AI for Markdown Optimization: When to Discount and by How Much

An apparel retailer selling about $45 million annually through their online store had a simple markdown strategy: if a product did not sell through 60% of its inventory by the midpoint of its planned selling period, it got a 25% discount. If it still had inventory at 75% of the selling period, it went to 40% off. Clearance was 60% off.

When they analyzed the results across 1,800 seasonal styles, they found that 34% of the marked-down products would have sold through at full price or with a smaller discount if they had been given more time. Another 22% needed steeper discounts earlier because the slow start indicated a fundamental demand problem, not a timing issue. The one-size-fits-all approach was costing them an estimated $2.1 million per year in unnecessary margin erosion.

Why Standard Markdown Rules Destroy Margin

The standard approach to markdowns treats all products the same way based on sell-through velocity alone. Product A has sold 40% of its stock with 50% of the selling window remaining, so it gets a discount. But this ignores crucial context. Was the product newly launched and still building awareness? Is it in a category where sales accelerate toward the end of the season (like holiday decor)? Has the product been reviewed positively but just not promoted enough?

Different products have fundamentally different demand curves, and applying the same markdown trigger to all of them guarantees suboptimal results. Products with front-loaded demand (trending items, products tied to a specific event) need aggressive early markdowns if they miss their initial window, because demand will only decrease. Products with back-loaded demand (seasonal basics, gift items as the holiday approaches) can sustain full price longer because demand will naturally accelerate.

The cost of discounting too early is invisible but real. A 25% markdown on a product that would have sold at full price two weeks later destroys margin on every unit sold during that period. If 200 units sell during the markdown window that did not need discounting, and the average selling price drops from $50 to $37.50, that is $2,500 in unnecessary margin loss on a single product.

How the AI Model Approaches Markdowns

An AI markdown optimization model evaluates each product individually based on its specific demand trajectory, remaining inventory, remaining selling time, and the expected response to different discount levels. The model answers three questions: should this product be discounted at all right now, if yes what is the optimal discount depth, and when should the next markdown be applied if the current one does not achieve the target sell-through.

The demand trajectory analysis fits a curve to the product's sales history and compares it to similar products from previous seasons. If a winter jacket has sold 300 out of 800 units in the first six weeks, the model checks how similar jackets performed at the same point in previous years. If similar products typically sold 35-40% of inventory by week six and then accelerated, the current pace is normal and no markdown is needed. If similar products that sold only 37% by week six historically ended up needing 50%+ markdowns to clear, early action is warranted.

Price elasticity estimation tells the model how much additional volume a given discount will generate. A 20% discount on a product with high price elasticity might increase weekly sales by 80%, while the same discount on an inelastic product might only increase sales by 20%. The model learns product-level elasticities from historical markdown data, looking at what happened to sales volume when similar products were discounted in the past.

The optimization then calculates expected margin at each possible discount level. No discount means higher margin per unit but the risk of unsold inventory at season end. A 20% discount reduces margin per unit but accelerates sell-through. A 40% discount accelerates further but at a steep margin cost. The model selects the discount that maximizes total expected margin across the remaining selling period, including the probability of needing a deeper markdown later.

Category-Specific Considerations

Fashion and apparel products have the highest markdown exposure because styles go in and out of favor, sizes fragment the remaining inventory (plenty of XS and XXL, sold out of M and L), and the selling window is relatively short. The AI model needs to account for size curve depletion: even if total sell-through looks adequate, a product that has sold out of popular sizes may need to be marked down because the remaining sizes will sell slowly.

Consumer electronics follow a different pattern. New product launches by competitors can make your existing inventory less desirable overnight. The model needs to incorporate product lifecycle stage and competitive launches as features. A laptop that launched 10 months ago and is about to be replaced by a new model needs aggressive markdown timing that differs from a laptop in its third month on market.

Home and garden products often have strong seasonality with relatively predictable demand patterns. The markdown model for these categories can lean more heavily on historical seasonal curves because the year-over-year demand shape is more consistent than fashion. A patio furniture set follows roughly the same demand curve every year, shifted by weather patterns.

The Implementation Approach

Most retailers implement AI markdown optimization in phases. Phase one replaces the static rules with model-generated recommendations that the merchandising team reviews and approves. The model produces a weekly report showing each product that it recommends for markdown (or removal from markdown), the suggested discount level, and the rationale. The merchandiser either accepts or overrides each recommendation.

Phase two introduces automation for lower-risk decisions. Products below a certain inventory value or in categories where the model has proven accuracy can have markdowns applied automatically within guardrails (maximum discount of 30%, minimum margin floor, no more than one change per week). The merchandiser reviews automated decisions after the fact and can override for the next cycle.

Phase three extends the model to optimize markdown timing across the entire seasonal portfolio. Rather than evaluating each product independently, the model considers how promotional events interact (running markdowns on too many products simultaneously dilutes their impact) and allocates markdown budget across the assortment to maximize total margin recovery.

The data requirements are manageable. You need daily sales data by SKU for at least two years to capture seasonal patterns, historical markdown events with the discount level and date for each, current inventory positions, planned selling periods for seasonal products, and product attributes (category, brand, price tier) for the similarity analysis.

For ecommerce retailers managing seasonal inventory, the difference between a good markdown strategy and a bad one is often 5-15% of gross margin on the affected products. Applied across a seasonal assortment worth millions in retail value, that margin recovery adds up to a substantial number, often enough to fund the optimization system several times over while leaving a meaningful impact on the bottom line.