Big Data, Machine learning and Pricing

Big Data and real facts instead of intentions 

One of the most common approaches used in the past to infer the value of a product was based on the use of Survey Data.  Where Marketing Analysts show consumers several product profiles, varying Price and other product attributes (ideally orthogonally designed) and ask them to rank these profiles. Then Analysts uses Full profile Conjoint Analysis, Hybrid Conjoint Analysis or Discrete Choice Analysis, to infer the importance of Price (and other product attributes) to the customers and estimate its Price elasticity. The Pricing problem is then framed as an optimisation problem to maximise Revenue or Profit. Tools such as Excel Solver engines (Simplex, GRG and evolutionary) where then used to find the best Price. 

Today, the Big Data enable us to be more factual. Indeed, we no longer need to use declarative consumer's intentions about buying or not and at which price. We can simply use server log data, containing detailed informations about the real choices that customers did. 

Feature engineering for Pricing 

Once we collected all the transactions log data, building the right Dynamic and Customized Pricing tactics is first a matter of Feature Engineering : which means defining and selecting the most important attributes to take into consideration in the Pricing process. 

Those features could be Product related, such as brand, price, color, technical features of the product. They could be Customer related, such as sociodemographic attributes and customer’s purchase history. 

The Pricing could be built using more Advanced features, using the mouse clicks data, the identification of the acquisition channel, and the customer’s interactions with the company's communications campaigns via Direct emails, facebook, twitter etc.. 

Machine Learning Techniques 

Once the Feature Engineering is operated, using the available data sources at the company level, comes the Modelling tasks. 

To the Pricing problem can be modelled as a two-step classifier: At first we predict whether the user will buy the product, and if the prediction is positive, then we predict the price he might be willing to pay. The two classifiers can be trained as a supervised machine learning problem, learning from the historical data and using the engineered features. 

Many classifiers can be used, such as Logistic regression, generalized linear model (GLM), Extreme Gradient Boosting (xgboost), Neural Networks. 

The choice between those techniques depend mainly on the trade-off between the accuracy, scalability of the algorithms we might need.

Aucun commentaire:

Enregistrer un commentaire