Bayesian Decision Engine for store analytics
Where sales value leaks between the price tag and the till, and what can be done about it.
Why this matters
Every product has a price tag, but a store rarely keeps all of it. Discounts, schemes and product mix pull the net sale below the tag. The share of the price tag a store keeps is its net realization (net sales value / gross sales).
Two stores selling the same category can differ by many points, and a few points across many stores is a large amount of money. Store analytics finds those differences, separates persistent gaps from noise, and tells the store and category teams where to look first.
below peers on every Rs 100 of price tag. With 6+ months of history and high confidence, this would make the review list.
Illustration only, not client data. Net realization = net sales value / gross sales.
What we have built
This has been run end to end on a large multi-brand apparel retail chain with over three years of transaction history. Client identity is withheld.
- Data pipeline. Transaction-level sales are pulled from the ERP (SAP or another system) into an analytics database by scripted extracts that check their own output and keep a log.
- Data-quality screens. Staff and shared accounts, non-product codes, thin-history cases and mixed category coding are handled before any analysis runs.
- Five analyses, each run two ways so results can be cross-checked, with plain-language guidance on what may and may not be concluded.
- A web console to run each analysis and read its results.
What the analyses answer
| Analysis | Question it answers | Output | Main data needed |
|---|---|---|---|
| Store realization review Main focus | Which stores keep clearly less of the price tag than peers selling the same category? | Ranked store-category list with the size of each gap and how sure we are | Gross and net value by store, category and month |
| Price response | If a category's selling price falls, how much do units rise? | Response by category, with a range; "cannot say" where data is thin | Units and net price by product and month |
| Discount-day effect | Do days with heavy discounting sell more than normal days? | Lift by category, and which cases cannot be judged | Daily units, discount depth, stores trading |
| Repeat-purchase segments Where applicable | Which customers are likely to come back? Used only for categories bought often. | Customer groups with observed repeat rates | Customer ID and purchase dates |
| Brand response For infrequently repurchased categories | When a brand's net price or discount changes, how much do its units and share of the category move? | Response by brand and category, with a range; "cannot say" where data is thin | Units and net price by brand, category and month |
Repeat-purchase segments apply only to categories customers buy again and again. For durable, infrequently repurchased categories (for example fixtures or appliances) it is skipped and brand response is used instead. Brand response applies the price-response method at brand level and has not yet been run on the reference chain. The same method can compare stores within a channel (for example exclusive stores, dealers, online) or region, if the ERP carries that field.
What the store review found on the reference chain
Aggregated results.
- More than 19,000 store-category combinations were reviewed.
- A plain "below the category average" ranking would have flagged about 45% of them. Many had too little history or too small a gap to act on.
- After screens for history (at least 6 months), size of gap (at least 5 points) and statistical confidence, about 6% remained: a short ranked list a store team can review.
Illustration only, not client data: Store X, category Y keeps Rs 62 of every Rs 100 while peers keep Rs 80, with 14 months of history and high confidence.
How it would work for a retail outlet firm
Four phases, from first data extract to a monthly scorecard.
Data and definitions
4–6 weeksAgree the ERP extract. Confirm what net value, discount and category mean. Run the quality screens.
Output: Validated data set and a short data-quality note
Store review
4–6 weeksRun the realization review and walk through the ranked list with store and category teams.
Output: Ranked store-category list with gaps and confidence
Pilot
8–12 weeksFor the largest gaps, split each into discount depth versus product mix, take one action, and measure.
Output: Rupees recovered, measured per action
Monthly cycle
OngoingRefresh monthly. A scorecard tracks each flagged store-category and any recovery.
Output: Monthly scorecard
Timings are indicative and would be agreed after a first look at the data. They are counted from the date data access is granted and include buffer for extract validation, changes in field definitions, and availability of store and category teams for reviews. Price, discount and customer analyses can be added once the fields exist.
What to expect
- It shows where to look and how sure we are. The discount-versus-mix split in the pilot says why.
- Value is shown only by measured pilots.
- Where data is too thin, the analysis says "cannot say" instead of guessing.
What not to expect
- Findings are patterns in past data, not proof that a change will work.
- Gaps are indicative. Part of a gap can be product mix and cannot be recovered, so we do not quote a gap as recoverable revenue.
- Pilot results depend on enough trading months after each action to measure the change; a pilot with too few months is reported as inconclusive.
- Margin and cost to serve are not included unless provided.
What we need from IT to start
- Read access to line-level sales for at least 2–3 years: date, store, product, category, quantity, gross value, discount and net value; customer ID and channel if available.
- A data owner who can confirm field definitions (net value, discounts, returns, category hierarchy).
- A place to load the extract, and agreement on data handling.
- Contacts in store operations and category management to review the first list.
Common questions
What is net realization in retail?
Net realization is the share of the price tag a store keeps: net sales value divided by gross sales. Discounts, schemes and product mix pull the net sale below the tag.
Does store analytics say why a store is below its peers?
It shows where to look and how sure we are. It does not by itself say why; the discount-versus-mix split in the pilot does.
What data is needed to start?
Read access to line-level sales for at least 2–3 years: date, store, product, category, quantity, gross value, discount and net value; customer ID and channel if available.
How long does it take?
Data and definitions take 4–6 weeks, the store review 4–6 weeks, and a pilot 8–12 weeks, counted from the date data access is granted. A monthly cycle then tracks recovery.
Find where your stores leak value
Start with a data extract and a short data-quality note. We will tell you what the data can support.