A Practical Guide to Forecasting New Products - YouTube

Imagine you have been given a new task at work: You have been asked to forecast sales

of a new product that your engineering team has been working on for the past 18 months.

Most of us, who have lived the world of corporate market research have been there: come up with

an accurate forecast for the new product. Hi, my name is Miklos Kremser, - I am the

Founder and Principal of Choice Based Market Insights and we specialize in helping companies

navigate the challenging world of new product forecasting.

If you say that the challenge of forecasting new products does not cause you heartburn

or other stress related symptoms – I beg to differ. Forecasting new product launches

is stressful. Most of us know that historically, most new product launches are considered a

failure. 75% of consumer package goods and retail products fail to earn $7.5 million

during their first year – which is often the measuring stick for success. To make it

more stressful, there is a myriad of unknowns: Is distribution going be sluggish? Is the

budget going to be adequate? Is there a precise understanding of the target market? In an

article, published by the Harvard business review by Joan Schneider and Julie Hall, the

authors listed 40 different factors that heavily influence the success of a product launch.

To add to your heartburn, there is usually a significant internal pressure to provide

an optimistic forecast. Think about it, your engineering team has just spent 18 months

designing and producing a new product, wholeheartedly believing that this product would disrupt

the industry as we know it. By the time they task you with providing a forecast, it’s

likely the company has already spent millions in man hours and materials. Right now, the

only thing that stands between them and eternal engineering glory… is you. So hurry up!

And hurry up is exactly what you should NOT do. In this video, I am going to provide you

a framework for new product forecasting. I will give you some tools, steps and some equations

all of which will give you more confidence in providing more accurate forecasts.

The very first thing – however – that you will have to do is ask the right questions.

Is there a market for this product? Because if the answer is no – or not very clear

– then the recommendation must be a swift and quick death. Now, the second question

then must be: if yes, how big? Again, if not big enough – the project must be killed

– and you must warn your organization – no matter how unpopular you may become. Only,

if there are signs for a big enough market – can you proceed with the actual forecasting.

Measuring market appeal – and forecasting sales are two completely distinct phases in

this process. And you should never skip the first.

First, I am going to give you some tools for the first phase: Measuring Market Appeal.

Measuring market appeal cannot happen without tapping into the market and actually asking

the market about it. In other words: market research.

Long gone are the days of asking survey respondents “how likely are you to purchase this product?”

These are stated likelihoods – that tends to be biased. For example: if you end of with

a result of 43% “Somewhat likely” – can you turn that into a market size? What does

“somewhat likely mean?” There is a much more useful approach – in

which we calculate market appeal or potential market size by making survey respondents choose

among product alternatives – and by analyzing the choice patterns, we can set up a statistical

model. This is a choice-based conjoint methodology – and there are lots of scientific publications

that claim these methods to be less biased and more accurate.

Here is how it works: First you set up a survey based choice experiment

that survey respondents participate in. Make sure all the relevant product attributes are

included in the experiment. Then, based on the respondents’ choice patterns, you create

a statistical model – that you can use in a market simulation to measure the appeal

of the potential product. This was very high level, so let me dive a bit deeper:

What does this choice experiment look like? Well, let’s imagine a product has three

important attributes about it. For example, these important attributes may be: flavor,

size and price. Or any other relevant attribute. For simplicity sake, let’s pretend these

three important attributes are: triangle, square and circle. Each of these attributes

can have different levels: in our case let’s pretend the triangle can be yellow, or blue

or orange or green. In the choice exercise we use these attribute

level combinations – and present seemingly random configurations as products to respondents,

and ask them to select the one out of these products – that they would be most likely

to purchase. Although it’s best to strive for ‘realistic’

combinations of these attributes – most often it’s fine to have as random combinations

as much as we’re able to suspend our disbelief. After the respondent selected an option, these

options go away – and another seemingly random combination of options come up – and

the respondent, again, is asked to make a choice. Remember, this is not a product concept

test – so it may not even be necessary at this phase to show the actual product you’re

looking to forecast. This may come as a shock to some. But more on it later…

The only purpose of the choice experiment was to use these hundreds and often thousands

of choices respondents made – among seemingly random product configurations – and be able

to calculate the “value” or utility of every one of the attribute levels. This is

where the statistical model gets calculated – a multinomial logistic regression model

often using a hierarchical Bayesian technique. Now that we have the model, we can use these

utilities and simulate what the market looks like when our product is in the market.

In this illustration, we imagine that the market is made up of three players. Product

1, Product 2 and Product 3. Using these attribute level utilities (also known as “part worth

utilities”) we can calculate a product’s “total utility” – or total attractiveness.

By exponentiating these total utilities – and using the formula on the screen, we can then

calculate a Share of Preference for each product. What does Share of Preference mean? What it

means is: in our choice experiment, if we had shown these three products to the respondents

– 87 out of 100 people would have chosen Product 1; 1% would have chosen Product 2

and 12% would have picked Product 3. If we assume that our sample of respondents

was representative, and we assume that all three of these products are available in the

market, and everyone is aware of all these products – then the share of Preference

should be very close to a Market Share. Therefore, in our first phase – where we

only want to find out whether or not there is a market for your product – a share of

preference can tell you: in a perfect word, with perfect awareness and distribution – this

is the amount of customers that would pick your product out the products in the market.

