How I Forecast Performance for Six-Figure Marketing Campaigns
Forecasting marketing performance is always difficult, particularly in the noisy world of online marketing, where a million different variables can affect your campaigns — how much you bid, what your adcopy looks like, whether it’s raining… how are you supposed to know what affect all of these things will have on performance?
The truth is that it’s impossible to predict future performance — but try telling that to to your CEO when they ask you to justify a budget increase or predict how much you’ll spend by the end of the quarter, and at what return.
Having to send the CEO something, but not knowing enough statistics to work out anything advanced, most people turn to the simplest method: extrapolation — which works as follows:
The Extrapolation Method
This method is simple: find out your average daily spend, your average daily conversions, then multiply them by how many days left in the month and you’ll get an estimate of how much you’re due to spend, and how many conversions that will buy you. You also know your CPA (cost per acquisition/conversion) so if you know how much a conversion is worth to you, you can work out return on investment for your campaign.
Seem too simple? It is. Take a further look at the data:
The days highlighted in blue had a really high cost per conversion, which was typically sitting around the 3 cents mark— is the 13 cents CPA really representative of the performance you’ll see by month end? Additionally spend was very low all the way up until row 16, when it looks like we drastically increased our run rate, and therefore saw worse performance — how do we take that dynamic into account? This data looks so random… is our CPA even predictable?
The Efficient Frontier Method
The good news is that there is a better way to predict future campaign performance. I call it the “Efficient Frontier” method after the incredibly smart company I learned it from (who later became part of Adobe). They borrowed the term from the world of finance where it is used to predict the value of a portfolio of investments. Though the way they implement this method is more accurate (and expensive!), involving advanced algorithms, we can actually make our own basic version (for free!) in Google Spreadsheets or Excel. It’s as simple as taking our data and running a scatter plot diagram, to get the following:
As you can see, we’ve plotted the Conversions on our Y axis, and Spend on the X axis to get an interesting pattern… but what causes this shape?
Diminishing Marginal Returns
Any economist will recognise this shape as a representation of ‘diminishing marginal returns’ — what this means is that for every additional dollar (on the margin) we spend, we’re getting a decreasing (or diminishing) number of conversions in return. So to take the above example, when we spend around $400, we’re getting 4,000 conversions, but when we double that spend to $800, we’re only getting 5,000 conversions — 1,000 more than when we were spending $400. To think of this another way, we paid $400 for those additional 1,000 conversions, or 40 cents per conversion. This is an incredibly important concept in marketing because if you weren’t aware of this (and many aren’t), you might make the mistake of thinking you’d be able to double your conversions when you double your spend.
I see this shape in almost every marketing campaign I run — it’s pretty much a law of nature. It occurs in campaigns that are well-optimized for a simple reason — you spend on the best performing placements first, in the least competitive areas (sometimes called the low hanging fruit). As the volume of the campaign expands you start getting into less profitable and more competitive territory, so your performance (CPA) starts to suffer.
Note: if this pattern doesn’t emerge, either you need more data, your campaign might not be fully optimized, or your campaign is being affected by something big, such as a sudden increase in inventory, change in ad format, or general improvement in account structure.
Tweet me @2michaeltaylor if you want help interpreting your data.
How to use the Efficient Frontier method
Let’s take another look at what that chart tells us.
We can pick almost any spend point on this chart and guess at roughly how many clicks we’ll drive, and therefore our CPA. For example if our CEO asked us to spend $1,600 per day on this campaign, and expected us to drive much more than 6,000 conversions we can show them this probably won’t be possible. We can now run a much more plausible forecast to see what our CPA will be by the end of the month, by figuring out how much we want to spend per day, finding the point on the graph and extrapolating how many conversions we’ll get. Or we could set an acceptable CPA, then forecast how much we need to spend per day to get to that level. Day-to-day things will always fluctuate, but this is our way to use data to make an educated guess.
Top tip — plot CPC vs Spend to forecast how much you should be bidding to hit your desired daily spend level.
Example
Say we only wanted to spend around $100 a day — we can see that we’ll pretty reliably get close to 3,000 conversions per day. In fact, this chart indicates that even if we paused this activity we’d continue to get about 2,000 conversions a day — an indication of cannibalisation, which means that we’re attributing conversions to our marketing activity that we would have gotten anyway — you might want to down-weight your clicks by this amount in order to get a better idea of real marketing performance.
The Magic of Aggregate Data
But how can I predict performance on a specific day? This chart won’t help me if its raining and I sell ice cream, if my site goes down, or if my intern screws up and hurts performance by 20%… and you’re right — this method can’t possibly account for all of the possible things that could affect performance on any given day, just as I can’t possibly predict if you will get hit by a bus tomorrow. What I can actually tell you with more accuracy is how many people will get hit by busses tomorrow… and that’s the magic of aggregate data. What is impossible to predict with a sample size of one, is more accurately predictable once you have enough data. Though the Efficient Frontier method will be completely wrong with its prediction on certain days, over the course of a month it should actually be pretty spot on.
That’s all folks!
So that’s it — a simple scatterplot diagram can give you a decent idea of what to expect and you don’t have to learn any advanced statistics. Advanced technologies like Adobe Media Optimizer can do this at scale, modelling this dynamic for every keyword and placement, and adding a bunch of other advanced maths into the mix — so if you can afford a solution like that I’d say go for it. If you’re a small startup, with less than a six-figure marketing budget, you can’t go too far wrong with this method.
Feel free to tweet me if you have any questions, suggestions or if you need a bit of help interpreting your data, always happy to help! @2michaeltaylor
