02 Diving deeper: category-specific insights
Finding a fit
So, our investigation delved further, systematically evaluating different categories to discern whether the overarching pattern of Double Jeopardy applies universally.
The more rigorous approach we undertook was to repeat the analysis in the following paper: If the Model Fits, Use It: Methods and Benchmarks for Evaluating NBD-Dirichlet Goodness-of-Fit by Carl Driesener, Melissa Baelis and Cam Rungie. The NBD-Dirichlet is the long-established way to model buyer behaviour – it is used to create buyer-based benchmarks like the one described earlier. This paper sets out measures to evaluate whether the NBD-Dirichlet model for a defined category is a “good” fit for a category.
We followed the “good” fitting criteria set out in the paper and examined over 190 UK grocery categories over every year. The category definitions are used by all our clients to measure the grocery market. Private Label is included but has been amalgamated into one to make it like a national famous brand, i.e., it can be bought in almost every retailer.
The findings were telling: approximately two-thirds of the categories exhibited a good fit.
We found similar “good” fit results every year, although, in the latest year, we see more variation than in previous years, with a particularly low pass rate for categories with under 25% penetration.
Splitting the data by category penetration shows that the fit improves as category penetration increases. However, even when the data is at its strongest — when over three in four people buy the category at least once per annum — still, one in five categories are not passing the “good fit” test.
To illustrate good and bad fits, we show multiple categories of varying sizes with individual brands shown against the Dirichlet fit.
To be a good fit: we used the rules set out in the paper: correlation - >=0.9 for Penetration - >=0.6 for frequency; AVE % <=5% for Penetration, <=10% for Frequency; RAAE <15% for Penetration, <20% for Frequency; MAPE <20% for Penetration, <20% for Frequency. For more details: If the Model Fits, Use It: Methods and Benchmarks for Evaluating NBD-Dirichlet Goodness-of-Fit by Carl Driesener, Melissa Baelis and Cam Rungie.
Why do we find many categories that don’t seem to fit?
There could be a myriad of reasons that would be uncovered in a deeper analysis, but as we know, sometimes certain brands are not as national as others — they are strong in a region; sometimes brand sales are more seasonal; sometimes they have functional qualities that create more frequency (loyalty) than you would expect for their attraction (penetration).
Category conundrums
It is clear, for all these reasons and more, that category and brand definitions used really matter. This is not unlike a regression modeller making calls on how to transform/include/exclude explanatory variables to build a “best fit” model. In our case we have only one model to fit – where we can include or exclude brands and their specific formats to find what works best.
For example, to illustrate the analytical work required to find a predictable relationship these would be our simple next steps:
Men’s Skincare: The next step would be to exclude the very low frequency brands, which are more seasonal, hence they have a lower frequency than expected.
Flavoured Milk: It just passes. Would it work better to split the market by dairy and non-dairy to find a better fit?
Instant Hot Snacks: Is Pot Noodle an outlier, and the reason why the fit doesn’t work well for small brands? Is Nissin causing the issue – is that in a different category or does it have less distribution?
Antiseptics and Liquid disinfectants: These are good examples. We expect them to always work. It's worth checking if all brands have similar sub-offers selling in the same proportions—is this why they work so well?
Chocolate Biscuit Bars: Another good example, and how we expect it to work. Look how close McVities (Club) and KitKat are on the chart. But it shows Tunnocks off the line – with a higher loyalty. This makes sense due to the regional strength it has in Scotland.
Toothpaste: All the smaller brands are below the line. Is this because Colgate or Sensodyne is stronger than expected? The latter has more loyalty – it’s a premium brand with strong positioning around reducing teeth sensitivity. Would the fit improve if we split out the sensitive market, as all the leading brands have sensitive products? Are there other potential splits?
Instant Coffee: Another example where all the small brands fall under the line. Is Kenco and Nescafe doing better than expected? Instant Coffee is a mixture of different formats – from sachets to powder to granules to more expensive micro powders. Is this total overview giving us the right benchmarks?
