If you enjoy explaining what changed in a business metric and tracing it to something the team can investigate, start with business data analysis. If you prefer defining a prediction target and checking against later data to see whether your approach could help the team make decisions, try the modeling side of data science. Before applying, ask which deliverables the team expects you to own over time.
Some employers combine both kinds of work. Tapai Group's recruitment for 2027 graduates includes a “Data Scientist (Data Analyst)” position that covers metrics and BI dashboards as well as forecasting and anomaly detection. This example from a manufacturer shows why job titles and tools alone cannot draw a firm boundary. In this article, “analyst” means a business analysis or business intelligence role, while “data scientist” refers to work focused on validating predictions. Pinduoduo's 2027 graduate data analyst posting also asks for SQL, Python, and mathematical statistics. When a posting mentions dashboards, metrics, or business reviews, ask who uses and signs off on the work. For modeling, forecasting, or anomaly detection, ask what gets predicted and how the team tests it. The two roles can use the same tools. Tapai Group campus recruitment: expand the full data scientist job description to find position 3 · Pinduoduo graduate data analyst posting via Nankai University · IBM on data science and data analytics
| How the work divides around one business problem | Business analysis | Predictive validation |
|---|---|---|
| What the business asks | How much did net sales decline? Which channel accounted for most of the decline? | At order time, can we prioritize orders for manual review? |
| What you analyze | Sales by channel and confirmed refunds for two order weeks | A later batch of orders and their subsequent refund outcomes |
| Who gets the work | Operations or finance checks the figures and arranges further investigation | The review lead checks the queue. Data and engineering colleagues verify the inputs and results |
| What you deliver | A breakdown someone else can recalculate and a follow-up recommendation with an owner | A reproducible comparison of rankings and a recommendation to keep the approach or test it again |
| How it gets checked | Groups add up to the total, calculation rules are consistent, and conclusions stay within the evidence | Predictions use only information available at the time, and both approaches have the same review capacity |
Start With the Same Order Table and Define Both 14-Day Windows
This is a small fictional exercise. Each row is one order. It leaves out cancellations, return shipping, and partial refunds paid in installments. Use the confirmation date to determine whether a confirmed refund counts. In this exercise, the order date is day 0, and the 14-day window covers days 0–13 by calendar date in Beijing time. Copy the table into a spreadsheet for both exercises.
| Order | Order date (example) | Channel | Promotion coupon used | Order amount (yuan) | Refund confirmation date | Confirmed refund amount (yuan) |
|---|---|---|---|---|---|---|
| A1 | March 2 | Search | Yes | 360 | March 7 | 50 |
| A2 | March 2 | Search | No | 240 | — | 0 |
| A3 | March 2 | Other | Yes | 230 | — | 0 |
| A4 | March 2 | Other | No | 170 | — | 0 |
| B1 | March 23 | Search | Yes | 300 | March 28 | 90 |
| B2 | March 23 | Search | No | 260 | — | 0 |
| B3 | March 23 | Other | Yes | 240 | April 2 | 20 |
| B4 | March 23 | Other | No | 160 | — | 0 |
A and B are complete synthetic samples for two order weeks. The table contains all orders placed in those weeks. In business analysis, the 14-day window determines how much to subtract for each order before totaling results by order week. In predictive validation, it determines whether each order had a confirmed refund, a yes-or-no outcome. B3 subtracts 20 yuan in the business report. In the validation exercise, it gets the same “yes” label as B1, which had a 90-yuan refund. The prediction target is which orders are more likely to be refunded. It does not measure the size of the actual loss.
Exercise 1: Give Operations a Report They Can Reconcile
Calculate net sales as order amount minus confirmed refunds within 14 days. Assume you review the data on April 7, after both observation windows close and all refund records are in. Check the totals by channel:
| Channel | Week A order amount − refunds | Week B order amount − refunds | Change from A to B |
|---|---|---|---|
| Search | 600 − 50 = 550 | 560 − 90 = 470 | −80 yuan |
| Other | 400 − 0 = 400 | 400 − 20 = 380 | −20 yuan |
| Total | 1,000 − 50 = 950 | 960 − 110 = 850 | −100 yuan |
The 80-yuan decline in Search breaks down into 40 yuan less in order value and 40 yuan more in confirmed refunds. That gives operations a concrete follow-up: “Check the mix of order values in Search first. Ask after-sales support to verify why B1 was refunded, look for the same reason in other channels, and recalculate next time using the same window.” Search accounts for 80% of the decline in this example. That figure is an arithmetic breakdown. It cannot establish whether search ads, coupons, or actions by operations caused the drop.
Submission check: Include a table by group, one line explaining the calculation, and two sentences of recommendations. Ask a classmate to remove B3's refund and recalculate. The Other channel's net sales should return to 400 yuan and the overall decline should shrink to 80 yuan. If the channel detail changes but the total stays the same, fix the formula. In a real project, if a week's refund records arrive late, flag the incomplete data and explain the impact before deciding when to issue a report with comparable figures.
