How does markdown optimization work for retail?

Most retailers discount by feel. A product sits on the shelf for six weeks, somebody decides 20% off sounds right, and if that does not work they go to 40%. The problem is that intuition leaves a predictable and measurable amount of money behind, and the better approach has been around long enough that the research is settled.

Markdown optimization is the practice of using your own sales data to calculate when to discount, how deeply to discount, and how to sequence price changes across a product's remaining shelf life. Retailers who do this well recover 10–30% more revenue from clearance inventory than those who do it by hand (McKinsey, 2021).

A markdown is any planned price reduction on inventory that needs to move: end-of-season clothing, perishable food, consumer electronics before a new model ships. The goal is never to discount as much as possible. It is to find the lowest discount that still clears the shelf before the product loses all value.

Markdown optimization automates that calculation. Given how many units you have, how many days remain before the season ends or the product expires, and how demand responds to price changes in your specific store, the system recommends a discount depth and start date. It also updates those recommendations as real sales come in, adjusting if early sell-through is faster or slower than expected.

The original adopters were large apparel chains and grocery retailers. Fashion retailers face hard season-end deadlines: unsold spring inventory cannot roll into fall. Grocery stores face literal expiration dates. But the same logic applies to any product with a sell-by window: consumer electronics, home goods, even hotel rooms. The McKinsey Retail Practice estimated in 2020 that markdowns account for 30% of all retail revenue in the US, making this one of the largest levers available to any merchant.

The core math is a sell-through curve. For a given product, the model estimates how many units will sell at each possible price point over each remaining day of the season. It then finds the price path that maximizes total revenue while hitting a target inventory of zero by the deadline.

Say you have 500 units of a jacket with 40 days left in the season. At full price, the model predicts you will sell 8 units per day, leaving 180 jackets unsold. At 20% off, you sell 14 per day and clear everything with a week to spare. At 30% off, you sell 18 per day and clear out two weeks early, but you sacrificed margin on units that would have sold at 20%. The optimization finds the path that avoids both outcomes: leftover inventory and unnecessary early discounting.

What makes the model more than a spreadsheet is that it incorporates real-time feedback. If week-one sell-through comes in 30% below forecast because of an unexpected cold snap, the model immediately recalculates and may recommend pulling the first markdown forward by two weeks. A buyer making that same decision by hand would need to notice the trend, estimate the pace, calculate the revised projection, and propose the change through an approval process. That process takes days. The model updates overnight.

A 2019 study by the Journal of Retailing found that algorithmic markdown tools reduced end-of-season inventory by an average of 18% compared to buyer-led decisions, while improving gross margin by 6 percentage points. The margin improvement comes from timing: the algorithm discounts just enough to sell through, rather than overcorrecting with deep cuts that wipe out margin on items that would have moved anyway.

The model needs four categories of input to produce reliable recommendations.

Transaction history is the foundation. Daily unit sales by SKU, store location, and price point, going back at least two to three full selling seasons. One season of data produces rough estimates. Three seasons lets the model distinguish between a slow week caused by weather and one caused by a genuine demand problem with the product.

Inventory positions tell the model what it is working with: current stock on hand, stock in transit, and any replenishment orders already placed. Without this, the model cannot calculate whether a given sell-through rate will clear the shelf or leave a surplus.

Price and promotion history lets the model isolate price elasticity. If you ran a 25% promotion on this same product two years ago and sales tripled, the model learns that this category responds strongly to discounts. If a similar promotion produced a 20% lift, elasticity is lower and smaller discounts will do the same job.

External signals improve accuracy but are optional. Weather forecasts matter for seasonal apparel. Local events matter for grocery and convenience. Web search volume for a product category can signal whether consumer interest is rising or falling ahead of the data showing up in your transaction log. Retailers who layer in even basic external signals typically see forecast accuracy improve by 8–12% (Gartner, 2021).

The data requirements are why large chains adopted this technology before small ones. You need enough history and enough SKU volume for statistical patterns to be reliable. A store with 50 products and two years of records can run a simpler version of these models. A chain with 5,000 SKUs across 300 locations can run much richer ones.

The honest answer changed considerably between 2018 and 2022. Before cloud-based analytics became standard, building a markdown optimization system meant hiring a data science team, standing up your own infrastructure, and running a multi-year implementation project. That was a realistic option for retailers doing $500 million or more in annual revenue. Below that threshold, the ROI math rarely worked.

Today the economics look different. Software-as-a-service pricing has brought markdown tools into a range where mid-sized retailers can evaluate them seriously.

For a retailer doing $20 million in annual revenue with a 30% markdown rate, that is $6 million in discounted merchandise per year. A 10% improvement in markdown efficiency, a conservative outcome given the published research, recovers $600,000. A $15,000 annual software subscription pays for itself in the first month of operation.

The catch for smaller retailers is data quality, not price. The tools are affordable. Getting clean, consistent transaction data out of a legacy point-of-sale system and into a format the model can use is frequently a two-to-three month project before any optimization starts. Retailers who have already invested in modern commerce infrastructure skip this step. Those running on older systems often need help extracting and cleaning their historical records before the optimization layer can do anything useful.

Timespade has built markdown optimization systems for retailers at both ends of this spectrum. For larger operations, that means integrating the model into existing data infrastructure and building dashboards that surface recommendations to buying teams. For smaller retailers, it often starts with cleaning up transaction history and building a simpler demand forecasting layer before any markdown logic runs on top. The engineering cost for the full stack, data pipeline, forecasting model, and recommendation interface, runs $25,000–$40,000 for a mid-market retailer. A US-based data science consultancy would quote $150,000–$300,000 for the same scope. The difference is a focused global team with no office overhead and no layers of account management between you and the engineers.

Buyers are not bad at their jobs. They carry years of product knowledge, vendor relationships, and market intuition that no model can fully replicate. The problem is that human judgment has predictable failure modes at scale, and they show up in the data consistently.

Buyers discount too late. A 2020 study by Boston Consulting Group found that retail buyers delay their first markdown by an average of two weeks compared to the timing an optimization model would recommend. Two weeks of full-price inventory that is already past its demand peak means lower eventual sell-through and a sharper final cut to move remaining stock.

Buyers also discount too uniformly. When a buyer decides a category is underperforming, they often apply a blanket discount across a whole assortment. The model treats each SKU separately. The black jacket in size medium may need 15% off. The tan jacket in size extra-large may need 40%. Blending those into a single promotion leaves margin on the first and risks stale inventory on the second.

None of this means replacing buyers. The model does not know that the vendor for this jacket line is launching a replacement in four weeks and the buyer should clear the current SKU faster. It does not know that a competitor just ran a similar promotion. Human judgment sits on top of model recommendations, not below them. The retailers who get the most out of markdown optimization treat the system as a calculator, not a decision-maker. Buyers review recommendations, override when they have information the model lacks, and track whether their overrides improve or hurt outcomes over time.

That feedback loop is where the real value compounds. Over two or three seasons, a buying team that regularly reviews model recommendations, even when they override them, develops a sharper read of when their judgment adds value and when they are just reverting to habit.

If you are evaluating whether a markdown optimization system makes sense for your retail operation, the right starting point is a data audit: what transaction history do you have, how clean is it, and what model complexity does that data support. Book a free discovery call and we will walk through it with you in 30 minutes.