Data analytics case study Independent portfolio project Fictional data
by Alfonso Cifuentes
Turning a complex publishing catalogue into defensible pricing insights.
This project grew out of a technical assessment I completed during a recruitment process for a data analyst role at an international educational publisher. The task was to structure a product catalogue, compare genuinely equivalent items, and explain which price differences deserved further investigation. I rebuilt the case with entirely fictional titles, series and prices so the approach could be shared publicly.
362 simulated products. 12 fictional series. One question: are the prices different, or are the products different?
An independent reconstruction using fictional data. No employer catalogue or commercial prices are included, and this is not paid or client work.
01 · The brief
Why a publisher can’t just sort its price list and read off the anomalies.
An educational book price list hides a lot. One course can come as a student book, a workbook, an answer-key edition, a digital licence or a bundle, in several levels and editions. Two titles may look similar and still be different products.
Before looking for price anomalies, I needed to recover each product’s identity and define what a fair comparison actually meant. Averaging across all of it gives a confident wrong answer: the mean of this fictional catalogue is €44.04, a figure that describes no product at all.
| Type | Count | Min | Median | Max |
|---|---|---|---|---|
| See type_stats in public/data/analysis.json | ||||
02 · The raw material
Every name and price below was invented by a seeded generator. Same seed, same file, byte for byte.
The raw file has 362 rows across 12 invented series, with two level systems. Each row carries only five fields: series, segment, sub-segment, title and price.
Enrichment adds fourteen attributes the comparison needs. Some can be read from the title; others, like edition and access duration, never appear there and have to come from metadata. That difference matters later.
03 · My approach
Four steps, in this order. Skipping the first is how analysis quietly destroys evidence.
The raw file is never edited. Every correction lives in a separate reporting view that can be switched off, so any number can be traced back to the record it came from.
A rule-based parser reads titles only. It matched the reference labels on level, pack, answer key, book part, audio and digital access for all 362 rows, and abstained on product type twice (“TB”, “Gleebook”) rather than guess. The rules were written for this synthetic catalogue’s own title patterns, so that score checks the rules; it says nothing about accuracy on real data.
Not recoverable from titles: edition · access duration · extra resources · segment
Group products that are the same in every respect except the level, then ask whether the price changes with the level. The formal definition is below.
A price difference is a question for the commercial team, not a verdict. Each exception is classified as a data issue, a plausible pricing decision or an open question, and left in the data so the reader can disagree.
The analytical question
Comparable group: products sharing series, type, role, book part, pack, answer key, both digital rights, extra resources and edition, present at two or more levels. Uniform price: every product in the group has the same list price.
68 groups in total · 60 span two or more levels and can be tested · 8 exist at a single level and are set aside.
Naive substring rule
"ebook" in title
1 false eBook: Fable Road Level 3 Gleebook Resource
Word-boundary rule
\be-?book\b
0 matches. A printed resource stays printed.
04 · The discovery
The interesting errors don’t show where a record belongs. They show where it landed.
Observation, row 20: level B1 appears twice in this group, and one B1 title says Mosaic English. Interpretation: what looks like a €17 price cut is another series’ product filed under the wrong name. Action: keep the source; re-file it in the reporting view. Row 155 is the same pattern (a Glasshouse English study pack filed as Kindred Horizons).
As recorded
54/60
90.0% of comparable groups hold one price
Reporting view
56/60
93.3% · same 60 groups, source file untouched
The denominator was recomputed after re-filing, not assumed: both views yield 68 groups and 60 comparable ones, because each misfiled record fills an empty slot in its real series. The classification was corrected for analysis; no price was changed.
05 · Explorer
Each row is one configuration; read across the levels. A highlighted price differs from the row’s reference price, its most common price, and shows the difference.
Pick a series. The grid shows every fictional product in it, priced by level; nothing is aggregated. Switch to the reporting view to see the two misfiled records move to their own series.
06 · Evidence
Each comparison holds series and level fixed and checks edition labels. Strict and indicative matches are reported separately.
Split volumes
Strict (same series, level and edition label): 10 of 12 levels. Little Lanterns €62.00 vs €54.49 (+13.8%); Wildflower Words +13.9%. Indicative: 2 Little Lanterns levels where parts and volume carry different edition labels show the same gap.
Entry level
Mosaic English and Silverline Skills price Starter student books −€8.00 and workbooks −€4.00 below their other levels (−14.0% to −19.5%). Observation: a consistent pattern. Interpretation: an entry-level policy is plausible. List prices alone can’t confirm intent.
Answer keys
24 of 24 workbook pairs, matched on series, level and edition, share one price. Parity only means something because the match is strict.
07 · Quality register
Stet: let it stand. Query: ask the owner. Normalise: alias in reporting only. Investigate: a pattern without a known cause.
Each issue was planted in the generator on purpose, then found, measured and logged. None was silently fixed.
Rows 20 and 155 carry another series’ name. They create two of the six price “exceptions”. The source keeps the recorded value; the reporting view re-files them by title prefix.
“Second Edition” and “2nd Edition” appear on six Little Lanterns rows. Alias them for reporting. Because edition is part of the comparison key, these rows sit outside every cross-level group. They are six of the 8 single-level groups; the other two are the TB pack and the Gleebook resource, each the only product of its kind.
A printed resource whose name contains the letters “ebook”. A substring rule grants it digital access; a word-boundary rule doesn’t. The record is correct, so it stands. The parser was the problem.
Mosaic English calls its study pack “+ Online Access”; every other series says “with Digital”. Probably the same thing, but merging them is a business decision, not a parsing one.
Row 61, Fable Road Level 1 Pupil Book: segment Primary, sub-segment Secondary. Every other Fable Road row says Primary. A hierarchy error that would skew any segment report.
The only price above the upper IQR fence (about €100). “TB” could mean teacher’s book or something else entirely, so its category stays unresolved; it is not re-classified from its price. Flag it; don’t delete it to tidy a chart.
Six digital workbooks state 18 months; sixty-six say nothing. Blank means not stated, not unlimited. Treating the gap as “no limit” would invent a product feature.
Buying both halves costs about 14% more than the complete volume at every level of both split series. Ten levels are strict matches; at two Little Lanterns levels the parts and the volume carry different edition labels, so those two are indicative only. Possibly intended; worth confirming before anyone calls it an error.
Two series, student book and workbook only. Consistent enough to look deliberate. The data shows the pattern; the reason has to come from a person.
With-key and without-key workbooks always share a price when matched on series, level and edition. A clean result, logged because a looser match would have hidden it.
08 · Limits
Said plainly, so nobody has to infer it.
Not in the data
So there is no recommended price here, no claim about profitability and no estimate of financial impact. The €44.04 mean describes this synthetic catalogue only, not any market. The parser’s perfect score is a check of rules written for synthetic titles, not a production benchmark. And replacing names and prices is not, on its own, a legal opinion on confidentiality. What the analysis can do is hand a commercial team four precise questions:
09 · Videos and files
Everything the page says can be checked in these files. Rerun the generator and you get the same bytes.
Transcripts and captions: walkthrough (VTT) · short version (VTT)
10 · Takeaway
The most striking number is not always the real problem. Two “price exceptions” here were filing errors; correcting the classification, not the prices, changed the reading from 54 to 56 uniform groups out of 60.
What the case shows for an analytics role: structuring messy catalogue data, defining a fair comparison before measuring, keeping lineage from source to report, separating observation from interpretation, and delivering it as reproducible Python, a live-formula Excel workbook and a clear story.