Do popular product pairs stay popular? A market basket analysis example

Only seven of the original 20 product pairs appeared in the next year’s independently ranked top 20. That is 35.0% of the shortlist. Yet both products in every original pair still appeared in later eligible invoices, and every pair still met the same minimum counts for inclusion in the ranking.

These observations answer different questions. Staying in the top 20 means a pair remains near the front of a competitive list. Appearing together means there is still evidence of co-purchase. A product pair can meet the second condition while falling well down the ranking.

We selected the original pairs using invoices dated December 1, 2009 through November 30, 2010. We then kept that list unchanged and followed the same product codes through December 1, 2010 to November 30, 2011. The later top 20 was selected afresh from all eligible pairs in the later year, including pairs outside the original list.^[2]

The comparison uses 19,645 eligible invoices in the first year and 18,877 in the second. Invoices with only one qualifying product remain in those totals. Before exploring the complete lists below, consider what happened to the first year’s leading pair.

The first year’s leading pair was the Sweetheart ceramic trinket box, product 21231, and the Strawberry ceramic trinket box, product 21232. The analysis plan specified the original number-one pair as the worked example before its later results were calculated.

The two boxes appeared together on 766 of 19,645 eligible invoices in the first year, or 3.9%. In the later year they appeared together on 320 of 18,877, or 1.7%, and ranked 153rd among eligible pairs. The pair had moved far down the list while still appearing together hundreds of times.

Now narrow the question to invoices containing the Sweetheart box. In the later year, 492 eligible invoices included it, and 320 of those also included the Strawberry box. Dividing 320 by 492 gives 65.0%. That percentage describes co-purchase within the Sweetheart-box invoices. It does not describe the share of all invoices containing both products.

Reversing the question changes the denominator. The Strawberry box appeared on 791 later invoices; 320 also contained the Sweetheart box. The result is 40.5%. Neither direction establishes which item was chosen first. These are two views of the same unordered pair.

The product counts also changed. The Sweetheart box appeared on 960 eligible invoices in the first year and 492 in the second; the Strawberry box appeared on 1,582 and 791. Those counts supply context for the lower joint count. The records do not explain why the products appeared less often.

For a merchant, this separates two decisions that can otherwise be blurred. A broad campaign may need to consider a pair’s share of all eligible invoices. A suggestion shown beside one specific product raises a narrower question about invoices already containing that product. Historical co-purchase can inform which idea to examine, but neither percentage estimates the extra sales a suggestion would create.

A claim such as “this recommendation converts at 65%” would therefore misread the evidence. We have no recommendation exposure, campaign assignment or comparison group here. The observed 65.0% is 320 out of 492 historical invoices, not a conversion rate for an intervention.

The first table keeps all 20 original pairs in their original order. “Together” counts eligible invoices containing both products. The share divides that count by all eligible invoices in the relevant year: 19,645 in the first year and 18,877 in the second. Open the supporting tables to inspect individual product counts and the independently selected later list.

Read the count, share and later rank together. A count describes the number of invoices containing a pair; the share also accounts for the different number of eligible invoices in each year. Rank describes its position against other qualifying pairs. None of these measures tells us whether a promotion worked.

The cutoff was unambiguous in both years. The final selected pair had no equally counted pair left outside the list. Changing the tie-break at the later cutoff therefore leaves the overlap at seven. This is a check of the ranking rule, not a statistical confidence interval.^[3]

The later shortlist answers a separate question: which pairs were most common together in the second year under the same eligibility and ranking rules? It is not a rearrangement of the original 20. Every eligible later pair could enter it.

Keeping the two lists separate prevents a subtle change of question. If we replaced the original list with the later winners, the report would describe a new shortlist without showing what happened to the one a team might have carried forward. Both views are useful: one audits an earlier choice, while the other identifies candidates for a new review.

The complete later table includes each product’s individual invoice count as well as the joint count. These counts do not establish that the products make a sensible bundle. Descriptions identify the recorded products; they are not evidence of compatibility, margin, stock availability or a shared customer need.

For a merchant planning a collection or cross-sell campaign, this case supports a practical question: what evidence would justify carrying an old pairing into the next brief? The answer needs a date, a defined set of invoices and a clear reason for the proposed placement.

Start by preserving the shortlist that informed the earlier decision. Keep its product identifiers and the period used to select it. When new records arrive, follow those same pairs before replacing them with current favorites. That makes a change in the evidence visible and gives colleagues a shared basis for discussing it.

