The New Science of Retailing

1. 1Name the business problem before the technology. The passages tie RFID to better data capture and a better match of supply and demand. Write one sentence naming the data gap (for example, stockouts you cannot see) and the metric it moves (in-stock rate, sales, labor).
2. 2Size the benefit with the Wal-Mart method. Start from the measured stockout reduction, then haircut it. Most retailers the authors consulted assume sales rise by only 25% to 50% of the stockout reduction. Adjust upward if stockouts hit popular items or complementary items (a sold-out sheet set takes the matching pillow covers with it). Adjust downward if your measurement window was late in the day. Wal-Mart's was 2 to 10 p.m.
3. 3Convert the sales lift to gross margin dollars. Apply your own margin to the haircut lift. Wal-Mart's roughly 1% lift is worth about $791 million in gross margin.
4. 4Add the benefits the sales lift misses. Follow METRO's approach and count cost savings, such as labor from manual steps that RFID eliminates, alongside the stockout reduction.
5. 5Build the full cost side. Include tags, readers, software, and training. Flag tag price as the key cost driver, since the passages identify it as such. Compare benefit to cost at each level where tags could go: pallet, case, or item.
6. 6Check the external dependencies. Confirm the frequency band for each market in your supply chain: 902–928 MHz in the U.S. (most passive tags near 915 MHz), 865.7–867.6 MHz in Europe, and unallocated in many emerging markets. Note that the EPCIS standard is early and slow to be adopted. Any gap here is a reason to limit scope, not to cancel.
7. 7Stage the commitment. Start at the level and in the geography where the numbers and the standards already work. Define in advance what evidence (benefit measured against cost, standards progress) triggers the next stage. Keep each stage small enough that stopping it is cheap.
8. 8Check fit with your company's life stage. The passages say rocket-science methods pay off when a firm, at roughly fifty stores, shifts from proving the concept to adding order and profitability. Before then, basic execution matters more than advanced technology.

Take the Wal-Mart analysis in the passages. RFID is expected to cut stockouts by 2.7 percentage points. A straight conversion would overstate the gain, so apply the 25% to 50% rule used by the retailers the authors consult: sales rise by 0.675% to 1.35%, roughly 1%. At Wal-Mart's scale, a 1% sales increase adds about $791 million in gross margin. Next, test the number's weak points. The stockouts were scanned only between 2 and 10 p.m., so morning stockouts may have been lower, which argues for caution. The lift could also be higher if popular or complementary items were hit, which argues the other way. Then set it against costs for tags, readers, software, and training, which the authors call substantial at Wal-Mart's size. METRO's analysis adds labor savings from eliminated manual steps. The decision rests on the net of all of these, not on the headline stockout figure.

• If the benefit case rests on a stockout reduction, then haircut it to 25–50% of that reduction before converting to sales and margin.
• If the benefit only covers sales and ignores labor, then add the labor savings from eliminated manual steps, as METRO did, before judging the case.
• If your supply chain runs through markets with no allocated frequency or a different band (U.S. 902–928 MHz, Europe 865.7–867.6 MHz), then confine the deployment to markets where the frequency works.
• If the case only clears at a lower tag price, then treat tag price as the swing variable and hold the next stage until the price supports it.
• If you are a young retailer still proving the concept and opening stores, then put basic execution first. Advanced technology pays off as you approach about fifty stores and need profit discipline.

• Judging the technology without a framework: executives who came up through operations or merchandising accept vendor claims or reject the technology outright. The tell: the pitch cites promise ("could change retailing as radically as the cash register") but no sized benefit.
• Counting the full stockout reduction as sales: the tell is a sales lift equal to the stockout drop, with no 25–50% haircut and no mention of when the stockouts were measured.
• Pricing only the tags: the tell is a cost line that omits readers, software, and training.
• Ignoring standards and frequency: the tell is a global rollout plan with no frequency check for each market and no view on EPCIS adoption.
• Buying software and analysts and expecting results: the tell is a project with no defined goal for your specific context and no plan for how to reach it.

• ▸Problem and metric: one sentence naming the data gap and the metric it moves (e.g., in-stock rate).
• ▸Stage and level: pallet, case, or item, plus the stores or lanes in scope.
• ▸Measured stockout reduction: the percentage points from your test, and the hours and days it was measured.
• ▸Haircut applied (25%–50% or other): the share of the stockout reduction you convert to sales, with a reason.
• ▸Resulting sales lift and gross margin dollars: lift times your margin.
• ▸Labor and other savings: manual steps eliminated, valued as METRO did.
• ▸Cost build: tags (price per tag), readers, software, training.
• ▸Frequency and standards check: band by market, and EPCIS status.
• ▸Next-stage trigger and stop rule: the evidence required to expand, and the point at which you stop.

1. 1Separate stable demand parameters from volatile product history. Seasonality, size mix, which product types sell best at which stores, price elasticity, and category growth trends endure for years even as products turn over. Estimate these from past sales history and treat the sales of any one prior style as nearly worthless.
2. 2Aggregate before you read trends. An average SKU sells about one unit per month per store (study of thirty-two retailers), so store-SKU data is too sparse to forecast from. Group similar stores and similar SKUs until the sales rate is high enough to show a pattern.
3. 3Record the conditions behind every sale. Log price and product availability always, and weather and competitive activity where they matter. Estimate how each factor moved past sales, then state your assumptions about how each will behave in the coming season.
4. 4Keep models as simple as the data allows. Count the parameters against the data you have. The substitution example cut thirty parameters to three, giving ten times more data per parameter, like flipping a coin 100 times instead of 10. Prefer the simple model when you believe substitution rates are similar across options.
5. 5Check estimated results against an independent signal. In the source example, a store's House 3-Low demand share correlated with zip-code income. If an estimate has no external confirmation such as demographics, treat it as unproven.
6. 6Build a weekly sales-curve table from past seasons for the category, such as the share of season sales that occur in each week. Once an item has sold, extrapolate: units sold to date divided by the cumulative fraction of the season they represent gives the season forecast.
7. 7Buy a little, sell a little, update, buy more. Use the replenishment lead time (typically twelve weeks) to reorder after week two on the updated forecast. Size the initial buy so you probably won't sell out in the first fourteen weeks but won't be left with a loss-making excess.
8. 8Pair every forecast with an error distribution and buy to the marginal-unit rule. Compute p = (cost − markdown price)/(normal price − markdown price). Add units while the probability that the next unit sells exceeds p.

A navy turtleneck costs $40, sells at $60, and marks down to $18. The target sales probability for the marginal unit is p = (40 − 18)/(60 − 18) = 0.524. In the demand table, the cumulative probability that demand is ≤ 93 is 0.442, so the 94th unit has a 0.558 chance of selling. That beats 0.524, so its expected profit is positive. The cumulative probability for ≤ 94 is 0.497, so the 95th unit has a 0.503 chance, below 0.524, and its expected profit is negative. The optimal buy is 94. For the early-sales update, an item that sold 11 units by the end of week two, when those weeks represent 11 percent of season sales, is forecast at 100 units for the 26-week season. Across 27 catalog items, preseason buyer forecasts carried 50–100 percent error, while forecasts extrapolated from two weeks of sales carried 10–20 percent.

• If a store-SKU sells about one unit a month, then aggregate across similar stores and SKUs before reading any trend.
• If you have an early-sales curve for the category and at least a couple of weeks of sales, then replace the preseason buyer forecast with the extrapolated forecast and reorder if lead time allows.
• If the next unit's probability of selling exceeds (cost − markdown price)/(normal price − markdown price), then buy it. If it falls below, stop.
• If a model has many parameters relative to the data and you believe the options behave similarly, then collapse it to fewer parameters.
• If inventory was inflated and prices cut in prior periods, then discount optimistic sales forecasts. Analysts' bias for such retailers averaged roughly 5 percent of sales.

• Buying more of whatever is selling well on raw store-SKU data. The tell: reorders chase one- or two-unit blips that reverse.
• Treating past sales as a guide without the conditions behind them. The tell: a promotion or stockout period is read as true demand, so the forecast is skewed by price or availability effects.
• Over-parameterized models fit to thin data. The tell: estimates swing wildly store to store and fail the external check, such as income-correlated shares.
• Relying on a single point forecast from buyers, which carries 50–100 percent error. The tell: no distribution of demand exists and the buy quantity is not tied to a marginal-profit threshold.
• Ignoring inventory and price effects on sales. The tell: forecasts run high after periods of heavy stock and discounting, because the effect of a 1 percent inventory change (about 0.224 percent of sales on average) is left out.

