For some strange reason we are fielding lots of inbound traffic from managers seeking a better model for baseline sales. This has been a persistent issue in the industry for at least 20 years, but only now are the calls coming…odd. For those of you not familiar with the concept for the CPG industry, a sales baseline is an estimate of sales in the absence of retailer price promotion. You can learn more about the concept in this sales baselines blog post and trade promotion optimization white paper.
Just know that without an accurate baseline, trade promotion optimization analysis will be inaccurate, and more importantly many of the decisions made from the analysis will be wrong. Some in the industry will say there are many ways to calculate a baseline (true), and given that, there is no way to tell which model or estimates are the best (false). There actually is a method for measuring which baseline estimates are the most accurate, and if you are interested in the technical details you can find them in this academic paper.
This is just a simple company blog, however. We know that you – our dear readers – don’t have time to decipher skull-crushing mathematical formulas and copious amounts of exclusionary academic jargon. That’s why we will make it easier for you to digest four ways you can measure whether your baselines are accurate.
- No correlation with promotional activity – A valid baseline model should eliminate any correlation between the presence of promotional activity and the baseline estimate. The industry term we use for this condition is “phantom spike,” and its existence is intuitively and theoretically impossible. The theory of “steady-state equilibrium” means that there should be no structural difference in baseline between promoted and non-promoted weeks. You can see from the example below that sometimes this bias in industry models is huge. TABS AccuBase® eliminates these phantom spikes.
- Minimal week-to-week volatility – Again due to steady-state equilibrium we expect the baseline of any given week to be pretty close to the baseline estimates in the weeks immediately before and after. If we do see a change or drift in the baseline it should be a gradual drift or a sudden, short-term shift due to some outside factor (such as an increase in distribution).
- No structural bias in the deviations – This means that when we look at baselines for non-promoted weeks, we should not see that baselines are consistently over or under-estimating baseline sales over continuous periods. Misses should be more-or-less random.
- Minimal error in non-promo weeks – This is pretty basic, no? If you know exactly which weeks a promotion did not happen, the baseline estimates should be pretty darn close to the actual sales, and again, per Point #3, any misses should be randomly distributed over vs. under.
Each one of these tests has a mathematical formula that can be applied to measure concepts such as “No Correlation,” “Minimal Volatility,” and “No Structural Bias.” Again, all that heavy math is in the academic paper link above, but math is just a tool that is proving our intuition and experience.
Once we gather the measures of each, we transform each of the measures to a common scale, weight each of the measures based on importance and then create a weighted average score. Of course, we can get more exotic and develop a Quadratic Loss Function to determine the best model. This accentuates the penalty for deviations from the perfect score.
OK, let’s get out of this rabbit hole and take a step back. It is a commonly held premise that it is impossible to evaluate the “best” model for baseline sales. The objective of this post was to disprove that notion. A best baseline model can be measured, and there is a lot of money at stake for companies to do the due diligence required to make sure they get the best model available.