Then connect the evidence to the actual placement. A collection page, a product-page suggestion and a creator demonstrating several products are different uses of a pairing. Ask what the customer will see, which product brings them into the experience, and what question the content should answer. Co-purchase records alone do not settle those editorial choices.

Creators can use the same distinction when reviewing a brand brief. A claim that two products are often bought together should come with its period and base. Ask whether the claim concerns all orders or only those containing one named item. That clarification helps a script describe the evidence accurately without promising a result the brand has not measured.

Before choosing the final products, the team still needs current information about stock, pricing, product fit and the proposed offer. Those are inputs for the business to supply, not explanations we can recover from this historical archive. If a campaign goes ahead, evaluate its effect using an appropriate comparison and the business’s existing measurement setup.

The worksheet below turns that discussion into a brief that can be reused. Complete the evidence review first, then transfer the selected idea into the linked social media campaign template. This keeps the reason for choosing a pair attached to the creative work without treating the old ranking as a recommendation engine.

Use this template for one proposed pairing. It is an ungated planning aid, not a scoring system. Enter your own dates and counts; the historical trinket-box figures are an explanation of the method, not suggested targets.

1. Decision and placement: [collection, product-page suggestion or creator brief] · [customer question the content should answer] · [person responsible for the choice]. Describe what will be shown and why the pairing is being considered.

2. Original evidence: [product A identifier and name] · [product B identifier and name] · [start and end dates] · [what qualifies as an invoice or order in this analysis]. Keep the original pair even if its later position changes.

3. Later evidence: [later start and end dates] · [all eligible invoices] · [invoices with A] · [invoices with B] · [invoices with both]. Use the same definitions across periods and retain one-product invoices in the all-invoice base.

4. Comparison: [original and later joint counts] · [joint count ÷ all eligible invoices for each period] · [later rank under the same rule]. If you add an anchor-product percentage in your own analysis, name the anchor and show its denominator. Record unavailable information as unavailable rather than filling it with zero.

5. Commercial check: [current stock and pricing] · [reason the products belong together in this placement] · [open question for the merchant or creator]. Keep this business judgment separate from the historical calculation.

6. Next decision: [retain, reconsider or gather more information] · [reason supported by the evidence] · [owner and next review date]. If you test a campaign, state separately how you will assess its effect. A change in a pair’s rank is not a campaign result.

A completed brief should let another person trace the proposed pairing back to its evidence and understand what still needs checking. It should also remain usable when the ranking changes: the record of the original decision is part of the review, not something to overwrite.

We used Daqing Chen’s Online Retail II dataset from the UCI Machine Learning Repository. Its source describes transactions from one UK-registered non-store giftware retailer, with many wholesale customers, from December 1, 2009 to December 9, 2011. Our comparison uses the two complete December–November periods and leaves the final partial December outside the analysis.^[1]

An analytical basket is one eligible numeric invoice with at least one qualifying product. Each distinct product counts once per invoice, regardless of quantity, repeated lines or worksheet overlap. A repeated customer can contribute several invoices. Missing customer identifiers do not by themselves remove an invoice, because this question concerns product pairs rather than unique people.^[2]

We use a conservative positive-product cohort. An invoice is excluded if an in-scope product line has a nonpositive or invalid quantity or price, or if specified identifier and timestamp checks fail. Cancellation-prefixed invoices cannot create a basket. We do not link cancellations back to earlier invoices or claim that the remaining records represent fulfilled orders. The package documents product-code scope and all exclusions.^[2]

In the first year, 23,269 numeric invoices minus 3,624 excluded invoices left 19,645 eligible invoices. In the later year, 21,192 minus 2,315 left 18,877. An invoice can fail more than one check, so the exclusion reasons overlap; their individual counts must not be added together. These are invoice counts, separate from the source’s line-record count.^[2]

For a pair to enter either year’s ranking, it must appear together on at least 20 eligible invoices, with each product appearing on at least 50. These were fixed display-support choices, not significance tests. We rank by joint invoice count, then product codes for ties. All original pairs remain in the later comparison whether or not they satisfy those later floors.

We fixed the periods, selection rule, outputs and worked-example choice before examining the pair results. Two separately written calculations reproduced the complete lists, counts, shares and tie checks. Results describe this historical archive; they are not estimates for today’s retailers or a representative sample of consumers.

Market basket analysis and checking patterns in later data are established practices. Our contribution is this reproducible audit of a specific dated shortlist, with every selected pair retained. It does not introduce a new algorithm or claim the first study of changing product associations.^[4][5]