How to create accurate sales forecasts.
You will learn here what options are available for making sales forecasts – with good results.
If you work in (or with) sales at a B2B company, you have almost certainly been asked your opinion on future sales. Something like, “How much do you think you’ll sell in the next quarter?”
Does that make sense? Do many salespeople’s rough estimates produce good sales forecasts? Be curious.
There are many ways to make forecasts: from gut-feel guided estimates to simple to very complex models of math and statistics. Some are just better than others.
Before we get into the various forecasting options, let us first address the following question:
Sales forecasts: what are we talking about?
What did you think of first when you read the word “sales forecasts”? Our guess is revenue forecasts. And that is correct. However, revenue predictions are only one of several types of forecasts that are useful to sales.
Think about it: what predictions about your B2B customers could make your job even easier?
What type of coffee your negotiation partner is most likely to drink? Entertaining it would be – but all kidding aside.
Predictions like “which price is your customer most likely to accept” (dynamic pricing) or “which customer is threatening to leave” (churn predictions), save you from nasty surprises and give you time to act.
Amazon has made another sales forecast type a celebrity: the “cross-selling forecast”.
We are sure you are familiar with the following phrases: “other customers also bought…” or “you might be interested in this as well…”.
The goal of Cross-selling is that a customer – in addition to the original product – buys another one as well. This feature accounts for over 25% of Amazon’s sales.
So, you see that turnover forecasts are not the only valuable predictions that can help sales.
How forecasts arise – the “human” option.
You would be amazed how many German B2B companies predict their turnover via sales employee surveys and through a good dose of gut feeling.
And even more astonishing is that these “pinch-a-thumb” predictions are often not that bad. But when does it get tricky?
When you have many hundreds of customers & products and therefore many transactions. The fewer customers and products you have, the easier it is to keep track. But with hundreds and thousands of them – mission impossible.
Another sticking point is detailed forecasts. What do we mean by that? A turnover forecast is a number that includes total sales over a given period. But what if it goes deeper? Future revenue per customer or product. In our experience, “human” predictions at this detailed level are no better than coincidence.
How forecasts arise – the “statistical models”.
To explain the principle behind these models, let’s take one of the simplest to help: a linear function.
As a quick example: a customer has been buying from you for €100 every month for a year.
I guess you can now predict quite well how much revenue you will make with this customer in the next three months. No rocket science.
What was needed to make this prediction?
Historical data: the customer has always purchased 100 € in the past months.
Pattern recognition: you assume that the customer will maintain the “pattern” (every month 100 €).
Extrapolation: based on this pattern, you calculate (100 € x 3) how much revenue the customer will bring in 3 months.
Unfortunately, not much follows this straightforward pattern. Human behaviour especially does not. And yes, purchases made in B2B also count as human behaviour.
How can “math” predict human behaviour?
By patterns in the past. Not as simple patterns as a linear function – but they are there, and the principle is similar (historical data, pattern recognition and extrapolation).
Because purchases do not happen randomly.
If you work in a wholesale business or with an industrial manufacturer, you know that your customers already have certain regularities and preferences in their purchases.
To name just a few:
how often they buy,
how much they buy,
how much they have paid in the past,
at what intervals they buy.
And much more.
These are all criteria from which a forecast can be “built” for each customer.
Which statistical model to use depends on the goal of the prediction. So, what do I want to know?
Whether the customer is at risk of churn? Here, you should calculate the probability of churn. That can be done, for example, using a probability tree.
What price could he accept? Here, a price corridor per customer and product would have to be created.
As you can see, these projections are very time-consuming and no longer feasible manually for hundreds of customers and products.
So how does it work?
The use of artificial intelligence has increased the quality of predictions tremendously in recent years.
The reason for this is that the algorithms behind it can include very many characteristics in their projections. They are masters at recognizing patterns in large data sets (machine learning).
That is also how the word “predictive analytics” (or specifically related to sales: “predictive sales analytics”) has become established.
There are software programs that are fully specialized in making the most accurate sales predictions possible.
So, our tip is, don’t bother with manual Excel spreadsheets. Also, the time of professionals in the field of data analysis is valuable and limited. Automate sales forecasting through software. This way, you and your sales team can entirely focus on your customers.
The minimum requirement for sales forecasts.
One thing must be made clear. Forecasts are never 100% accurate. We cannot predict the future with pinpoint accuracy because none of us knows it or has a magic ball.
The goal of forecasts and all the models behind them is to get as close as possible to future reality.
To explain this, let us retake Amazon.
You remember the product suggestions “what else might interest you…”? Below that, Amazon suggests several items to you. And any product you do NOT click on would be a false prediction.
Quite a lot of wrong predictions, isn’t it? If you only take a few individual forecasts into account, yes. BUT across the masses, this Amazon suggestion algorithm is so good that it brings in ¼ of the total sales!
So, what is the minimum requirement for a predictive model? It must be at least better than coincidence.
Amazon has about 500 million products for consumers. So, the minimum requirement for the Amazon algorithm is, that these suggestions are better accepted than completely randomly offered products.
And this algorithm achieves this with bang and trumpet.