Tuesday, 20 February 2018

DIY Volume Forecasts from concept tests

At the 2010 AMSRS conference, a paper was presented dealing with the difficulties of using purchase intent scores to estimate sales for a new product.

The claim was made we can’t simply take these scores, however much we may modify or recode them, and use them for this purpose.  So an approach was described, whereby the purchase intent scores for a similar/comparable product that had been launched were used to devise a calibration factor for the scores obtained for the test product.

That’s not a bad idea … however, it’s not quite so simple.

Purchase intent, properly considered, measures only the respondent’s likelihood that they will try the product.  There is no guarantee that, once having tried, the respondent would go on to adopt the product into his/her repertoire (i.e. keep on buying it).

Indeed, if the product turns out to be a ‘dog’, then there is very little likelihood that the respondent will ever purchase that product again.  [Why are dogs used to represent something bad?  I know some very nice dogs.]

So, properly considered, purchase intent scores are only one part of the estimation process.  This is why, some considerable number of years ago, “trial-repeat” modelling arose.  It’s an old approach, but I have yet to find one better.

The underlying premise of trial-repeat modelling is that the steady-state long-run market share achieved by a (new) product will be the product of the long-run levels of trial and repeat purchasing it attains.

Thus (slightly simplified):

MR = TR * RR

where:            MR = steady-state long-run market share
TR  = long-run cumulative trial rate
RR  = long-run repeat purchase rate.

Further (again, slightly simplified):

• Trial comes through initial purchase
• Initial purchase depends on the level of awareness (brought about by advertising and promotion) and on the availability of the product in stores.

Thus:

TR = TP* A* D

where:            TP = long-run probability of trial
A  = long-run level of awareness
D  = long-run level of distribution.

All the above parameters are reasonably readily able to be estimated, except RR - the long-run repeat purchase rate.

TP can be estimated from the proportion of respondents who indicate at the initial interview of a product test that they would buy or choose the test product.

D and A can be specified based on experience or by some other means.

RR - the long-run repeat purchase rate – can actually be viewed as the equilibrium, or steady-state solution of a first-order, two-state Markov chain.

In the table below, R1 is the probability that someone would be in a state of not having bought the test product, and would move into a state of having bought the test product.  R2 is defined similarly.

Some people will stay permanently in the ‘not buy’ state and others will stay permanently in the ‘buy’ state.  And some will be in a constant state of flux, moving periodically from ‘not buy this time’ or ‘buy this time’ to ‘not buy next time’ or ‘buy next time’.

In the long term, for a group of people it can be shown that mathematically things settle down to a steady state situation, RR, the long-run repeat purchase rate mentioned above.

In fact, RR turns out to be exactly computed as[1]:

RR = R1  /  (R1 + 1 - R2).

Where do R1 and R2 come from?

In some simulated test market models, where respondents are interviewed both before and after the product is placed with them to try, estimates of R1 and R2 are based on the behaviour (or preferences) at the follow-up interview:

• R1 is the proportion of those respondents who did not buy/choose the test product at the initial interview, who did buy/choose at the follow-up interview.
• R2 is the proportion of those respondents who did buy/choose the test product at the initial interview, who did so again at the follow-up interview.

But if we are not actually undertaking a full-scale simulated test market modelling study, what do we do?

From a one stage product test (i.e. where product is not actually placed with respondents), we actually have some information in the answers obtained to the purchase intent questions.

We can assume, for example, that:

“I would definitely buy this product” = 70% probability of purchase
“I would probably buy this product” = 30% probability of purchase.

An estimate of R1 can then be computed directly from this information, perhaps with some additional tweaks.  R2 can take an assumed value, normally around 25% to 35% (depending on the category) and the trial-repeat calculations then worked through to give the long-run market share estimate MR.

In reality, we should not make fixed assumptions about any of the various parameters discussed above.  Ideally, we would conduct a number of Monte Carlo simulations, where values for each parameter are sampled from a specified distribution.  That’s not too hard to do, and it gives us the option of providing an estimated product sales figure that sits between an upper and lower bound, or (better) 25th, median and 75th percentile estimates (based on the estimated values for MR, in combination with population data, and weight and frequency of purchase).

My experience in doing this over very many projects is that the eventual actual sales often lie pretty much in the middle of that range, i.e. more or less coinciding with the median forecast.

[1] William Feller, An Introduction to Probability Theory and its Applications: Vol 1, 3rd edition, Wiley 1968.