame({
    "year":    [2024]*5 + [2025]*5,
    "region":  ["North", "Central", "South", "East", "West"] * 2,
    "revenue": [15.2, 22.4, 14.1, 17.8, 11.3, 15.9, 23.1, 14.6, 18.4, 13.9],
    "cogs":    [11.5, 17.0, 10.7, 13.5,  8.6, 12.1, 17.6, 11.1, 14.0, 11.5],
})                                   # $ millions
sales["margin_pct"] = (1 - sales["cogs"] / sales["revenue"]) * 100
total = sales.groupby("year")[["revenue", "cogs"]].sum()
total["margin_pct"] = (1 - total["cogs"] / total["revenue"]) * 100
print(total.round(1))
wide = sales.pivot(index="region", columns="year", values="margin_pct")
wide["change"] = wide[2025] - wide[2024]
print(wide.round(1).sort_values("change"))

      revenue  cogs  margin_pct
year                           
2024     80.8  61.3        24.1
2025     85.9  66.3        22.8
year     2024  2025  change
region                     
West     23.9  17.3    -6.6
North    24.3  23.9    -0.4
Central  24.1  23.8    -0.3
East     24.2  23.9    -0.2
South    24.1  24.0    -0.1

The overall margin fell from 24.1 percent to 22.8 percent, a drop of 1.3 percentage points on revenue of about $86 million, worth roughly a million dollars of gross profit. The regional breakdown shows that the change is concentrated in one place. Four regions moved by less than half a point, but the West region's margin fell from 23.9 percent to 17.3 percent while its revenue jumped from 11.3 million to 13.9 million. The diagnosis has narrowed from "margins are soft" to "something happened in the West, and it came with a lot of new revenue." That points toward the large-contract hypothesis, though it does not prove it. The next step is to drill into West's customers and products, which Maya does in lesson 7.

Two habits stand out in this small example. First, the data were reconciled to the ledger before any analysis: the regional figures add up to the reported revenue and cost of sales. Second, the analysis compared like with like and expressed the change in a way the reader can judge, in both percentage points and dollars. A margin change of 1.3 points sounds small until it is translated into a million dollars.

Professional skepticism and judgment

Analytics does not replace professional judgment; it depends on it. Data can be incomplete, mislabeled or manipulated. A strong pattern can have a boring explanation, such as a change in how a field was coded or a system migration midway through the year. An absence of exceptions can mean that controls work, or that the test was designed so it could never find anything. Skepticism in analytics means asking, at every step, how the result could be wrong.

• Where did the data come from, who controls that system, and could anyone have changed it?
• Does the data reconcile to an independent total, such as the general ledger or a bank statement?
• Is the population complete, or did a filter, a join or an extraction parameter drop records?
• Is the result plausible given what I know about the business? If not, is the business different from what I thought, or is the analysis wrong?
• What other explanations fit the same pattern, and what test would distinguish them?

Judgment also matters in deciding how far to go. Analytics has costs: staff time, software, data access and the risk of drawing conclusions from noise. The aim is not the most sophisticated analysis but the simplest one that answers the question reliably. A pivot table that settles the matter in ten minutes beats a machine learning model that takes a week and that nobody can explain.

Tools of the trade

Accountants use a layered set of tools, and part of the skill is choosing the right one for the job. Spreadsheets remain the most widely used analytics tool in finance because they are flexible, visible and understood by everyone. Their weaknesses are scale, auditability and the ease of introducing silent errors, such as a formula range that misses new rows. Modern spreadsheet features such as Power Query in Excel reduce some of these risks by recording each data preparation step so it can be repeated.

This course uses Python with pandas and SQLite for its worked examples because they are free, run anywhere and make every step explicit, so you can rerun each program and get exactly the output shown. The concepts transfer directly to Excel, Power Query, Power BI, Tableau, and the audit tools that firms license. Where a spreadsheet technique is the natural choice, the lessons describe it.

Why this matters for your career

Data analytics is now part of the core expectations of the profession. Since January 2024, the Uniform CPA Examination has followed the CPA Evolution model, with core sections that test data and technology skills alongside accounting, auditing and tax, and a Business Analysis and Reporting discipline section that covers data analytics in more depth. Firms and employers increasingly hire for the combination of accounting knowledge and data skills, because an accountant who understands both the ledger and the query can find problems that neither a pure accountant nor a pure data scientist would see.

The rest of the course follows the analytics cycle. Lessons 2 to 4 cover acquiring and preparing data. Lessons 5 and 6 teach SQL. Lesson 7 covers descriptive and diagnostic analysis, and lessons 8 and 9 cover visualization and dashboards. Lesson 10 introduces prediction. Lesson 11 applies everything to audit, and lesson 12 covers governance, ethics and artificial intelligence, then ties it all together in a capstone.