Merchandising data checklist: ten fields AI planning tools need

AI planning tools need ten kinds of data: product identifiers, attributes, sales, stock, prices and promotions, returns, channel, lifecycle dates, location and supply constraints. If any of these is missing or inconsistent, the tool either cannot run or produces confident but unreliable results. The checklist below gives the fields, why each matters and the most common defect.

What are the ten fields?

Why do attributes and identifiers come first?

Uphance's review of AI demand forecasting for apparel says the tools need consistent style attributes such as fabric, silhouette and fit in one system, not split across product lifecycle management, e-commerce platforms and spreadsheets. An arXiv paper on forecasting demand for new fashion items from the Indian platform Myntra used item attributes such as colour and material and derived features such as style age. Identifier hygiene is equally basic: SPS Commerce explains that a style number identifies a design and is not globally unique, and that each colour and size combination needs its own GTIN.

Read also

Machine learning basics for merchandisers: models, features and training data

What do sales, stock and returns data need?

• Sales at the lowest level you can: style, colour, size, day or week, channel.
• Stock by day, so that out-of-stock days can be identified and adjusted, as in Inventory Planner's size curve method.
• Returns posted promptly: Uphance says returns should reach inventory within days rather than weekly.
• Wholesale commitments separated from direct channel stock, which Uphance names as a requirement for accurate planning.

What do prices, promotions and lifecycle dates add?

The Myntra paper used merchandising factors such as discount relative to brand and platform averages, list-page visibility and promotion timing, and found that tree-based machine-learning models clearly beat a naive baseline for new items, though accuracy was moderate and some future inputs were unknown for new items. Without price history and lifecycle dates, a tool cannot separate a style's decline from the effect of a markdown or a launch week.

What should you do before buying or building a tool?

1. Audit the ten fields for your top categories and note which are complete, partial or missing.
2. Fix identifiers and attributes first, because they are cheapest to fix and affect everything else.
3. Check you have at least two seasons of clean history for replenishment styles.
4. Plan for new products: decide how comparable items will be chosen.
5. Separate wholesale commitments from direct stock in the data.
6. Ask any vendor to run a pilot on your data and compare with your current method.

It is also worth asking how a tool handles missing data. Some tools silently fill gaps with averages, which hides problems; others flag them. Ask for a report of data quality issues from any pilot, and treat a long list as useful information, not a failure.

Lastly, remember that planners remain accountable. Uphance observes that a merchant's judgement anchored to comparable styles remains competitive for new fashion drops. A good tool supports that judgement with data and makes assumptions visible, which is the standard to apply when evaluating one.

Read also

How Abercrombie and Fitch plans assortments and inventory: what is public on AI

How can you score your own readiness?

A simple score helps decide where to start. For each of the ten fields, rate it 0 for missing, 1 for partial (some categories, some channels or inconsistent) and 2 for complete and consistent. A total below about ten suggests that data work should come before tool purchase; above sixteen suggests you are in a position to pilot. These thresholds are a rule of thumb rather than a standard, so adjust them to your business.

• Score each field per category, because data quality often differs between womenswear, menswear and accessories.
• Weight the fields that matter most for your business: for a wholesale-led brand, supply constraints and channel separation carry more weight.
• Re-score every season, and track the trend.

The Myntra paper is a reminder of the limits even with good data: accuracy was moderate, with item-level correlations around 0.4, and some future inputs such as promotions and visibility are unknown for new items. Better data raises the ceiling, but it does not remove uncertainty. A planning tool should therefore present ranges and let planners override at style, colour and size level, as Uphance notes is needed in practice.

A final practical point: assign an owner to each field. Data that nobody owns degrades. The owner does not need to enter it, but must be accountable for its completeness and for fixing defects at source.

Start the audit with a single category that matters commercially, such as a core replenishment line. Pull two seasons of data, check each of the ten fields and write down the gaps with an owner and a date. Doing this once for a narrow scope is faster than a company-wide review and shows what the full effort would cost. It also gives a realistic basis for conversations with vendors about what a pilot requires from your data.