Markdown Optimization

The predictive core of any markdown engine, using time series forecasting to project inventory depletion at current price points. Models ingest historical sales velocity, seasonality indices, and price elasticity coefficients to identify when a product's trajectory will miss its target exit date. Techniques like Temporal Fusion Transformers capture complex multi-horizon patterns, while Gradient Boosting Machines handle non-linear demand curves. The forecast gap between projected and target sell-through triggers the optimization routine.

The objective function that distinguishes markdown optimization from standard dynamic pricing. Rather than maximizing per-transaction margin, the system solves for total lifecycle revenue given a hard constraint of zero inventory at end-of-life. The algorithm must balance:

• Depth: Larger discounts accelerate sales but destroy margin
• Timing: Early markdowns cannibalize full-price sales; late markdowns risk forced liquidation
• Cadence: Gradual reductions preserve margin perception; steep cuts clear inventory faster This is typically framed as a constrained optimization problem solved via dynamic programming or reinforcement learning.

A mathematical model encoding how a product's value degrades over time. For fashion apparel, decay follows seasonal trend curves with steep drop-offs at collection boundaries. For electronics, decay maps to technology refresh cycles and successor product announcements. For groceries, decay is literal—tied to expiration dates with spoilage risk. The decay function serves as a baseline for markdown schedules, with the algorithm accelerating or decelerating discounts based on real-time demand signals. Weibull distributions are commonly used to model these non-linear deterioration patterns.

Advanced markdown systems optimize across a network of locations rather than treating each store independently. The algorithm evaluates whether transferring slow-moving stock to high-demand locations yields higher net revenue than local markdowns. Key considerations include:

• Transshipment cost: Shipping and handling vs. margin loss from discounting
• Demand correlation: Whether two locations share similar demand patterns
• Cannibalization risk: Whether moving inventory saturates the destination market This transforms markdown optimization from a single-node problem into a multi-echelon inventory routing challenge.

A critical constraint preventing markdowns on one product from eroding sales of full-price items. The system models cross-elasticity of demand between products within the same category or collection. When a markdown is proposed, the algorithm simulates substitution effects: will discounting last season's jacket cannibalize current-season outerwear sales? Techniques like uplift modeling isolate the incremental revenue from a markdown by predicting what the customer would have purchased absent the discount. This ensures markdowns generate true incremental clearance rather than shifting demand from full-price alternatives.

Modern systems increasingly use Deep Q-Networks (DQN) or Proximal Policy Optimization (PPO) to learn markdown policies through simulated market interaction. The agent observes state variables—current inventory, days remaining, demand velocity, competitor pricing—and selects a discount action. The reward function combines revenue generated with a penalty for residual inventory at end-of-life. Unlike rule-based or static optimization, RL agents learn adaptive policies that respond to non-stationary demand patterns, discovering non-obvious strategies like strategic inventory withholding to create scarcity signals.

The statistical foundation for any markdown strategy. This technique quantifies how demand responds to price changes, expressed as a ratio. A product with an elasticity of -2.0 will see a 20% demand lift from a 10% price cut.

• Inelastic goods (0 to -1): Necessities where markdowns fail to stimulate volume
• Elastic goods (< -1): Discretionary items where strategic discounts drive revenue
• Cross-elasticity: Measures cannibalization risk when discounting substitutes

Accurate elasticity curves are the primary input for calculating revenue-optimal markdown depth.

A specialized markdown model where remaining shelf life becomes the dominant variable. The algorithm applies time-decaying discounts to maximize recovery before a product becomes unsellable waste.

• Spoilage curves: Predict the probability of waste at each time horizon
• Salvage value: The floor price, often near zero for expired goods
• Markdown cadence: Optimal frequency of price drops (e.g., 30% at 3 days, 60% at 1 day)

Common in grocery, floral, and fashion with strict seasonal windows. The objective shifts from pure profit maximization to waste-minimized revenue recovery.

A predictive model that quantifies the probability a markdown on one product will erode sales of your own full-price items rather than generating incremental revenue.

• Substitution matrices: Map which SKUs compete within your catalog
• Incremental lift isolation: Separates true new demand from diverted demand
• Category halo effects: Measures if a doorbuster markdown lifts overall basket size

Without this scoring, aggressive markdowns can create a zero-sum game where total category revenue remains flat while margins collapse.

Integrates real-time stock levels, holding costs, and sell-through velocity directly into the markdown calculation. The algorithm dynamically adjusts discounts based on the urgency of liquidation.

• Days of supply: Triggers aggressive markdowns when inventory exceeds demand forecasts
• Holding cost accrual: Factors warehousing costs into the profit function
• Stockout risk penalty: Reduces discount depth when scarcity is high

This prevents the common failure mode of marking down items that would have sold at full price while overstocked items languish.

Moves beyond static elasticity curves by deploying algorithms like Contextual Bandits or Q-Learning that learn optimal markdown policies through continuous trial-and-error in the live market.

• State space: Inventory level, time remaining, competitor prices, demand signals
• Action space: Discrete markdown percentages or continuous discount factors
• Reward function: Revenue recovered minus holding and opportunity costs

The agent explores discount depths early in the lifecycle and exploits learned patterns later, adapting to concept drift as consumer preferences shift seasonally.

A real-time calculated lower boundary that prevents margin-eroding algorithmic decisions. The floor is typically a composite of:

• Cost of goods sold (COGS): The absolute minimum to avoid unit loss
• Liquidation value: What a third-party buyer would pay for bulk inventory
• Competitive indexing: The lowest price at which brand equity is maintained
• Channel conflict threshold: Prevents e-commerce markdowns from undercutting wholesale partners

The markdown optimizer can recommend any price above this floor but is hard-constrained from breaching it, protecting brand integrity.