Price management is no small task for B2B pricing owners. While machine learning can set smarter pricing strategies, understanding how it works in the context of B2B and knowing how to leverage it to deploy highly-relevant pricing guidance, and ultimately, create a more strategic pricing approach isn’t as straightforward. In this four-part series, Zilliant Director of R&D, Pricing Science Amir Meimand will share a summary of how predictive analytics, prescriptive analytics and machine learning contribute to a smart pricing strategy for B2B companies.
Machine learning (ML) algorithms are categorized into three main groups based on their interactions with the environment:
- Supervised learning: Apply what has been learned in the past to new data using labeled examples to predict future events.
- Unsupervised learning: Explore the dataset to recognize and discover a hidden structure/pattern from unlabeled data.
- Reinforcement learning: Learn decision-making heuristics from the environment by taking action and analyzing consequences (i.e., errors or rewards).
Supervised Learning ML
Supervised learning models require training datasets comprised of data and the correct labels. Generally, the larger the dataset, the better the model performs for unknown cases. During the training process, the model should learn to generalize the solution to cope with unknown cases to avoid “overfitting” the data, as illustrated in Figure 1.
Typically, supervised ML is used to solve two major types of analytic problems:
- Classification: Learn to categorize data into pre-defined classes. For example, when initiating a contract with a new customer, it may be useful to know which existing customers are similar.
- Numeric Prediction: Learn to predict a numeric quantity instead of a class. For example, it may be necessary to develop a model to predict the price of a house based on features such as location, square footage, number of bedrooms, etc.
At the heart of supervised machine learning are three main algorithm categories:
- Regression-Based Models: Regression models are used primarily in cases in which the target variable is continuous and the relationship between the target and response variable(s) is linear or can be transformed into a linear form. However, there are some versions of regression models (logistic/non-linear regression) which can be used for classification or are designed to handle non-linear relationships.
- Decision Tree-Based Models: Decision tree models are based on a framework of decisions and their possible consequences. Observations about an item (represented by branches) lead to conclusions about the item’s target value (represented by leaves). Tree models that enable the target variable to take a discrete set of values are called classification trees; in these structures, leaves represent class labels and branches represent conjunctions of features that lead to those class labels. Decision trees where the target variable can take continuous values are called regression trees.
- Bayesian-Based Models: The foundation of these models is the Bayesian probability in which probability is interpreted as the reasonable expectation that depends on both observation and prior information rather than just frequency. These sorts of models are very useful when data are structured in groups and clusters. In this structure, each group has its own prediction; however, the predictions are related due to the hierarchical data structure. Bayesian models enable information to be shared among groups by using the same prior parameters, yet the behavior of each group is influenced by its own evidence.
Applications of Supervised ML in B2B
As discussed in the previous post, many factors affect market prices in the B2B world. In such a complicated situation, a single model may not be able to explain all behaviors accurately. Zilliant employs ensemble learning models which use multiple predictive algorithms and report the aggregate result. The aggregate result of multiple predictive models is less noisy than the result of a single model. In addition, ensemble learning models help prevent overfitting, which is less likely among a majority of models than in a single model. Figure 2 illustrates a comparison of market prices predicted by an ensemble model and a single model in terms of mean absolute percent error (MAPE).
Accurate market price prediction is extremely important in the price optimization process. In Figure 3, the actual market price is P0 and the optimum market price is P*. If the market price is incorrectly predicted as either P1 or P2, then it will incorrectly recommend a price change, either in direction or magnitude.
The ML model also can be used in B2B pricing to estimate price elasticity. Price elasticity is the core of pricing science, as it provides direction on how to adjust and modify behavior to achieve P&L objectives. In general, B2B companies have fewer transactions than B2C companies, thus data sparsity is a real issue that can lead to overfitted models when estimating price elasticity. Zilliant uses sophisticated ML algorithms such as the hierarchical Bayesian model to deal with data sparsity. Since data are structured in a hierarchical format (i.e., product hierarchies), information can be shared among similar segments to yield robust and accurate estimations of price elasticity (see Figure 4).
In Figure 4, each product (represented by an individual box) does not have enough data to estimate the relationship between the change in price and change in quantity. Zilliant’s hierarchical Bayesian model can estimate the price elasticity for each product by sharing the information among all products within a product group.
Stay tuned for Amir's next post as he continues to dig into ML applications into B2B pricing.
About the AuthorFollow on Linkedin Visit Website More Content by Amir Meimand