• ▸Item / category and season length (weeks): identify the item and the number of weeks in the selling season.
• ▸Cumulative share of season sales by week (from past seasons): the fraction of total sales typically complete by the current week.
• ▸Units sold to date and weeks elapsed: actual sales through the current week.
• ▸Updated season forecast: units sold ÷ cumulative share for the current week.
• ▸Preseason forecast and gap: the original buyer forecast, and the difference from the updated one.
• ▸Conditions log: price changes, stockouts, weather, competitor activity during the weeks sold.
• ▸Cost, normal price, markdown price: the inputs for the marginal-unit threshold.
• ▸Target sales probability p: (cost − markdown price) ÷ (normal price − markdown price).
• ▸Replenishment lead time and reorder decision: weeks of lead time, weeks left to sell on arrival, and units to reorder (add units while probability of selling exceeds p).

1. 1Measure the gap before changing anything. Record two numbers per category: the share of sales sold at markdown (the passage cites markdowns rising from 8 percent in 1970 to nearly 30 percent in the 1990s) and the in-stock rate. Both must move; a retailer can have piles of marked-down goods and still lack the item the customer came for.
2. 2Convert in-stock improvement into profit terms. Multiply your gross margin by the expected sales lift from a higher in-stock rate. Compare the result with current net profit margin. This sets the business case and the priority against other projects.
3. 3Break sales down by size, style, color, and store, then look for variation across stores. The clothier analysis found a surprising variety in size mix by store. The chain had shipped the same size mix to every store, so some stores ended the season with leftover small sizes and others with leftover large sizes.
4. 4Turn the data into a one-page finding, not a report. The clothier's size-selling report was an eight-inch printout buyers never used. Reduce each finding to a decision: which stores get which size mix, and by how much.
5. 5Replace the single 'average' test store with clustered test stores. Group stores by taste, meaning the product mix customers bought in a prior season, not just by geography or demographics. Pick one test store per cluster (ten clusters for a ten-store test) so the test sample reflects the chain's diversity.
6. 6Forecast from the test, then correct for seasonality. Scaling test sales by weeks and store count (80 × 8 × 10 = 6,400) assumes a flat sales rate. A 24-week season can run a higher or lower weekly rate than a 3-week test, so adjust for that.
7. 7Back-test the forecast. Plot forecast (horizontal) against actual demand (vertical) for each style/color from a past season. Points scattered far from the diagonal mean the method fails. Tight clustering means it works.
8. 8Close the buyer-customer gap. Put the real customer in front of buyers: store visits, signs in the buying office ('you're buying it to sell, not to use'), or living a week on a typical customer's budget. Buyers who differ from customers misjudge assortments.

A clothier sent the same size mix of every product to all its stores. Analysis showed the mix of sizes actually sold varied widely by store, so some stores ended the season with many small sizes left over and others with mostly large sizes. The VP of planning said buyers already had the patterns and pointed to an eight-inch weekly size-selling report broken down by size, style, color, and store. He then acknowledged that buyers had never used it. The fix is to turn that report into store-specific size allocations. The profit logic comes from a jeweler example in the same passages: gross margin of 42 percent and net profit of 1 percent. Raising in-stock from 90 to 93 percent lifts sales 3 percent and adds 1.26 percent of revenue in gross margin (0.42 × 3%). If nearly all of it reaches the bottom line, profit rises by more than 100 percent. The same logic applies to the clothier: stockouts in one store and leftovers in another are two sides of the same mismatch.

• If the stores in your chain sell visibly different mixes (sizes, styles, tastes), then stop shipping identical assortments and allocate by store or cluster.
• If a test-store forecast plots far from actual demand, then do not scale a single 'average' sample. Cluster stores by prior-season product mix and test one store per cluster.
• If a test window is much shorter than the selling season, then adjust the scaled forecast for seasonality before committing quantities.
• If your gross margin is high but net margin is thin (jewelry: 42 percent vs. 1 percent), then rank in-stock gains above cost-cutting. Incremental gross margin largely drops to the bottom line because fixed costs are already covered.
• If a report is too large for buyers to act on, then reduce it to exceptions and recommended allocations rather than adding more data.

• Treating data volume as insight: buyers receive a huge size-selling report and never use it. The tell: reports exist, but every store still gets the same size mix.
• Using one 'average' test market and multiplying up (the 'if one Peoria is good, twenty-five are perfect' approach). The tell: forecast vs. actual points scatter widely.
• Ignoring seasonality when scaling a short test to a full season. The tell: forecasts systematically miss in seasons whose weekly rate differs from the test weeks.
• Optimizing only against markdowns or only against stockouts. The tell: marked-down piles sit next to empty slots for the items customers ask for.
• Buyers who do not resemble the customer (wealthier, different tastes) buy for themselves. The tell: assortments skew away from what the store's actual shoppers purchase.

• ▸Category / season: name the product group and selling period under review.
• ▸Markdown share of sales: percent of sales sold at markdown, with the prior-year comparison.
• ▸In-stock rate: current percent, plus the target and the expected sales lift.
• ▸Gross margin and net margin: current percents, used to compute incremental profit (gross margin × sales lift ÷ net margin).
• ▸Store variation check: list the dimension (size, style, color) and the stores where the sold mix differs most from the shipped mix.
• ▸Store clusters: group stores by prior-season product mix; name one test store per cluster.
• ▸Forecast method and seasonality adjustment: describe how test sales scale to full season and how the weekly rate is corrected.
• ▸Back-test result: forecast vs. actual for last season by style/color; note which points miss most.
• ▸Buyer-customer action: one concrete step this season to bring buyers closer to the real customer.

1. 1Start with what your customer buys on, not with a store-execution checklist. The source's test: Nordstrom-quality service does not help if customers want Target-level prices. Write down the one or two reasons a shopper picks you over a competitor, and rank every execution fix against them.
2. 2Look at the store the way the customer does. Walk the floor for the item a customer came to buy. The source's signature failure is a store full of marked-down merchandise that lacks the one wanted item. Log each case where the item is missing and the shelf is crowded with excess.
3. 3Treat the inventory record as the decision input. The source holds that accurate information is the lifeblood of data-driven retailing and that you cannot make good decisions without good data. Before trusting replenishment, forecasting, or availability reports, compare what the system says is in each store with what is physically there.
4. 4Separate 'no stock' from 'stock that is not where it should be.' Excess inventory and missing items can coexist. For each missed sale, record whether the product was absent from the store, in the store but mislocated, or recorded as present but not physically there. Each cause has a different fix.
5. 5Track inventory against same-store sales, not on its own. Cutting inventory improves cash and turns, but you must watch whether sales fall with it. Review the two series together each period.
6. 6Benchmark your turns against competitors before deciding inventory is too high or too low. The source's Home Depot case treats 'normal' as turns similar to competitors, and the investor view of your numbers is a useful check on your own.
7. 7Find where execution breaks by comparing stores. The source's research program on store execution studied what separates stores that convert shoppers from those that lose them. Rank your stores on availability and record problems, then inspect the best and worst side by side.
8. 8Align incentives with the execution behaviors you want. The source warns that new methods fail when no incentive exists for people to adopt them. Tie store-level goals to accurate counts, shelf placement, and customer experience, not only to cost or cash measures.

Home Depot under Bob Nardelli, as told through investor David Berman's DeeBee Report (June 10, 2003). Nardelli focused on cash flow and dramatically lowered inventories. Cash balances rose, but same-store sales fell sharply and the stock dropped from around $40 to $22. Under pressure he reversed course and built inventory. Since Q2 of the prior year, inventories had climbed until they were up 25 percent year over year. Same-store sales improved, and the stock rose from $22 at the start of 2003 to $36 by year-end. Berman noted that Home Depot had only returned inventories to 'normal,' meaning turns similar to competitors, and cautioned that the inventory-driven sales gain was arguably a one-time benefit. For store execution, the lesson is that stripping stock out of stores cost sales because the product customers wanted was no longer there. Measure availability and inventory against sales together, with peer turns as the reference point.

