Cart Abandonment Dataset For Funnel Analysis Data: Dataset Structure and Field Coverage

Most public retail data stops at receipts, and a receipt only shows who converted - it never shows the shopper who added a product to the cart and walked away. One record in Datadory's apparel-retail slice closes that gap: eCommerce Behavior Data from Multi-Category Store (Apparel Segment). It stores one row per clickstream event, with an event type field that distinguishes view, cart, remove_from_cart and purchase, so each funnel stage is directly countable rather than inferred.

The scale is what makes a funnel statistically usable. The corpus covers about 285 million users' events over seven months (October 2019 through April 2020) from a large multi-category online store operated by REES46, and Datadory scores it 9 of 10 - tied for the highest quality score among the industry's seven primary records. Every row carries product, brand, price, user and session identifiers, so sessions can be replayed end to end: first view, add-to-cart, then removal or purchase.

For contrast, consider what the rest of the slice offers. The Kaggle Fashion MNIST Mirror supplies 70,000 labeled 28x28 images but no behaviour at all; DeepFashion's 800,000+ photos annotate garments, not shoppers. Only this one record carries the event stream a funnel needs, which is why it anchors every section below.

The event schema turns the standard funnel into a group-by exercise rather than a modelling problem:

1. Group rows by session. The session identifier ties every event back to one browsing visit.
2. Flag stages per session using the event type: any view, any cart, any remove_from_cart, any purchase.
3. Count distinct sessions at each stage, then divide each stage by the stage above it to get stage-to-stage conversion rates.
4. Split deliberate removals from passive abandonment. A session with a cart event but no purchase is abandoned; whether it also contains a remove_from_cart tells you which kind it was.

Because coverage runs October 2019 to April 2020, monthly slices show how the funnel shifts across seven consecutive months, including the November-December peak season. Every figure in the output derives from logged behaviour rather than survey recall - the reason growth teams prefer clickstream over aggregate benchmarks when quantifying drop-off.

Two practical caveats keep the numbers honest. First, define the session window before counting: the corpus's session identifier does the grouping for you, but decide whether a session that ends in a remove_from_cart counts as recovered or abandoned before you publish a rate. Second, remember the store is multi-category rather than apparel-only, so treat product-level filters as part of the query, not an afterthought - the tags on Datadory's record include both funnel-analysis and cart-abandonment precisely because this filtering step is the core workflow.

Not directly, and knowing why saves wasted effort. Both UCI records in the slice are invoice-level: UCI Online Retail holds 541,909 invoice lines from 01/12/2010 to 09/12/2011 in a 22.6 MB Excel file, and UCI Online Retail II extends the same UK non-store retailer to 1,067,371 lines across December 2009 to December 2011 in 43.5 MB. Neither contains views, cart adds or removals; a cancelled order surfaces only as a negative-quantity line and browsing behaviour is invisible.

What they contribute to a funnel project is the post-purchase half: order values, repeat-purchase gaps and returns, both under commercial delivery terms 4.0. Online Retail II's two-year span captures two full annual cycles - enough to separate seasonal effects from structural ones in whatever converts. The practical split: use the REES46 events to quantify where shoppers drop off before paying, and the UCI invoices to model what the customers who do pay go on to do.

Government series cannot see individual carts, but they set the denominator that tells you whether a conversion swing is site-specific or market-wide. Three free sources cover it:

• US Census Monthly Retail Trade Survey (MRTS) - monthly sales by NAICS kind of business down to apparel subcodes 448, 4481, 44811, 44812, 44814, 44819 and 4482, from January 1992 onward in mrtssales92-present.xlsx (~0.44 MB, 34 annual sheets), with quarterly e-commerce series from Q4 1999. It is the only primary record whose format list includes api_json, and it is commercial delivery terms.
• UK ONS Retail Sales Index - Retail Industry - 622 time series in drsi.csv including dedicated Textile, clothing and footwear splits, monthly from January 1988 and current through July 2026 under OGL v3.0. Its internet sales series begins November 2006, isolating the online channel your funnel lives in.
• US Census Retail Historical Data Portal - about 1,010 advance reports back to October 1953 plus 107 Quarterly E-Commerce PDFs from Q4 1999; long-run context rather than month-to-month benchmarking.

All three refresh monthly, and none requires payment or registration.

Ranked against the questions a funnel analyst actually asks, using quality scores from Datadory's catalog:

1. Measuring view-to-cart-to-purchase drop-off - eCommerce Behavior Data (Apparel Segment): the only record in the slice with all four funnel stages logged as events, ~285 million users' sessions, quality score 9.
2. Separating deliberate removals from passive abandonment - the same corpus: remove_from_cart is its own event type, so the two behaviours are distinguishable without guesswork.
3. Post-purchase repeat and returns analysis - UCI Online Retail II: 1,067,371 invoice lines over two full annual cycles, quality score 8; drop to UCI Online Retail (541,909 rows, also score 8) when a lighter file suffices.
4. Benchmarking against the wider market - MRTS or the ONS Retail Sales Index, both quality score 9, both refreshed monthly, one per side of the Atlantic.

The rule that falls out: event-level questions get the clickstream, money-level questions get invoices, and trend questions get official statistics. No single file answers all three.

1. Pull one monthly CSV.GZ archive first rather than the whole ~4.6 GB version 8 bundle. Validate the four event types and the session identifier before committing disk space to all seven months (~16.5 GB uncompressed-equivalent).
2. Deduplicate sessions. Count distinct session IDs at each stage; raw row counts inflate stages where shoppers view many products before deciding.
3. Compute the three stage rates - view-to-cart, cart-to-purchase and overall view-to-purchase - then recompute them monthly across October 2019 to April 2020 to expose drift and seasonality.
4. Segment by price band. The price column rides along on every event row, so abandonment can be read against product value instead of reported as one blended rate.
5. Overlay the market backdrop. Pull MRTS NAICS 448 series or the ONS Textile, clothing and footwear series for the same months, so a falling conversion rate can be judged against sector-wide demand.

Steps one to four stay inside the clickstream; step five is what turns an internal metric into a defensible claim about the category.