Exercise 2: Give the Review Team a Ranking Comparison With a Time Cutoff
Treat March 20 as the proposed rollout decision date. The observation window for week A has closed and no week B orders have been placed. Assume the confirmed week A labels are already in the database. Use week A to discuss the rules, then lock in two simple approaches before seeing the week B outcomes: the existing rule sorts orders from highest to lowest amount. The candidate puts coupon orders first and sorts by amount within each group. Coupon use and order amount are available at checkout. Refund date and amount are off limits when ranking. In a real project, check when each field actually becomes available in the database. A date field alone does not establish that the data was available at the time. Coupon-first is a candidate rule for learning how to compare rankings. One of the two coupon orders in week A was refunded and one was not. Four orders cannot establish whether coupon use has predictive value. scikit-learn documentation on time-series splits
Assume reviewers can check only two orders from this simulated batch. Once all week B outcomes are available, compare the same orders with the same number of review slots:
| Rule | First two orders | Which were later refunded | Hit rate among the first two |
|---|---|---|---|
| Order amount | B1, B2 | B1 | 1/2 |
| Coupon first | B1, B3 | B1, B3 | 2/2 |
There is no model training here. If B3 had no refund, both queues would flag only B1 correctly. Even if the candidate rule flags one more order that ends up being refunded, ask three questions before piloting it. How many of those refunds could manual intervention have prevented? How many normal orders would it flag incorrectly, and how much reviewer time would that take? Does the difference hold on a new batch? Four orders illustrate how to run the comparison. On later batches, you could run both rankings without acting on them, keep review capacity fixed, and check hits and false positives. Then evaluate the intervention separately to measure the benefits and costs of manual review.
Submission check: Show the two queues and their hits side by side. Record the decision date, which fields are allowed, and which conclusions need another batch to test. Ask a classmate to remove the refund columns and recreate both queues. If they cannot, the rules may be incomplete or depend on fields unavailable at prediction time.
See How the Two Assignments Look in Code
Import the eight orders into an SQLite table called orders. Use order_id, order_day, channel, amount, refund_day, refund_amount as the columns. Enter dates as YYYY-MM-DD (use 2026 for this example) and leave the refund confirmation date empty when an order has no refund. Grouped by order week and channel, the expected net sales are 550 yuan for Search and 400 yuan for Other in week A, and 470 yuan for Search and 380 yuan for Other in week B.
SELECT substr(order_id, 1, 1) AS cohort, channel,
SUM(amount - CASE
WHEN refund_day >= order_day
AND refund_day < date(order_day, '+14 days')
THEN refund_amount ELSE 0
END) AS net_amount
FROM orders
GROUP BY 1, 2
ORDER BY 1, 2;For the ranking, give Python only the week B order ID, coupon status, and amount available when each order was placed. The rules were agreed on March 20. Orders enter the simulated batch only after they are placed. Once you have both rankings, check the hits against the refund outcomes you stored separately.
orders = [
("B1", True, 300), ("B2", False, 260),
("B3", True, 240), ("B4", False, 160),
]
baseline = sorted(orders, key=lambda row: -row[2])[:2]
candidate = sorted(orders, key=lambda row: (not row[1], -row[2]))[:2]
print([row[0] for row in baseline]) # ['B1', 'B2']
print([row[0] for row in candidate]) # ['B1', 'B3']These examples show what each assignment produces and when you can use the data. Check the team's job description and ask in the interview which tools they use and who owns the work.
Use the Exercises to Decide Where to Apply
Give each assignment the same amount of time. Notice where you get stuck and which feedback you are more eager to follow up on. Then ask your target team, “What's the first assignment a new hire owns from start to finish? Who signs off on it? How often does it come up?” If the team mainly evaluates metric definitions and whether the business uses recommendations, lead with the business report. If it asks more about baselines, when data becomes available, and when a method stops working, lead with the ranking validation. If the job needs both, ask how your first three months would be split and who can coach you on the areas you know less well. One practice exercise gives you questions to test against the job.
Compare the broader work in FermatMind's business intelligence analyst profile and data scientist profile. As you prepare to apply, use this guide to comparing two job postings to check what roles in your city actually require. This job search guide for applicants whose major differs from the role can help you present practice work separately from real business results.
Frequently Asked Questions
Can I Apply for Data Analyst Roles Without Knowing Machine Learning?
Start with roles centered on business metrics, retrieving data, analysis, and visualization. Use a report others can recalculate to show how you think. Then check whether the job explicitly requires experimentation, statistical modeling, or forecasting, and fill gaps based on the posting.
Do Data Scientists Train Models Every Day?
The work can include defining the problem, checking data, comparing baselines, investigating errors, and discussing what has to be true before rollout. Ask what the team delivers routinely, who reviews it, and how much time goes to model validation.
Can I Use the Same Exercise to Apply for Both Roles?
Yes. For business analysis, show the breakdown, calculation rules, and recommended follow-up. For modeling-oriented roles, show the timeline, baseline comparison, and plan for testing the next batch. In both versions, state that the data is fictional and the work was never deployed.