• If customers choose you on price, then put execution effort into availability and cost discipline before service extras. If they choose you on service, then protect service levels first.
• If shelves hold marked-down excess but the wanted item is missing, then treat it as a placement and replenishment problem, not a total-inventory problem. Do not just add stock.
• If inventory is cut and same-store sales fall, then restore inventory toward peer-level turns. Do not keep optimizing cash.
• If the inventory data is not trusted, then fix record accuracy before running forecasting or replenishment analytics. Poor data produces poor decisions.
• If a new execution method is introduced, then pair it with a store-level incentive. Otherwise people revert to their routines.

• Optimizing cash at the expense of sales. The tell: inventory and turns improve while same-store sales and the stock price fall, as at Home Depot.
• Adding stock to solve availability. The tell: markdowns grow but the item shoppers ask for is still missing.
• Trusting the system count without checking the shelf. The tell: the system shows stock, the shelf is empty, and replenishment never triggers.
• Investing in the wrong customer promise. The tell: premium service is offered to price-driven shoppers, or low prices to service-driven ones, and conversion does not improve.
• Rolling out new tools without incentives. The tell: analytics software is installed but store behavior and results stay unchanged.

• ▸Store / date: identify the store and the audit date.
• ▸Customer purchase driver: state whether shoppers choose you on price, service, or something else, with the evidence.
• ▸Item sought: name the specific product used for the walk-through test.
• ▸System on-hand vs physical count: record both numbers for the item and the gap.
• ▸Miss cause: mark one of absent from store, in store but mislocated, or recorded but not physically there.
• ▸Excess alongside gap: note any marked-down or surplus stock on the same shelf while the item is missing.
• ▸Same-store sales vs inventory trend: record year-over-year change in each for this store.
• ▸Turns vs peer benchmark: record your turns, the competitor reference, and the difference.
• ▸Owner and incentive: name who fixes the issue, the incentive tied to it, and the review date.

1. 1Separate service level from inventory level. Service level is the level of product availability you aim for. It is not reported in public financials and is rarely tracked even internally, so you have to start tracking it yourself.
2. 2Pick sales-capture as the headline measure and keep the raw in-stock rate as a diagnostic. Your working definition: of the sales customers would have made, how many did you actually make? The source gives no formula, so write yours down and hold it fixed across periods.
3. 3Record the conditions behind every sales observation, not just the units. Price and availability always shape sales. Weather and competitor activity often do. Log what was on shelf and at what price when each sale happened or failed to happen.
4. 4Aggregate before you read availability. An average SKU sells about one unit per month per store, so a single store-SKU history is too thin to show a pattern. Group similar stores and similar SKUs, then measure capture at the group level.
5. 5Split inventory into the part that serves customers and the part that does not. Obsolete stock sits on the balance sheet but adds nothing to service level. Count only sellable, relevant stock when you ask whether more inventory is buying more availability.
6. 6Check the mix, not just the quantity, by store. Size mix varies a lot across stores. Shipping every store the same mix leaves some stores with leftover small sizes and others with leftover large ones. Measure capture by size, style and color within each store group.
7. 7Benchmark inventory against availability, not turns alone. Turns differ by segment (Kroger about 14 a year, Gap about 7), so cross-segment comparison is meaningless. Ask whether each added dollar of inventory raised availability or only covered inefficiency. That is your position relative to the inventory-service frontier.
8. 8Use software to read the data. A retailer with 10,000 SKUs at 500 stores generates about 5 million numbers in a weekly size-selling report. Push availability and capture into algorithmic exception reports instead of printouts.

The source describes a clothier that sent the same mix of sizes of every product to all its stores. The authors found that the size mix customers bought varied widely by store. At season's end some stores held mostly small sizes and others mostly large. This is the failure the construct targets: the chain carried inventory, yet customers at each store could not find their size. The VP of planning said the buyers already had the patterns, pointing to an eight-inch printout of sales by size, style, color and store. That report was useless in practice. A single week's sales could not reveal patterns, and buyers had neither the time nor the inclination to read it. The fix is to aggregate stores with similar size mixes, estimate each group's stable size mix from history, and measure sales capture by size per group. A store holding a lot of inventory but capturing a low share of its size-level demand shows the problem. The total inventory number does not.

• If inventory is rising and capture is flat, treat the added stock as covering inefficiency or obsolescence until you can show the extra units serve customers; explain it to investors before the build, not after.
• If a store-SKU sells about one unit a month, then do not judge its availability from its own sales history. Aggregate similar stores and SKUs first.
• If you compare inventory efficiency across retailers, then compare within a segment. Do not set Kroger's roughly 14 turns against Gap's roughly 7.
• If sales rose or fell, then check price and availability conditions before crediting demand. Both always affect sales, and raw sales hide them.
• If markdown piles coexist with missing items customers want, then the problem is assortment and allocation mix, not total inventory. Fix the mix before buying more.

• Reporting a raw in-stock rate as if it were availability. The tell: high in-stock numbers while customers still say they cannot find the item, or sales stay flat when items were present only in unwanted sizes or variants.
• Using inventory turns as a stand-in for the right service level. The tell: you defend inventory against a turn benchmark from a different segment and never state a target availability.
• Counting obsolete stock as protection. The tell: inventory is large and growing, markdowns are heavy, and sales capture does not improve.
• Reading availability from thin store-SKU data. The tell: you chase one-unit-a-month swings in single stores and infer trends that are noise.
• Drowning in reports without synthesis. The tell: weekly size and style printouts exist, yet buyers say they have neither the time nor the inclination to use them and the same mix still ships everywhere.

• ▸Store group and SKU group: name the aggregated cluster, since individual store-SKUs sell too slowly to read.
• ▸Period and conditions: record price points and any weather or competitor events in the period.
• ▸Observed sales: units actually sold in the period.
• ▸Estimated demand: units customers would have bought had the product been available. State your estimation method, since the source gives none.
• ▸Sales-capture rate: observed sales divided by estimated demand; keep the same formula every period.
• ▸Raw in-stock rate: share of the time the item was on shelf, kept as a diagnostic only.
• ▸Mix check (size/style/color): list where wanted variants were missing and where unwanted variants are left over.
• ▸Obsolete inventory share: portion of stock that cannot serve current demand.
• ▸Inventory-service read: did added inventory raise capture, or only cover inefficiency? Cite the evidence.

1. 1Define satisfaction with one overarching measure plus specific drivers. Use the study's design: a 10-point question, 'How satisfied were you with your store visit?', plus specific questions on greeting, assistance, associate knowledge, finding everything, out-of-reach items, prices, aisle debris, and checkout speed.
2. 2Collect the survey cheaply at the point of sale. Print the invitation on the receipt, offer a lottery entry with a cash prize, and take answers by an 800-number keypad survey. Record yes/no questions as the monthly percentage answering yes per store. Record 10-point questions as the monthly average score per store.
3. 3Add objective store data to the survey data. Gather monthly sales by store, planned versus actual payroll, and mystery-shopper results on in-stock levels and associate friendliness. Mystery shoppers check whether a policy, such as greeting customers by last name, is actually followed.
4. 4Remove noise before testing drivers. Deseasonalize sales and compare each store with its own average. The case used seventeen months of data across about five hundred locations.
5. 5Rank the drivers by their effect on sales. Sort stores into quintiles on each driver and compare deseasonalized sales to the store average. In the case study, payroll above plan was the biggest driver. Associate knowledge rating and the percentage who 'found everything' came next.
6. 6Tie each driver to an operational lever. More staff at registers shortens checkout. More people checking shelves and moving stock from the back room raises the share who find what they wanted. More floor staff raises the chance a customer gets help when needed.
7. 7Test costly policies against the margin they earn. Compare the incremental gross margin on the extra sales with the added payroll cost before you commit. The source states that the margin dwarfed the extra cost in the study.
8. 8Fix supply as well as the floor. 'Found everything' depends on product being in stock. Improving in-stock rate through better forecasts and inventory management raises sales and satisfaction. Align incentives where the supply chain caps availability.

A home furnishings and small appliances retailer with about five hundred stores supplied seventeen months of data. Each month the data held gross sales, payroll versus plan, and survey results. Customers answered an overall 10-point satisfaction question and specific questions on greeting, help, associate knowledge, finding everything, and checkout. Stores were grouped into quintiles by actual payroll relative to planned payroll. Deseasonalized sales soared above average in the months when payroll exceeded plan. Payroll was the single biggest sales driver. Next came the customers' rating of associate knowledge and the share who said they found everything they wanted. The source says that getting store execution right can produce double-digit rises in comp sales. It contrasts this with remodeling, which is only a one-time stimulus. The lesson: satisfaction levers are staffing, knowledge, and availability, and you can measure each against sales.

