Data Science That Is Useful for Business

Artem Duplinsky25 May 2024

Problem

A team has sales data with price and marketing spend by date and wants to choose prices or a marketing budget that maximize profit. A model fitted for prediction answers "what will sales be", not "what will happen if we change the lever". Marketing and prices usually move with the economic cycle and seasonality, and companies set them in reaction to sales. A model trained on such data credits marketing for the effects of the cycle, so the budget decision is wrong.

Example from the article: sales and marketing rise and fall together, so marketing looks very effective. Add a macro trend that follows the same rhythm and the apparent effect may be an illusion.

Solution in brief

Separate three kinds of data tasks: describe the past, forecast what you do not control, and build the future by changing levers. The third needs a causal model: draw a causal graph of the variables from domain knowledge, then estimate the effect of each lever with methods that handle omitted variables and simultaneity. A more complex predictive model does not fix this.

Questions this case answers
  • How to estimate what-if effects of price and spend levers from observational data that is confounded by a shared seasonal cycle
  • How to decide whether a decision task needs a predictive model or a causal model before choosing methods
  • How to estimate what-if effects of business decisions (price, marketing budget) from observational data
  • How to tell whether marketing really drives sales or both just follow the economic cycle

Tools and methods

  • Causal graph — variables as nodes, influence as arrows; records the analyst's assumptions
  • A/B testing — breaks the dependence between customer characteristics and the treatment decision
  • Instrumental variables — used for simultaneity when an experiment is impossible
  • CCE — panel-data method for structural coefficients when some variables are missing
Continue reading — it's free

Google or a one-time email code. No password.