How to build a successful markdown application

Why are markdowns set?

Markdowns are an important pricing strategy for two main reasons:

1. The value of a product often decreases over time, perhaps due to fashion trends or seasonality. If a reduction in perceived value is not met with a reduction in price, customers will be unlikely to purchase. For example, trying to sell a padded jacket at full price in summer is a difficult task. By applying a markdown, more people will be encouraged to buy, resulting in additional revenue that may not have been achieved had the item remained at full price.
2. Retailers have a finite amount of space in their stores and warehouses to hold their inventory. If the current season’s stock is not cleared, there will be no room for the new season’s stock. Markdowns can be used to increase sell through (the speed at which items sell), freeing up space for newer merchandise.

Most retailers want to achieve as much profit as possible from their stock, whilst ensuring it is cleared efficiently. Setting the right markdown price at the right time is crucial to accomplishing this. This is where data-driven markdown optimization comes in! 

Getting started

For a markdown application, we require a price-response function. This function aims to model the relationship between the price and demand of a product. At Peak, we typically use a dual model approach to build a price response function. This involves:

1. Estimating price elasticities for each product
2. Forecasting the demand for each product at its current price for a specified time period

Once we have both components, we combine them to create our price response function. With this function, we’re able to predict the quantity demanded of a product at any price. By incorporating this information into an optimization algorithm alongside business constraints and guardrails, we can generate a set of optimal markdown prices that aim to maximize a desired objective function, such as profit.

Calculating price elasticities

What is price elasticity? A price elasticity measures a product’s change in demand in relation to a change in its price. Let’s say we drop the price of a jacket from $50 to $45 and sales increase from 100 units to 115. A 10% reduction in price has resulted in a 15% increase in demand, therefore the jacket is elastic with a price elasticity of 1.5. 

For our Markdown application, we require a price elasticity for every product. At Peak, we like to incorporate additional features into our elasticity models that may be influencing the price-demand relationship. Examples of such features include stock levels and seasonality. 

To illustrate this, let’s look at the jacket example again. Perhaps the jacket was out of stock in all but a few sizes resulting in constrained sales. Had the jacket been readily available in all sizes, a 10% drop in price may have yielded a 30% increase in sales. By including this information in our models, we ensure that our elasticities are a true reflection of the price change itself rather than other external factors.

Forecasting current price demand

Once we have a price elasticity for each product, we require a model that can accurately predict the demand for each of these products (at their current price) over a specific time period. 

It is up to you as the data scientist to choose the most appropriate time step to forecast for. If a markdown sale is four weeks long, you may want to forecast ahead weekly and sum the four demand values together.

Alternatively, you may find that predicting demand for the total four weeks results in a better performing model. Consider which approach would be more appropriate for the team you are building the application for. If they report on trade weekly, they may wish to have a view of weekly forecasted sales which will aid your decision.

As for which forecasting model to choose, we recommend comparing the performance of a multitude of models against a baseline. Whatever model the business already uses to forecast demand is a great baseline to use.

Another option is to use a naive approach, such as taking the last time period’s sales and projecting them forward.

Combine and optimize

When we know a product’s price elasticity, its current price and its forecasted demand at that price, we can use the price-response function to calculate the predicted demand for any new price. By building this calculation into an optimizer, we can predict the quantity demanded for any given price.

While stakeholder involvement is important throughout the application build, it’s particularly important when setting up the optimization. One user may want to maximize profit subject to clearing a pre-defined level of starting inventory in a given time frame. Another may want to maximize both profit and sell-through rate at the same time. The end users might want something entirely different. 

How to build a successful markdown application with Peak

Now you know how to create a successful markdown application, let me explain how we do it on the Peak platform!

Firstly we have Dock. Dock is where data scientists create AI-ready datasets. For markdown, this typically involves cleaning and transforming transactional, inventory and product data. Transactional data is vital as without it we’d be unable to see the quantity of items sold and at what price they sold for.

Inventory data is useful for adding context to the price-demand relationship and product data is particularly important for identifying items with similar product attributes. This data is usually ingested via feeds within Dock so that our models are trained on and applied to recent data. In turn, our outputs are relevant and reliable. 

Once our data is AI-ready, we head to Factory. Here we can visualize the data, investigate anomalies and explore the price-demand relationship using R/python. Factory is where we build and iterate upon our price elasticity models, demand forecasts and optimization algorithms. It’s also where we incorporate business rules and guardrails into the application to ensure the price outputs are fit for purpose.

For example, the business may require all colorways of an item to be marked down to the same price. We write tailored code within Factory to ensure all of these guardrails are met. Once the steps of a markdown application are built, we use workflows to chain them together resulting in a robust end-to-end application; the output of which is a set of recommended markdown prices.

Work is where the model’s recommended prices are surfaced and made easy to action. If the end user wants to configure the model settings themselves, we may decide to build a dashboard within Work to house the application. For example, the end user might want to select the number of days to forecast ahead for or exclude certain countries from the optimization.

Alternatively, we can easily deploy an API and send the new prices to the user’s own pricing software each time they are requested. Whichever way we choose to surface the new prices we facilitate the end user to action them easily, enabling them to move from insight to decisions to impact rapidly.

If you’d like to learn more about how Peak could help you, there are two ways to get started:

1. Peak build: Leverage our industry experience and our Markdown application, which is scoped and built out for you on Peak. We have done this many times previously and it’s typically the quickest approach.
2. Self build: With access to Peak, your team of data scientists can take these steps and build out your own markdown application.