• If a driver's quintile analysis shows sales rising with it, then fund it, but first compare the incremental gross margin on the extra sales with the added cost.
• If stores run below planned payroll in a month, then expect lower sales and lower satisfaction, and treat the shortfall as a cost, not a saving.
• If the 'found everything' rate is weak, then look first at in-stock levels and shelf and back-room replenishment before adding promotions.
• If you adopt a service policy such as the 10-foot greeting or last-name greeting, then verify compliance with mystery shoppers and test whether it moves satisfaction and sales.
• If you want a sales lift, then prefer ongoing execution gains over remodels, which give only a one-time stimulus.

• Assuming a hospitality ritual raises satisfaction without testing it. The tell: the policy exists, but no analysis links greetings to satisfaction scores or sales.
• Cutting payroll to hit budget. The tell: actual payroll runs below plan, checkout lines lengthen, and fewer customers report getting help.
• Reading raw sales without adjusting for conditions. The tell: store comparisons ignore seasonality, so a quintile ranking reflects timing rather than execution.
• Treating satisfaction as a floor-only issue. The tell: associates are friendly and knowledgeable, yet customers do not find what they want because of stock-outs or overstocks of the wrong items.
• Collecting survey and POS data but never turning it into drivers. The tell: plenty of numbers, no ranking of which factors move sales ('awash in data and starved for information').

• ▸Store and month: identify the unit and period, with sales deseasonalized.
• ▸Overall satisfaction (10-point average): record the store's average answer to 'How satisfied were you with your store visit?'
• ▸Percent found everything: record the share answering yes to 'Did you find everything you were looking for?'
• ▸Associate knowledge rating (10-point average): record the average score for associate knowledge.
• ▸Percent greeted / percent assisted: record the yes rates for greeting and for help when needed.
• ▸Checkout quick and easy (%): record the yes rate for checkout speed.
• ▸Actual payroll versus planned payroll: record the ratio and the quintile the store falls into.
• ▸Deseasonalized sales versus store average: record the gap in percent.
• ▸Incremental margin versus added cost: estimate the extra gross margin from the sales lift and compare it with the added payroll cost to decide whether to act.

1. 1Measure the availability gap first. Record the in-stock rate by category and store, and count lost sales from customers who entered with a clear intent and left empty-handed. The KSA survey found that half of the 70 percent who arrived with a clear purchase idea left without buying because they couldn't find the item.
2. 2Translate any in-stock gain into profit before you fund it. Multiply the expected sales lift by gross margin (0.42 × 3% = 1.26% of revenue for the jeweler), then compare that with current net margin (1%). Fixed store costs do not rise with the extra sales, so most of the incremental margin reaches the bottom line.
3. 3Forecast from aggregated data, not store-SKU data. An average SKU sells about one unit per month per store, so group similar stores and similar SKUs until patterns appear. Record the conditions behind each sale: price, availability, weather, and competitor activity.
4. 4Mine history for the demand parameters that outlast individual products: seasonality, size mix, the product types that sell best at given stores, price elasticity, and category growth trends. Use them to set purchase quantities even when the specific style is gone.
5. 5Localize mix where stores differ. If size sales vary by store, stop sending the same size mix of every product to every store. Match sizes to each store's pattern so small sizes do not pile up in one location while large sizes sell out in another.
6. 6Evaluate assortment by marginal productivity, not average sales per SKU. Rank SKUs by dollar sales rate, plot cumulative revenue as SKUs are added, and find the categories where the weakest SKUs barely sell. Delete those SKUs and move the space and purchase dollars to categories where the worst performers still sell well.
7. 7Correct sales data for assortment distortion before reading demand from it. A flavor that is poorly placed or stocked can show low sales despite strong true demand (the second-ranked flavor in the Yummy Cakes example). Estimate true demand, then build the assortment with a greedy rule under the SKU-count constraint.
8. 8Check that sales growth is clean. Compare inventory growth with sales growth each period, and treat inventory growing as fast as or faster than sales as a warning that revenue is being bought with markdowns.

A jeweler has a 42 percent gross margin and a 1 percent net profit margin, in line with the 2000–2006 average for publicly traded jewelry retailers. Better forecasts and inventory management lift its in-stock rate from 90 to 93 percent, which raises sales 3 percent. Gross margin on the extra sales is 0.42 × 3% = 1.26% of revenue. Store outfitting, provisioning, and staffing costs are already paid, so nearly all of it drops to the bottom line. Profit rises from 1 percent of revenue to about 2.26 percent, an increase of more than 100 percent. The goods cost no more to produce or ship than unwanted goods, which would sell at perhaps fifty cents on the dollar. That is why the passages call this the difference between bounty and bankruptcy.

• If shoppers arrive with a clear purchase intent and leave empty-handed, then fix in-stock rate and forecasting before adding traffic drivers.
• If a store-SKU sells about one unit a month, then do not forecast at that level; group similar stores and SKUs first.
• If a category's weakest SKUs add almost no revenue, then cut them and move the freed space and dollars to a category whose weakest SKUs still sell well.
• If store size or product-type sales patterns differ by location, then localize the mix instead of shipping one chain-wide distribution.
• If sales grow but inventory grows as fast or faster, then treat the revenue as low quality and expect markdowns, as in Saucony, Bombay, and Jos. A. Bank.

• Buying more of whatever sold last week. The tell: slow sellers (about one unit per month per store) produce noise, and buys chase random fluctuations.
• Drowning in reports. The tell: an eight-inch weekly size-selling report that buyers have no time to read, and uneven size leftovers across stores.
• Discarding history because styles change. The tell: planners hold that old-product sales are worthless, while seasonality, size mix, and price elasticity persist for years.
• Ranking categories by average sales per SKU. The tell: a few star SKUs prop up a category whose weak SKUs hold space and capital that a more evenly productive category could use.
• Reading assortment-distorted sales as true demand. The tell: a product that is poorly stocked or placed shows low sales even though customers love it, and gets cut.

• ▸Category/store cluster: name the scope you are analyzing.
• ▸Current in-stock rate: percent of demand-days with product on shelf.
• ▸Target in-stock rate and expected sales lift: the planned change and its percent effect on sales.
• ▸Gross margin %: use to compute incremental margin (margin × lift).
• ▸Current net margin %: compare with incremental margin to estimate profit change.
• ▸Causal factors for the period: price, availability, weather, competitor activity.
• ▸Lowest-ranked SKU sales rate: the weakest SKU's dollar sales rate, to test marginal productivity.
• ▸Size or product-mix variance by store: where mix of sales differs from mix shipped.
• ▸Sales growth vs. inventory growth: the two percentages; flag if inventory is growing as fast or faster.

1. 1Write down the baseline: gross margin % and net margin % of revenue. The gap between them is your fixed-cost drag. Jewelry retailers in 2000–2006 ran 42% gross margin and 1% net, so nearly all margin was absorbed before it reached profit.
2. 2Size the prize before buying any analytics. Estimate the in-stock gain, convert it to a sales lift, then compute incremental gross margin = gross margin × sales lift. Divide that by current net margin. The passage assumes all of the incremental margin drops to the bottom line, because the fixed costs are already paid.
3. 3Compute two unit costs for every item. The underbuy cost is normal price minus cost, the margin lost per missed sale. The overbuy cost is cost minus markdown price, the loss per leftover unit. Underbuy is an opportunity cost and never appears on the income statement, so you must calculate it yourself.
4. 4Audit last season's buys item by item. For each item, record sales = min(buy, actual demand), markdown units = buy − sales, and lost sales = demand − sales. Then price them: markdown loss = overbuy cost × markdown units, and lost margin = underbuy cost × lost sales. Put the two totals side by side. Expect lost margin to be larger.
5. 5Replace the single-point forecast with a demand distribution. For each candidate quantity, test the marginal unit: (price − cost) × P(it sells) − (cost − markdown price) × P(it doesn't) = 0. P(it doesn't sell) is the cumulative probability that demand is at or below the prior quantity. Buy until that expected value turns negative.
6. 6For seasonal items, split the decision into an initial buy and a reorder. Convert early-period demand to its share of season demand (14 weeks ≈ 0.55 of 26). Test the marginal unit using the early-period cumulative probability for the sell side and the season cumulative probability for the markdown side. Automate it in Excel rather than reading a chart.
7. 7Watch the profit signal investors watch: inventory growth against sales growth. Falling inventory on flat sales foreshadows margin expansion. Inventory growing as fast as, or faster than, sales foreshadows markdowns and missed earnings.
8. 8Add order in proportion to scale. Around fifty stores, profitability discipline starts to matter. Add it to product selection, pricing and buy quantities, without removing the creativity that draws customers.