This is Phase 1. Is there or is there not a market appeal? There are quite a few steps

to accomplishing Phase 1: using the right attributes and levels as we plan the choice

task, then designing an experimental design that is balanced and will allow for a robust

model, then programming the choice task into a survey, then fielding it with a representative

sample, creating an accurate statistical model – and finally assessing the market using

a simulator. The image here shows a sample simulator to

assess market appeal for online trading services. In this illustrative example, we assumed the

market is made up of 4 key players with differing attributes in their offerings. Just a side-note

that this example does not contain real data for confidentiality purposes – so… no

need to pause the video as scribble down the most appealing levels.

So, let’s get to phase 2 now… the forecasting part. The result of phase 1 is some potential

size of customers who would pick your product over competitors. But we also had to assume

100% awareness, distribution – which are unrealistic. In the real world, awareness

and distribution take a long time to build – and we need to incorporate this into our

forecast. The question you need to ask as you head into Phase 2 is: “What is the path

to reaching the product’s potential?” In order to do that – I will now take a

theoretical detour – and discuss two common data-distributions: the S-curve and its cousin:

the Normal distribution. First, let’s talk about the S curve. Those

knowledgeable about statistics may be familiar with the S curve – especially in the context

of logistic regression. However, when it comes to forecasting that is NOT the context I would

like you to think about. The S curve, is also cumulative distribution

– of the normal distribution. For example, this curve shows how intensive the sun is

during the day – hour by hour. While this curve is the cumulative exposure to the sun

– the total sun rays I would absorb if I sat all day under that sun. Now, the true

distribution looks more like this – but the point is still the same: the cumulative

distribution of the normal curve looks like an S-curve.

The normal distribution, as a function of time tends to reflect a period of beginning

(or ramp up), a period of intense activity (the peak) and a period of decline activity.

Interestingly, the normal distribution and its cousin the S curve are all around us:

Here we see the growth of the sunflower plant – and we can clearly notice a slow ramp

up at the beginning, a period of intense growth and a slowing down after two and a half months.

Children learning vocabulary follows the cumulative normal distribution.

And artists’ productivity follows the cumulative normal distribution – here is the cumulative

works of Mozart – even though Mozart died young. The last dot, that is above the line

is the year Mozart died – already suspecting his death he worked around the clock…

So what does this detour on the normal and the S curves have anything to do with forecasting?

Everett Rogers hypothesized that new products get adopted – according to a normal distribution

that is a function of time. That the first adopters, the innovators, are promptly the

first 2.5% of adopters – who then influence the next 13.5% called early adopters – after

which an intense growth happens when the early majority adopt – which is called “closing

the chasm” or “tipping point.” Rogers’ curve has been used for forecasting

purposes – thousands of times since the theory was published; however, there are several

limitations: First, the model is based on a distribution

about the mean time of adoption. Which you really don’t know until the whole adoption

process is completed. In other words, you don’t know if you’re an early adopter

until we’ve accounted for the early majority, late majority and laggards, and only then

can we designate what phase was what time. Second, Rogers’ curve makes the assumption

that innovators can only adopt at the very beginning of the process. Empirical evidence

however shows that innovators can adopt a product at any phase of the diffusion. Albeit

a lesser degree. An objectively superior forecasting approach

was developed by Frank Bass in which he attributed product adoption to two main influences:

1. One that is due to external influences – such as advertising or believing in the

cause of the product. Obviously a larger impact in adoption at the beginning – they tend

to diminish with time. 2. And one that is due to internal influences

– such as word of mouth, peer pressure, etc…

Looking at the cumulative adoption curve, the equation to estimate cumulative sales

has three unknowns: - The size of the total market or m

- p – the coefficient of innovation – how big is the effect of external inflencers

- q – the coefficient of imitation – how big is the effect of internal influencer?

Well, we know from our phase 1 activities what the market potential for our product

is. In fact, that is the whole reason we conducted phase 1. But what numbers do we use to find

p and q? Here, I wish there was a handy technique that

would calculate the most appropriate p and q. Instead, what we have are standards and

guidelines. A typical p value is around 0.03 – rarely

if ever above 0.04 and often less than 0.01. A typical q value however is around 0.38 – with

a range between 0.3 and 0.5. These are based on several examples from different industries.

Finding the right figures does make a difference. The blue curve shows the uptake using the

typical q of 0.38 and typical p of 0.03. However, changing the q to the upper limit of 0.5 will

create the orange line which will achieve a certain sales level many years quicker.

Notice however, that swings in the p value make less of a difference in the accuracy

of the curve. So when it comes to using the p and q values

– here are some suggestions: - take a look at historical product launches

within your organization – and try to fit a curve to them, adjusting the levels of p

and q. - if there isn’t much history – start

out with typical values, - and, it is always smart – and helps save

on heartburn medication – if you create conservative and liberal forecasts.

- And finally, keep track of sales – and use the equation to continue to adjust your

p’s and q’s. What you’ll find is eventually locking in on figures you feel good about.

Hopefully with all this information you now feel more confident in helping your organization

approach new product launches in a scientifically sound and robust way. First quantifying market

appeal for your product and then using the Bass formula to quantify the path to reaching

the market potential. I hope this was helpful and I hope you’ll

reach out to me if you have any further questions.