Welcome to the Listening Post!

To Price or Not to Price, That Is the Question: Predicting the Effect of Price Increases - A Non-Technical Explanation

Tweet

 

Eliot Roth, Senior Manager of Pet Custom Research for Del Monte, began saying, "From the beginning of time, man has been consumed with one great question: How much do I charge?" Businesses have succeed or failed on this.  Like the Chinese curse, we live in interesting times. Del Monte has seen hyperinflation in commodities and then deflation, causing a rethinking of pricing research. The new world challenge is "out of bounds" pricing. The hyperinflation of commodities has pushed us outside the range of past models.

The three traditional pricing models are cost-based pricing (which Finance likes), competitive pricing (which Marketing likes) and economic pricing of the supply and demand curve (which economists like). In the real world, multivariate regression can be used to model pricing effects. For elasticity models, start with a good data source (for CPG, IRI and Nielsen are the suppliers, and Del Monte is a Nielsen house) and then control for other market forces (e.g., store size, seasonality, promotions, direct and indirect competition). 

Key things to keep in mind:

• Use relative pricing, not absolute pricing. What happens when everyone raises their price? Increased fuel costs and commodities costs have raised the pricing of entire categories. The effect differs depending on the category: if it is not discretionary, people still buy it (gasoline doubled but consumption dropped only 5%), but other categories see shifts to substitutes. 
• Relative pricing needs to consider item elasticity and category elasticity.  We did the math of the price increase separately: 3% price increase * 1.5 elasticity (4.5%) for item + 4% category component * 0.5 elasticity for 6.5% elasticity. This was the most predictive model for us.
• Use interaction levels to calculate category price increases. The relative category is defined by evaluating the item's competitors that interact most with it, using a weighted evaluation by interaction.
• Avoid over-fitting models. People overfit the model, with too many variables providing better fit statistics (one model had 4,800 factors) at the expense of predictivity.
• Do not mix sizes. Mixing different priced (different sized) SKUs introduces noise: for instance, a discount on the larger size can increase the average price.
• Transform variables. Avoid overly simplistic models, such as traditional linear models. Data models need to be transformed or normalized to have a better fit and to be more predictive.