Take a jeweler with a 42% gross margin and a 1% net margin. Better forecasts and inventory management lift in-stock rate from 90% to 93%, which raises sales 3%. Incremental gross margin is 0.42 × 3% = 1.26% of revenue. Because store outfitting, provisioning and staffing costs are already fixed, that margin probably drops straight to the bottom line. Profit rises from 1% to about 2.26% of revenue, an increase of more than 100% from a 3-point in-stock gain. The cataloger case shows the same asymmetry at item level. Its blue vest cost $33 and sold at $65, with a $32 loss per leftover unit. Overbuying left 66 vests against demand of 29, a $2,112 markdown loss that exceeded the margin earned. The red cardigan carried an $83 margin ($160 − $77). Buying 86 against demand of 132 left 46 units unsold, a $3,818 lost margin that never shows on the income statement. The lost margin on the cardigan exceeded the markdown loss on the vest.

• If gross margin is high but net margin is thin, then treat demand-supply matching as a profit lever. The incremental margin from extra full-price sales is not diluted by fixed costs.
• If the expected net profit of the marginal unit, (price − cost) × P(sell) − (cost − markdown) × (1 − P(sell)), is positive, then buy it. Stop at the first unit where it goes negative. For the navy turtleneck, the 94th unit is +$1.44 and the 95th is −$0.87.
• If you must choose where to look first, then check lost margin before overstock. Lost margin dwarfs the cost of ditching overstocks, even though managers complain mostly about the leftovers.
• If inventory grows as fast as sales, or faster (Saucony late 2004; Bombay up 50% against 19% revenue growth), then expect margin pressure and markdowns. If inventory falls on flat sales, then expect gross margin to rise.
• If you are pre-fifty stores and still proving the concept, then prioritize execution and growth. Once you reach roughly fifty stores, add profit discipline, but stop short of stifling creativity.

• Counting only markdowns as inventory cost. The tell: the team can name its overstocked items but cannot state lost margin by item, because opportunity cost is not on the income statement.
• Expecting the 3% sales lift to produce only a 3% profit gain. The tell: the business case ignores the fixed-cost leverage, so it understates the payoff, which is more than 100% on a 1% net margin.
• Buying the committee's average forecast as a single number. The tell: huge overbuys on one item (66 leftover vests against 29 demanded) and large shortfalls on another (86 cardigans against 132 demanded).
• Inserting too much order. The tell: customers leave because the creative, exciting assortment is gone. Too little order loses profit; too much loses customers.
• Reading revenue growth as health while inventory races ahead. The tell: sales are up (Bombay +19%) while inventories are up far more (+50%), and management goes quiet on the numbers.

• ▸Gross margin % and net margin % of revenue: enter both; their gap shows the fixed-cost drag.
• ▸Current in-stock rate and realistic target: enter both (e.g., 90% → 93%).
• ▸Expected sales lift %: enter the resulting sales increase (e.g., 3%).
• ▸Incremental gross margin % of revenue: gross margin % × sales lift %.
• ▸Profit uplift: incremental margin ÷ current net margin, assuming full drop-through. Note any reason it would not.
• ▸Per-item underbuy cost: normal price − cost.
• ▸Per-item overbuy cost: cost − markdown price.
• ▸Last season audit per item: buy qty, demand, sales (min of the two), markdown loss, lost margin.
• ▸Marginal unit test: candidate quantity, P(sell), P(not sell), and expected net profit = underbuy × P(sell) − overbuy × (1 − P(sell)). Raise the quantity until this turns negative.

1. 1Adopt the investor's vantage point first. List only what an outsider can see: reported inventory, sales, gross margin, SG&A, guidance and public commentary. The chapter stresses that investors work from a subset of your internal data, so build your read from that subset, not from your internal dashboards.
2. 2Compare inventory growth with sales growth year over year, not inventory in isolation. Saucony is the template: sales were flattish while inventories fell about 20% year over year, and Berman read that as a forecast of better sales, earnings and share price. The reverse pattern, inventory building while sales weaken, is the warning.
3. 3Do not rely on raw inventory turns. The source says to modify the metric so it describes your actual inventory levels and lets you speak to sharp-eyed investors. Add a trend line and the sales context to the turns figure, and say which adjustment you made when you present it.
4. 4Convert any inventory or in-stock change into profit with the fixed-cost logic. Take gross margin multiplied by incremental sales, and assume nearly all of it reaches the bottom line because store costs are fixed. Compare that figure with current net margin, and express the result as a percentage change in profit.
5. 5Check markdown exposure as the other side of the same ledger. Markdowns rose from 8% of sales in 1970 to nearly 30% in the 1990s. Excess stock sells at perhaps fifty cents on the dollar, while goods customers want sell at full price. Rising markdowns alongside a store full of unwanted items signal that supply is mismatched with demand.
6. 6Stress-test the earnings outlook for a miss. Berman expected a miss in Q4 because inventory and sales were already weak. Ask whether guidance depends on a consumer 'saving' you. If it does, expect repeated guidance cuts, not one.
7. 7Run the same scenario test on two peers. The source compares retailers A and B, both with 40% gross margin and $180 million SG&A, and varies same-store sales growth. Hold margin and cost constant, vary one driver, and see which retailer's earnings and valuation move.
8. 8Plan your communication. Even if you are private, record your inventory story (level, trend, cause, fix) in a form an investor like Berman could verify against your public numbers.

Take the source's jeweler. Publicly traded jewelers in 2000 to 2006 averaged 42% gross margin but only 1% net profit. Raise the in-stock rate from 90% to 93% through better forecasting and inventory management, and sales rise 3%. Incremental gross margin is 0.42 × 3% = 1.26% of revenue. Almost all of it drops to the bottom line because store fixed costs do not change. Profit goes from 1% to about 2.26% of revenue, an increase of more than 100%. An investor who sees this capability expects higher earnings, which is the logic behind Wall Street paying more for well-run inventory. The reverse is Berman's short case. After early-November sales weakness was admitted, the stock fell 20% to $10. Weeks later earnings guidance was cut and it fell another 20%. After Christmas guidance was cut again and it fell another 20%. Berman called it the 'classic inventory/earnings relationship.' In about a month and a half the stock was down 50%, driven primarily by inventory mismanagement and weakening sales.

• If inventory is falling while sales are flat or rising, then treat it as a leading indicator of better earnings and valuation (the Saucony 2003 call). Verify that the decline is not driven by stockouts.
• If sales are weakening and inventory is not shrinking, then expect an earnings miss and assume guidance cuts will come in steps, not once.
• If you are estimating the value of a service-level improvement, then multiply gross margin by the incremental sales and compare the result with current net margin. Thin net margins mean the percentage profit gain is large.
• If markdown rates are climbing while shoppers cannot find the item they came for, then diagnose assortment and forecasting, not demand volume.
• If you are presenting turns to investors, then show the modified version with its context, because they will adjust it themselves.

• Reading inventory declines as automatically good. The tell: sales are falling or in-stock rates are dropping alongside the inventory decline, so you are mistaking lost availability for efficiency. The Saucony case worked because sales were flattish, not falling.
• Using raw inventory turns as the whole story. The tell: you cannot explain what inventory level or trend sits behind the turns figure when asked.
• Assuming one bad quarter is the whole correction. The tell: guidance was cut and the stock fell, yet inventory and sales trends are unchanged. Berman's example shows three successive 20% drops.
• Sizing improvement benefits against revenue instead of profit. The tell: a 3% sales gain is described as a 3% gain, when with 1% net margins it can double profit.
• Ignoring the investor's data limits. The tell: your internal story relies on detail an outsider cannot see, so public numbers contradict your message.

