Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Marketing Data & Analytics Library

Practice Datasets

Every dataset in this section is synthetic: generated from stated rules with a fixed seed, with patterns planted on purpose so that exercises have answers. This hub explains why synthetic data is the right material for practice, how each dataset was built and labelled, which dataset suits which exercise, and the limits that real data imposes and generated data does not.
SectionbeginnerUpdated

Learn this as a path

Prefer a guided order? This learning path reads pages from this section in sequence, with exercises and checkpoints.

In this section

Why synthetic datasets

Learning analysis needs data, and real marketing data is the wrong material for three reasons. It is private: an export from an ad account or an order system contains customer information and commercial detail that cannot be published. It is unexplained: a pattern in real data may be a tracking fault, a seasonal effect, a competitor's launch or noise, and a learner cannot tell which, so an exercise has no answer key. And it is uneven: a real export rarely contains a clean example of every technique a guide teaches, so the learner practises whatever happened to occur. Synthetic data solves all three. It contains nobody's information, every pattern in it was put there on purpose and is documented, and it can be built to contain exactly the cases the exercises need.

The cost is that synthetic data is tidier than reality, and a learner who only ever sees generated tables is unprepared for the duplicate rows, renamed campaigns and missing days that real data contains. The datasets here plant some of that mess deliberately, each page says what was planted, and the limits section of every dataset page is honest about what was not. The parent data library explains the labelling rules: every dataset is labelled modelled, every sample row on this site is illustrative, and the distinction between those labels and sourced evidence is the point of the library.

How the datasets are generated and labelled

Each dataset is produced by a generator with a stated set of rules, a stated seed and a stated list of planted patterns. The dataset page documents all three so that a reader can reproduce the file, predict what an exercise should find and know which anomalies are deliberate.

ElementWhat it meansWhy it is stated
Base ratesThe underlying rates and distributions the generator draws from, by channel or segmentA reader can check whether an exercise result is plausible against the rule that produced it
SeedA fixed starting value for the random drawsThe same generator and seed reproduce the same file; two readers see the same rows
Planted patternsEffects placed deliberately: a seasonal lift, a tracking gap, a creative change, a retention difference between cohortsExercises have answers, and the answer key is the list of patterns
NoiseRandom variation added on top of the base ratesResults look like data rather than arithmetic; small differences are not findings
Planted messDeliberate data-quality faults: a renamed campaign, duplicated rows, a day with missing spendThe reader practises the cleaning the data quality guide teaches
LabelModelled for the dataset; illustrative for any sample rows shown on a pageNothing here is sourced evidence, and the label says so wherever a figure appears

The generation method is described in words rather than code so that it can be reproduced in any language; where a page shows code it is inside a labelled example and carries the same status as the sample rows. A reader who wants to change a base rate and regenerate is encouraged to; the exercises still work, only the answers move.

The datasets

Two datasets cover the two shapes of marketing data most analysis uses. The sample campaign dataset is a single daily table: one row per day per campaign with channel, spend, impressions, clicks, conversions and revenue. It suits the questions a paid media analyst asks, cost per result by channel, return on spend (ROAS) by campaign, conversion rate (CVR) before and after a change, and it contains a seasonal lift, a tracking gap and a creative change for the reader to find. The sample e-commerce dataset is three related tables, customers, orders and order items, built for the questions a retention or ecommerce analyst asks: cohort retention by acquisition month, average order value by channel, repeat purchase rate, and the join patterns that produce them without double counting. The two datasets share channel names so that spend from the first can be matched to customers in the second for a simple acquisition cost exercise.

Both are small by design: small enough to open in a spreadsheet and inspect by eye, large enough that aggregates are stable and a window function has something to work on. They are not meant to test performance, and nothing about their size reflects any real business.

Exercises

Exercises are listed on each dataset page in increasing difficulty, and the sequence below is a reasonable path through both. Start with aggregation on the campaign dataset: spend, conversions and cost per result by channel for the whole period, then by month, which is the first pattern in the SQL for marketers guide. Move to comparison over time: find the month where conversions rose without a matching rise in clicks, and decide from the planted-pattern list whether it was the creative change or the seasonal lift. Then find the tracking gap, which is a stretch of days where spend continues and conversions do not, and practise the reasoning the data quality guide teaches about distinguishing a market change from a measurement fault.

On the e-commerce dataset, begin with the join: orders to customers, then order items to orders, checking row counts at each step so that fan-out is caught. Compute average order value by acquisition channel. Then build the cohort table, customers grouped by first order month with the share ordering again in each later month, which is the method the cohort analysis guide teaches and the planted retention difference between two channels is there to reward. Finish by joining the two datasets: spend by channel from the first, new customers by channel from the second, and a cost per new customer that the generator's base rates allow you to check.

Illustrative example — not a real result

A first week with the datasets for a junior analyst

A junior analyst at a garden supplies retailer is given both datasets on her first week. On the campaign table she finds that paid social spent £14,200 for 410 conversions and paid search £9,800 for 520, and writes down the cost per conversion for each before reading the planted-pattern list. She then charts conversions by day and spots a nine-day stretch in week six where paid search spend continues at around £110 a day while conversions drop to zero, which the list confirms is the planted tracking gap. On the e-commerce tables her first join triples the order count because she joined items before aggregating; she repeats it with orders aggregated first and gets 2,640 orders, matching the row count. Her cohort table shows the March cohort from email retaining 31 percent of customers into month three against 18 percent for paid social, which is the planted retention difference. By Friday she has a note of each finding and which pattern explains it, which is the exercise working as intended.

Common dataset problems

Most difficulties with the datasets are either a method problem the guides address or a misreading of what synthetic data is for. The table maps each to the resource that resolves it.

ProblemWhat is happeningResource
Row counts multiply after a joinFan-out: joining a many-row table before aggregatingSQL for marketers, the join exercises on the e-commerce dataset
Cohort shares do not sum sensiblyCohort defined on order month rather than first order monthCohort analysis guide
A metric is quoted as a benchmarkModelled data read as evidenceThe labelling rules in the data library
The tracking gap looks like a market dropMeasurement fault mistaken for a real changeThe planted-pattern list on the campaign dataset
Results differ from a colleague'sDifferent seed or a regenerated fileThe generation section of each dataset page
The data feels too cleanIt is; the planted mess is limited and listedThe limits section of each dataset page

Where an exercise result cannot be explained by any planted pattern, the usual cause is a query error rather than a discovery, and checking row counts at each step finds it faster than re-reading the data.

Terms used on this page

Related