This blog, in addition to a series of posts to follow, is based on a machine learning webinar hosted by the MIT Sloan School of Management on May 4, 2016 and presented by Professor Simchi-Levi of the civil and environmental engineering department.
How Machine Learning Can Help Your Business
Before delving into the specifics of the webinar, it’s important to understand the basics of machine learning and how it can help your business.
Artificial intelligence (AI), or the ability for a computer to learn without being programmed, is not an idea of the future; in fact, it is a very real part of successful businesses today.
Without realizing it, many people interact with some form of artificial intelligence on a daily basis. Anyone that uses speech recognition software has very certainly felt the effects. A great example of this is Apple’s Siri, that has the ability to learn pronunciations and accents.
For a brief moment, visualize how machine learning can add to your business. Think of the time that will be saved and the risk that will be reduced. Let’s get back to what the Professor has to say on the topic.
Forecasting, Learning & Price Optimization
This first part of this webinar covers Forecasting, Learning & Price Optimization, and demonstrates a case study.
Forecasting is introduced in the presentation by focusing on Rue La La, an online retailer and client of Professor Levi. Forecasting to Levi involves the generation of initial forecasts for products. He explained by relating this definition to Rue La La’s current pricing system and general business model.
Rue La La offers extremely limited-time discounts, also called “flash sales,” on designer apparel and accessories. These flash sale discounts are located on a “deals” page with all the current sales available and the remaining time one can buy the product at the flash sale price.
A highlighted flash sale creates a sense of urgency in the in the shopper. Their standard flash sale programming includes the following steps:
- Merchants purchase items from designers
- Items arrive at warehouse
- Merchants decide when to sell items
- Customer purchases items once the event is launched
- Rue La La sells out of the item and ends the promotion OR decides when to sell the merchandise again
However, to order a sufficient amount of inventory, Rue La La needs to utilize demand forecasting, which calls for two main facets: predicting demand for new products, and estimating lost sales.
How to Overcome Demand Forecasting Challenges
Two techniques are used to overcome these demand forecasting challenges: clustering, and models for regression.
Essentially, reverse engineering is used for demand forecasting. In order to learn about market demand by product, a combination of the internal data collected (historical) is consolidated with external data (competitor and general market data). These can both be leveraged to produce more accurate demand predictions.
Both clustering and models for regression are used to reverse-engineer the demand. Clustering is essentially the assignment of some set of observations into subsets called clusters, i.e brand, fabric, color, etc., so that observations in the same clusters are similar in some sense.
This is where real world artificial intelligence can easily be seen as a much-needed addition to your business. If the goal is raising revenue, as I assume it is, the question is: how can it be done while lowering loss? Typically items placed on sale are overstock or items that initially didn’t sell as expected. However, what customers really want is a great deal on what’s hot now.
Machine Learning Forecasting Models
Once Professor Levi defined and gave a general background on the phenomenon of clustering and using regression models in machine learning, he went on to explain the forecasting model used for Rue La La, which analyzes the variables associated with historical data on product sales including variables of product by:
- Color popularity
- Size popularity
- Brand type
- Brand popularity
This is combined with:
- Percent discount (1-Price/MSRP)
- Concurring events
- Number of styles in a subclass
- Relative price of competing styles
This clustering is a method used in reverse-engineering to derive a forecast.
The combination of the aforementioned clustered data, both internal and external, is used to forecast customer demand for a new product through the use of regression trees.
Regression trees involve a number of rules that, if followed, will provide insight into market demand. These various rules are introduced step-wise as a means of narrowing down the price.
Professor Levi explains that from a series of rules, a demand prediction can be produced once a series of logical theoretical hoops are jumped through. This can then inform a suitable price point for the new product. Once cross-referenced with a model for the supply/demand graph, including the elasticity of the product, this becomes a very powerful and accurate tool.
Simply said, your business will be able to run effectively from only one round of machine learning programming. Inventory and price points will be consistently updated, with little to no effort on your part.
Machine learning can determine what the market trends are for a particular product, help gauge the selling price according to competitors, and determine how much inventory is available for sale.
Regression trees use this logic to essentially explain the story of how pricing has determined demand in the past based upon data. This is powerful especially because rules at each logical gate can be tested and modified if they are proven to be outdated or to have changed. These rules can also be formulaic, so they are not exclusively subject to a binary process.
The main objective is a positive impact on revenue without a decrease in market share. To test this, Rue La La identified 6,000 styles, and divided them into five categorized based upon price points. Each style is then split into two – one is used as a control group or standard customer group, and the other is a test group to gauge the impact.
Demand Forecasting Methodology
Professor Levi noted all experiment results were compared against a control case where the pricing strategy of the client was used and compared to pricing system used by Professor Levi and his team. The same comparison is done for measuring the sell-through rate from the control group to test group.
This webinar provided a good example of the methodology behind this new technique of forecasting using machine learning. The main strength of machine learning is the algorithms and price decisions made by the system are self-improving.
This relates to your business because the system learns from old historical data and also from the data it generates as it produces new price points. Even though the lowest priced item did not have an increase in revenue the price point was increased.
Stay tuned to learn about raising the value, which we will explore in the next blog post.
To view a recording of Professor Levi’s webinar, click here.