• ▸Company/peer and period: name the retailer and the fiscal periods compared.
• ▸Inventory change YoY (%): taken from public filings.
• ▸Sales change YoY (%): total and same-store, if available.
• ▸Inventory-minus-sales gap: inventory growth less sales growth. A positive gap with weak sales is a warning; a negative gap with stable sales is a positive signal.
• ▸Gross margin % and net margin %: needed to size any profit impact.
• ▸Profit-impact test: gross margin × assumed incremental sales, divided by current net margin, to give the % profit change.
• ▸Markdown trend: direction of markdowns as a % of sales, and any evidence of stockouts alongside them.
• ▸Guidance status: any cuts, tone, or reliance on a consumer rebound.
• ▸Stance and next check: bullish, neutral or bearish, with the next data point that would change your view.

1. 1Define every SKU as a bundle of attributes before you forecast anything. For sheets, that means size, color, fabric type, and thread count. For the tire and snack-cake cases, use the equivalent attribute set. This turns a new item into a combination of things you have already sold.
2. 2Pull sales history by attribute value, not by SKU. Most products sell about one unit per store per month (the authors' 32-retailer finding), so store-SKU data is too thin to show patterns. Aggregate by grouping similar stores and similar SKUs until the sales rate supports a pattern.
3. 3Record the conditions under which each sale happened. Price and product availability always matter. Weather and competitive activity often do. Raw sales without these causal factors mislead, because a stockout looks like weak demand.
4. 4Estimate each store's market share for each attribute value (twin vs. queen, sateen, 400-thread-count, pink) from prior sales. Combine the shares into a demand forecast for any candidate SKU, including one never sold, such as pink twin sateen 400-thread-count sheets.
5. 5Use history for the demand parameters that endure for years while products turn over: seasonality, size mix, which product types sell best at which stores, price elasticity, and category growth trends. Do not discard history just because the style is gone.
6. 6Keep the buyer in the loop and encode the buyer's logic. The model formalizes what a talented buyer does: carry all attribute values, overweight the ones customers want. Leave the overarching strategic planning process alone and inject the analytics only into item selection, which keeps the project tactical.
7. 7Audit data accuracy in parallel with optimization. Check inventory records against physical counts, and classify discrepancies as receipt errors, sales/exit errors (including theft), or items in the wrong place (such as the back room). Run the replenishment algorithm on corrected data.
8. 8Treat the model as a microscope. Use it to see more clearly and to produce well-informed projections, not to replace judgment, experience, or common sense, and not to eliminate risk.

In the book's inventory-optimization case, a replenishment algorithm was already saving about $16,000 per week in one department. A physical inventory audit then ran, and recorded lost margin jumped from $20,000 to $42,000. Margin had not truly collapsed. Stockouts that had existed for some time, because the system believed items were in stock, were suddenly detected. The authors concluded the true cost before and after the algorithm was higher by that $22,000 of newly visible lost margin, and the weekly benefit of optimization stayed about $16,000. Once the audit ended, the algorithm optimized against accurate data and drove cost lower than before. This 'accurate data gain' was of similar magnitude to the optimization gain obtained on imperfect data. The lesson is that you pursue both optimization and data accuracy. Treating them as rivals is misguided.

• If a store-SKU sells about one unit a month, then do not forecast from its own sales; aggregate by grouping similar stores and SKUs first.
• If you must forecast a SKU you have never sold, then decompose it into attributes and build the forecast from the sales history of each attribute value.
• If sales data looks weak or odd, then check price, availability, weather, and competitor activity before concluding demand changed.
• If the system says an item is in stock but customers cannot find it, then investigate the three causes: receipt errors, sales/exit errors (including theft), or misplaced stock. Do this before trusting the algorithm's output.
• If cross-functional change is politically infeasible, then start with a tactical project that leaves strategy and structure alone and applies analytics only to item selection.

• Dismissing history because products change. The tell: planners cite 'History's a mystery; trend's your friend' and use only current-style sales, ignoring stable parameters like seasonality, size mix, and price elasticity.
• Chasing what is selling well from raw store-SKU sales. The tell: you chase noise because most items sell about one unit per month per store, and you ignore the conditions of each sale.
• Optimizing on bad inventory records. The tell: the algorithm reports strong results while customers still fail to find items the computer says are in stock, and an audit makes lost margin jump (here from $20,000 to $42,000).
• Debating optimization versus data accuracy as an either/or. The tell: investment goes to one while the other decays, so the gain from the neglected side is left unclaimed.
• Letting buyers' intuition drift from the customer, or overselling the model. The tell: buyers who do not resemble their customers buy heavily by guesswork at season start (the plus-size apparel retailer's case), or the quants-style model is trusted to remove risk instead of informing judgment.

• ▸Category and store group: name the category and the cluster of similar stores you aggregate to reach a usable sales rate.
• ▸Candidate SKU: the item you want to forecast, including any never sold before.
• ▸Attribute list: break the SKU into attributes (e.g., size, color, fabric, thread count).
• ▸Attribute value history: for each attribute value, record store-group sales share from prior sales.
• ▸Sales conditions: log price, in-stock status, weather, and competitor activity for the period.
• ▸Stable parameters applied: note seasonality, size mix, price elasticity, and category growth used.
• ▸Combined demand forecast: state the forecast for the candidate SKU built from the attribute shares.
• ▸Buyer override and reason: record any judgment adjustment and why, for later review against actuals.
• ▸Inventory record check: record the physical count versus system count and the cause of any gap (receipt, sale/exit, misplaced).

1. 1Screen the category first. Use the method only where demand patterns change slowly (grocery, hard lines, tires, snack cakes). For fashion apparel, where 'there is nothing as boring as last season's hot seller,' switch to the Urban Outfitters approach: buyers watch current in-season sales and flag which attributes are trending up or down.
2. 2Model each SKU as a bundle of attributes. Sort them into three types: functional (size; no substitution across values), price/value (good-better-best; substitution is likelier the closer a product's price is to the customer's preferred price/value point), and taste (flavor, color, fabric type).
3. 3Define attribute values and group the ones customers cannot tell apart. The authors collapsed many warranty levels into low, medium and high for tires, and many package sizes into single-serving and family for snack cakes. Fewer values means fewer parameters to estimate.
4. 4Set the substitution paths in the model. Functional attributes get none (a twin-sheet buyer won't take a queen). Price/value attributes substitute toward the nearest price/value point. Taste attributes substitute across values.
5. 5Estimate from recent sales history the share of demand for each attribute value, at the level you plan to act on: chain, cluster or individual store. Do this separately for each location, since 'we sell more black in Manhattan than in Madison' is exactly the variation you want to capture.
6. 6Forecast demand for any candidate SKU, including ones never sold, by multiplying the attribute-value shares (pink × twin × sateen × 400-thread-count). Run this across the full candidate list, not only current SKUs.
7. 7Optimize. Select the SKU set that maximizes forecast revenue (or profit) within the store's capacity, using the substitution paths. Produce add and drop lists per store, then run what-if comparisons of chain-wide versus localized assortments to size the gain from localizing.
8. 8Implement a portion of the recommendations, not all of them, and measure the lift. Discount the forecast, because demand patterns shift between the sales-history window and the implementation date.

The authors applied the method to a national tire retailer. The analysis recommended changes involving a list of forty-seven SKUs. The retailer implemented only a portion of them: it added eleven of the recommended SKUs and deleted eleven of the SKUs the analysis recommended dropping. The authors estimated these changes would raise revenue 13 percent. The retailer realized a 6 percent increase. The authors attribute the shortfall to demand patterns changing between the sales-history period and implementation. Even so, 6 percent is large next to the 3 to 4 percent annual revenue increase Canadian Tire reported from existing stores in 2005 through 2007. Plan on realizing roughly half the modeled lift, and read the gap as the cost of the stable-preference assumption, not as proof the method failed.

• If the category's demand patterns change slowly (grocery, hard lines), then forecast from sales history with the attribute model; if it is fast-moving fashion, then track currently trending attributes within the season instead of extrapolating last season.
• If two attribute values differ so slightly that customers likely can't tell them apart (many warranty levels, near-identical package sizes), then group them into one value.
• If the attribute is functional (size), then assume no substitution; if it is price/value, then assume customers trade to the nearest price/value point.
• If one store or cluster shows a different attribute mix from the chain (more plus sizes, more black), then estimate and optimize at that level and use what-if runs to test whether localizing pays.
• If you need to compare category space, then use Borders's relative sales per title (category sales in the store divided by titles carried): add titles in high-RST categories and cut them in low-RST ones, for example by ranking categories into thirds.

