Optimizing Markdown in Fashion E-Commerce with Machine Learning
The AI team at ASOS uses Machine Learning to improve the customer experience, improve retail efficiency and drive growth. We contribute to the global data science community by publishing our research and by sharing the outcomes of our work at conferences. This article describes work from our recent research paper Promotheus: An End-to-End Machine Learning Framework for Optimizing Markdown in Online Fashion E-commerce in ACM SIGKDD 2022.
We’re hiring! If you’d like to join us to join in our quest to build awesome AI products in fashion, we’re currently looking actively for machine learning scientists, as well as product and technology folks.
The importance of efficient markdown at scale
ASOS is a major fashion e-commerce company, with more than 26.7 million active customers worldwide and >100K products on our platform at any one time. Promotional pricing (“markdown”) is crucial in retail, and especially in fashion e-commerce, where retailers purchase stock far ahead of time, with little information about the trends and fast-changing micro-trends that can dramatically affect week-to-week demand for a given product. Efficient markdown management is crucial to building a profitable business at scale.
Retail price optimization is very challenging to get right. Many retailers, including ASOS, have leveraged the substantive domain expertise in operational teams to successfully implement rule-based markdown strategies. We show here that there are powerful gains to be had from building a price optimization system with machine learning.
Price Optimization involves causal inference
Price optimization is often framed as one of price elasticity: models are built to predict sales as a function of price, using historical data. At first glance, this demand prediction appears to be a standard supervised learning problem. However, price optimization requires price to be varied as a decision variable, with models generalizing beyond the historical pricing policies expressed in the data. In other words, demand prediction takes place in the context of partial information (rather than full supervision) and price optimization involves counterfactual inference (“what would have happened if…”): outcomes are un-observeable for prices other than those actioned historically, and demand curve learning inherently generalises across products with very different characteristics. As a consequence, careful protocols must be adopted for offline evaluation, and it is extremely important to evaluate price optimization systems using rigorous real-world tests (as opposed to offline numerical analysis).
Two novel markdown engines
We present two solutions for managing markdown, which are together deployed to manage markdown at ASOS.com. The first system, “Ithax”, enacts a rational supply-side pricing strategy without demand estimation, and can be usefully deployed as a “cold start” solution to collect markdown data while maintaining revenue control. The second system, “Promotheus”, presents a full framework for markdown optimization with price elasticity and profit optimization.
Our approach produced vast improvements in profitability, even compared to experienced operational teams. These gains were demonstrated in a rigorous real-world test, which also validated the veracity of our offline evaluation and decision making framework. These two novel systems have been deployed to manage markdown at ASOS.com, and both systems can be fruitfully deployed for price optimization across a wide variety of retail e-commerce settings.
Ithax: a cold start markdown optimization engine
Our first markdown algorithm ignores demand estimation and the partial information problem entirely, and instead focuses on the supply-side goal of “selling out all products”. It is economically rational in that it allocates resource disproportionately according to “need”, and aims to select and prices products for a markdown event while keeping within financial constraints. Importantly, Ithax is a powerful “cold start” solution for markdown optimization, as it requires only recent product information to operate with financial control, and can thus be useful for bootstrapping data collection for later price elasticity modelling.
Ithax is a multi-objective optimization algorithm, developed in close collaboration with the Retail Pricing operations team at ASOS. Using a core mechanic inspired by binary search, it successfully constructs markdown product sets that satisfy targets in the following two objectives:
- Stock value: total stock value, at full price (proxy for revenue)
- Stock depth: the ratio between the sum of discounted-price to the sum of full-price value (proxy for profit margin)
Promotheus: a full markdown optimization engine with demand estimation
Promotheus is our full framework for markdown optimization, which uses forecasting to optimize directly for (expected) sales and profit. We overcome the challenge of partial information with a careful offline procedure for model validation and decision making, and validate this procedure with rigorous online testing.
In this framework, Ithax is used to select products to include in the markdown event. Then a price elasticity model is used to forecast likely outcomes across the action space. We use carefully crafted offline validation processes to measure model quality, ensuring that the partial information problem is always present in the train/test split. Crucially, we also then use the offline validation data to constrain the feasible region for decision making. Finally, we optimise depths on a product level to maximise expected outcomes according to business objectives.
Both markdown systems achieve superior profitability compared to decisions made by our experienced operations teams in a randomized online test, with improvements of 86% (Promotheus) and 79% (Ithax) relative to manual strategies. These systems have been fully deployed to manage markdown at ASOS.com, and work continues to iterate and improve on the underlying algorithms.
Read the full paper here.
Eleanor Loh is Lead Machine Learning Scientist for Pricing at ASOS.com. She is a specialist in causal machine learning, and in her spare time enjoys dancing. The paper was co-authored by Jalaj Khandelwal, Brian Regan, and Duncan A. Little. This blog was adapted from the paper by Dawn Rollocks.