• Treating attribute independence as true. Queen buyers may prefer 400-thread-count more than twin buyers do, so multiplying shares overestimates demand for combinations like 400-thread-count twins. The tell: forecasts for niche combinations run well above what comparable combinations actually sell.
• Applying history-based forecasting to trendy categories. The tell: the model keeps recommending last season's winners in fashion.
• Leaving attribute values ungrouped, so the model carries too many near-identical values. The tell: many SKUs with tiny shares and unstable estimates.
• Expecting the full modeled lift. The tell: the plan books 13 percent while the realized result is about 6 percent, because preferences drifted after the history window.
• Leaving the decision to buyers' tastes or a data dump. The buyers' tastes may differ from the customers', and raw reports (the eight-inch size-selling printout) bury the pattern. The tell: the same size mix is shipped to every store and some stores end the season with leftover small sizes and others with leftover large ones.

• ▸Category and stability check: name the category and state whether demand patterns change slowly (use model) or quickly (use in-season trend tracking).
• ▸Attributes and type: list each attribute and tag it functional, price/value or taste.
• ▸Attribute values (grouped): list the values per attribute, merging values customers can't distinguish.
• ▸Substitution rule: for each attribute, write none (functional), nearest price/value point, or cross-value (taste).
• ▸Level of estimation: chain, cluster or individual store, with the sales-history window used.
• ▸Share by attribute value: percent of demand for each value at that level, from recent sales.
• ▸Candidate SKU forecast: for each candidate SKU, multiply its attribute-value shares and note whether it has never been sold.
• ▸Add/drop list: SKUs to add and SKUs to drop per store, with forecast revenue change.
• ▸Realization check: modeled lift versus actual lift after partial implementation, with notes on what shifted.

1. 1Quantify the cost of rigidity first. Pull prelaunch forecast error on new products (the source cites 50 to 100 percent) and your current lead time. Flexibility pays because it lets you buy less up front and chase winners, which removes both stockouts and end-of-season markdowns.
2. 2Benchmark your replenishment lead time against peers. In the source's study, Zara, World, and Destination Maternity could replenish a hot seller in two to three weeks, specialty retailers averaged six months, and department stores eleven months. Place yourself on that scale and set a target.
3. 3Build the four capabilities as a set, not one at a time: (1) efficient production of small quantities, (2) a wealth of product versions, (3) accommodation of variation in aggregate demand volume, (4) a short lead time. A short lead time is worthless if minimum batch size equals a six-month supply.
4. 4Attack setup time to make small lots economical. National Bicycle cut setup from about an hour to twenty seconds by fitting servomotors, computer-controlled, to old frame-building equipment, and scanning a bar code so the fixtures self-adjust. Find your longest changeover and ask what automation or reused equipment removes it.
5. 5Move the push-pull decouple point toward the customer. National's custom factory runs a pull protocol (batch size 1, two-week lead time) beside a push mass-production plant. Decide which steps are built to forecast and which wait for an order or an early sales signal.
6. 6Convert the lead-time gain into a buying policy. Split the single season buy into a preseason purchase plus an early-season purchase. Buy less initially, read early sales, and reorder winners. The source says going from one buy to two has a huge impact.
7. 7If you don't own the supply chain, negotiate flexibility with suppliers through collaborative planning, forecasting, and replenishment (CPFR), as Best Buy did with Hewlett-Packard.
8. 8Align merchandising and logistics together. Merchants who can't act on early sales data and logistics teams who won't shorten lead times each use the other as an excuse. Run one cross-functional project with a shared target.

The source simulated a 27-item category with a twelve-week lead time. Smarter, risk-adjusted buying (hedging by forecast margin of error instead of buying exactly the forecast and then chasing winners) raised margin by an amount equal to 3.5 percent of base-case revenue. Cutting lead time from twelve weeks to nine, as airfreight might allow, added a further 4 percent of revenue to margin dollars. The mechanism is that a nine-week lead time lets you buy less initially, which cuts markdowns, and jump on winners sooner, which avoids lost sales from stockouts. Note that the lead-time gain was larger than the gain from better buying alone. A three-week reduction, with no change in forecasting skill, was worth more than the analytical refinement.

• If your lead time exceeds the product's selling life, then you get only one buy. Shorten lead time enough to get a second, early-season buy before investing in forecast refinements.
• If your minimum batch size exceeds a few weeks of demand, then fix setup time and lot size before paying for faster shipping, since a short lead time cannot be used with big batches.
• If forecast error is 50 to 100 percent, then buy less initially and reserve capacity for reorders rather than buying the point forecast.
• If you don't own the factory, then pursue supplier collaboration (CPFR-style) rather than waiting for ownership; Best Buy and HP show it works.
• If merchants and logistics each cite the other's weakness as the reason not to act, then launch a joint project that improves lead time and early-sales reading together.

• Fixing only one of the four capabilities, for example a short lead time with large minimum batches. Tell: you can ship in two weeks but still must order months of supply.
• Treating shipping speed as the whole answer. Tell: airfreight is used but initial buys are unchanged, so markdowns persist.
• Functional standoff: merchandising says early sales data is useless because replenishment is slow, and logistics says faster replenishment is useless because buyers can't read trends. Tell: both functions have a reasonable excuse and nothing changes.
• Pursuing tactical, functional projects only. Tell: many small improvements consume organizational capacity, and nobody owns the cross-functional change flexibility needs.
• Assuming flexibility means variety only. Tell: you offer many versions but cannot absorb swings in aggregate demand volume.

• ▸Prelaunch forecast error (%): Record the typical error on new products, from history.
• ▸Current replenishment lead time (weeks): Measure order to shelf for a hot seller and compare with the 2-3 week best-in-class benchmark.
• ▸Product selling life (weeks): State how long the product sells, and whether lead time exceeds it.
• ▸Number of buys per season: Record the current count and the target (at least preseason plus early-season).
• ▸Minimum economic batch size (units or weeks of supply): Identify what limits small-lot production.
• ▸Longest setup/changeover time: Name the step and one automation or reuse option to cut it.
• ▸Product versions offered vs. demanded: List the variety gap customers expose.
• ▸Aggregate demand variation absorbed (%): State the volume swing you can handle without stockouts.
• ▸Decouple point location: Name the last step built to forecast and the first built to order or signal.
• ▸Owner and cross-functional team: Name the merchandising, logistics, and supplier leads accountable for one shared target.

1. 1Name the behavior you want before touching any metric. For each decision in the chain (stocking quantity, size-mix, lead time, packaging), write one sentence: 'Agent X should do Y.' Principle 1 says incentives exist to influence behavior, so an undefined target produces an arbitrary incentive.
2. 2Map principal and agent for each decision. The principal delegates and cannot dictate the action. Examples: Anna and Ralph, a retailer and its Indian exporter, a distribution chief and store managers. Write down what each agent is measured and paid on today.
3. 3Compare each agent's local marginal economics with the whole chain's. Use the Ralph test: his gain from one more copy is 20¢ and his loss on one unsold copy is 80¢. Anna's printing cost is 20¢. Then use the video test: a $3 marginal cost to the chain versus a $60 purchase for the retailer. A gap between the agent's break-even and the chain's break-even tells you the agent will under-commit.
4. 4Separate 'can't' from 'won't' before you spend on training or systems. The Indian exporters knew how to cut lead times and understood the systems. The cause was the cost of carrying fabric and yarn, so the fix was cheaper working capital, not knowledge transfer. Ask the agent directly what stops them.
5. 5Classify the cause using Principle 3. Is the action hidden (you cannot observe it) or is the information hidden (the agent knows something you don't)? Test contractibility: observable, verifiable, enforceable. If all three hold, contract the action directly. Anna could have simply contracted Ralph to stock 584. If not, you must shape incentives.
6. 6Check the diversion problem for any stocking-quantity clause. A retailer can divert part of its purchase to another retailer instead of stocking it, so quantities are usually uncontractible. Assume you must change the economics rather than mandate the quantity.
7. 7Design the change in metrics or rewards, and stress-test it for behavior you don't want. The teaching-to-the-test example shows the risk: a reward on one measure can crowd out the unmeasured ones. Ask what the agent will do to hit the number cheaply.
8. 8Quantify the whole-chain gain, then pilot on a small, unglamorous scope. Show the before-and-after in fill rate, sales, and profit, as Table 5-1 does, so doubters see evidence quickly.

Anna Sheen's Hamptonshire Express shows the mechanism. Printing, demand, and retail price are identical in both scenarios. When Anna sells directly, she stocks 584 copies, with a 98% fill rate, 488 copies of expected daily sales, and $331.40 of expected daily profit. She then moves to an agent, Ralph, at an 80¢ transfer price against a $1 retail price. Ralph earns 20¢ on each extra copy sold but loses 80¢ on each unsold one, so he stocks only 416. Fill rate falls to 81%, daily sales to 404 copies, and expected profit to $282.00. That is 168 fewer copies on the shelf, 84 fewer sales a day, and $49.40 of lost daily profit. The only variable that changed is Ralph's incentive. Because stocking quantity is uncontractible (diversion), Anna cannot simply order him to stock 584. She has to change his economics, for example by lowering his cost of an unsold copy.

• If an agent's local break-even differs from the chain's break-even (20 rentals versus 1 rental in video), expect under-stocking, and restructure the economics rather than exhort.
• If a partner says they know what to do but do not do it, treat the cause as incentives, not skill. Look for the cost they bear, as with the exporters' working capital, and reduce that cost.
• If the action is observable, verifiable, and enforceable, write it into the contract. If any of the three fails, shape the reward instead.
• If a manager is evaluated only on their own function's efficiency (the distribution chief), expect them to ignore downstream stockouts and markdowns. Add a downstream measure to their scorecard before approving their initiative.
• If a proposed metric can be hit by narrowing the work (teaching to the test), add a counter-measure or drop the metric.

• Treating a misaligned incentive as a training problem. The tell: the agent can explain the right practice and still does not do it (the Indian exporters).
• Optimizing a department on its own efficiency score. The tell: a proposal that raises warehouse picking speed, such as a uniform size-mix, while the owner says 'that was the store managers' problem.'
• Mandating a quantity or behavior that cannot be observed or enforced. The tell: the contract specifies a stocking level, but product is being diverted elsewhere.
• Designing the reward without testing for unwanted behavior. The tell: the measured number improves while unmeasured dimensions deteriorate, as with teachers drilling only basic skills.
• Designing without seeing the operations. The tell: packaging that cannot carry bottled drinks, so staff unpack and repack by hand. A ten-cent-per-case improvement was never made.

• ▸Decision/behavior at issue: state the one decision (e.g., stocking quantity, size-mix, lead time) you want changed.
• ▸Principal and agent: name who delegates and who decides.
• ▸Agent's current metric and reward: write exactly what the agent is measured and paid on.
• ▸Agent's local break-even vs. chain break-even: note the gain per extra unit sold and the loss per unsold unit for each (e.g., 20¢ vs 80¢, or $60 vs $3).
• ▸Can't or won't?: record whether the agent knows how and lacks the incentive, and what cost they cite.
• ▸Hidden action or hidden information?: note which applies, and whether the action is observable, verifiable, and enforceable.
• ▸Proposed change to metric/reward: describe the revised incentive in one sentence.
• ▸Unwanted behavior check: list how the agent could game it or what unmeasured work could be neglected.
• ▸Whole-chain impact estimate: record the expected change in fill rate, sales, and profit (as in Table 5-1) and the pilot scope.

1. 1Reject the payroll-as-percent-of-sales budget as your evidence. Most retailers set next month's payroll as a fixed percentage of forecast sales, so high sales and high payroll move together. Sales drive payroll, and the correlation tells you nothing about what extra labor would earn.
2. 2Find the natural experiment in your own data. Actual payroll deviates from planned payroll somewhat randomly: associates can't be hired, part-timers won't commit to hours, or banked hours are used in later months. Pull store-month records of planned payroll, actual payroll, sales and terminations.
3. 3Compute the ratio of actual to planned payroll for every store-month. Sort by it and group into quintiles. In the source study the lowest quintile ran about 14 percent below plan and the highest nearly 50 percent above.
4. 4Deseasonalize sales against each store's own average. Compare each quintile's sales to that average to see whether sales fall when payroll is low and rise when it is high. Deseasonalizing is what separates the labor effect from calendar effects.
5. 5Convert the result into sales per additional payroll dollar, then into gross margin per dollar. In the source, $9.52 of sales per extra payroll dollar at roughly 40 percent gross margin gives $3.81 of gross margin per dollar spent. Compare that return with your cost of labor.
6. 6If you can, run a deliberate test. Raise payroll 10 percent in about fifty stores and cut it 10 percent in another fifty, then compare deseasonalized sales. Use this to confirm the causal direction.
7. 7Tie staffing to what customers value. Survey which factors drive the shopping decision (fast checkout, in-stock on basics, easy to find items, courteous and helpful associates), and assign labor to registers, shelf checks and back-room-to-floor moves accordingly.
8. 8Budget for turnover. Store associate turnover is typically about 100 percent a year, so hiring is continuous. Treat hiring capacity and associate knowledge as constraints, and don't assume budgeted hours equal delivered hours.

The source retailer had more than 500 stores and 17 months of data, giving over 6,000 store-month observations. The authors sorted them by actual payroll divided by planned payroll and split them into quintiles. The bottom quintile averaged 14 percent below plan and the top nearly 50 percent above. Deseasonalized sales rose sharply relative to store average in the high-payroll quintiles (Figure 6-1). Averaged over stores and months, each additional payroll dollar produced $9.52 in sales. At about 40 percent gross margin, that is $3.81 of gross margin per dollar, or 40 percent of $9.52. The authors call this a great deal and ask why the retailer wasn't hiring more. Their answer comes from a store tour with grocery CEOs. After an acquisition, a regional store showed fallen staffing, dispirited employees and unhappy customers, consistent with a cost-minimizing strategy that held minimum staff at minimum wage.

• If gross margin per incremental payroll dollar (sales response × gross margin %) is well above $1, then add hours. If it is near or below $1, hold payroll.
• If payroll is set as a fixed percentage of forecast sales, then don't read the sales-payroll correlation as causal. Use the deviations of actual from planned payroll, or run a 10 percent up and 10 percent down store test.
• If sales in a store run above its seasonal average, then ask whether staffing explains it before crediting other causes.
• If customers rank checkout speed, in-stock, findability and helpful associates highest, then put added labor on registers, shelf checks and back-room replenishment first.
• If labor budget is being cut to hit a cost target, then check the sales-response estimate first. Cutting head count and wages can look efficient while losing the sales those hours produced.

• Reading the correlation between payroll and sales as proof that labor drives sales. The tell: payroll is set at a fixed percentage of forecast sales, so the two move in lockstep by construction.
• Skipping seasonality adjustment. The tell: high-payroll months are simply peak months, so the 'effect' vanishes once sales are compared with the store's own seasonal average.
• Managing labor as a pure cost to minimize. The tell: minimum staff at minimum wage, dispirited employees, and customers who look unhappy.
• Assuming budgeted hours equal worked hours. The tell: about 100 percent annual turnover, unfilled hires and part-timers who won't commit to hours, so actual payroll runs below plan.
• Optimizing for the wrong customer priority. The tell: you invest in service levels customers don't value, such as Nordstrom-style service where shoppers want Target-level prices.

• ▸Store/period set: number of stores and months of data (source: 500+ stores, 17 months).
• ▸Planned vs. actual payroll: per store-month, record both and compute actual ÷ planned.
• ▸Quintile bands: average actual-to-plan ratio for each group (source: about -14% to nearly +50%).
• ▸Deseasonalized sales lift: each quintile's sales minus the store's average sales, seasonally adjusted.
• ▸Sales per extra payroll dollar: the sales change divided by the payroll change (source: $9.52).
• ▸Gross margin %: the retailer's margin on incremental sales (source: about 40%).
• ▸Margin return per payroll dollar: sales per dollar × gross margin % (source: $3.81); compare with $1.
• ▸Customer-priority ranking: top four factors from your own survey, and the associate tasks that serve each.
• ▸Turnover and fill rate: annual turnover %, hiring shortfall vs. plan, and banked hours carried forward.