It's goal-setting time, and you need to be realistic. If you double your ads budget you won't double your sales, but how much of an uplift would you expect? What would that do to your ROI?

GoolyBib Office – Second to Last Week of the Quarter

You have been procrastinating on this task, but you can avoid it no longer. It's that time of the quarter where you need to submit a business case to support your ongoing marketing budget for next quarter.

Hi,
I need the budget proposal by end of week to send onto finance.
LMK if any issues.

Building a budget proposal is one of the hardest problems in marketing. Ask for too much budget without any evidence to back it up, and finance (and your boss!) will question your competency. Ask for too little, and you won't be able to hit your targets.
With the right modeling techniques you can predict what a realistic target would be based on a given budget. This gives you the upper hand in negotiating budget and goals, helping fast track your career.
However most marketers don't know how to accurately model or predict future marketing performance, so they make decisions on gut feel or simply extrapolate by X%. This can leave you in hot water when it leads to missed targets, or if you encounter a more senior marketer who knows how to forecast properly.

Budget Proposal

A common mistake is to assume your CPA (Cost per Acquisition) or ROI (Return on Investment) won't change, even as you increase budgets. In reality, most channels give diminishing returns, which means efficiency decreases at higher spend levels.

Whatever we're spending now​.​.​. let's double it.

Graphs are easier to read than numbers, so let's start with a chart. If we plot spend and sales by day, do we see any patterns in performance on high spend days versus low spend days?

Whenever you're doing data analysis, it can be helpful to visualize the data. Charts are easier for the human brain to process than spreadsheets, and can help us identify any patterns in the data.
The first thing we want to check for is the difference between high and low spend periods. If performance looks worse during higher spend periods, then we can identify a pattern of diminishing returns.
For this task, you'll get access to performance data for GoolyBib over 30 days, with spend and conversions broken out by date. Plot spend and cost per acquisition over time to identify any patterns.

Visualizing Data

What day did the daily spend break past $3,000?

What happened to the CPA when the spend increased?

Plotting spend and sales over time

We're reached the point where a little statistics is unavoidable. Is there a correlation between spend and conversions? Can we establish a linear relationship between the two?

You remember something vague about correlation from university, but it's been a while since you did any statistics. You message a friend to ask how to establish if there's a predictive relationship between spend and conversions.

Have you done a linear regression before?
It's easy to do in GSheets/Excel, it's just the LINEST function
Let me know if you run into issues

Linear Regression is a powerful statistical tool for finding the relationship between two or more variables. It works by estimating model parameters from the data, usually using a least squares approach.
In a simple linear regression, like our current situation, there's one dependent variable we want to predict (conversions) and another called the explanatory variable we think explains or predicts the first dependent variable (ad spend). The line of best fit is drawn to minimize the sum of the sum of the differences between the line and the data points we observed.
With a multiple variable regression, the same process is in effect, but just across multiple variables (it's hard to draw a graph with more than 3 dimensions, so you'll just have to trust the math!).

Linear Regression

In Excel/GSheets you can build a linear regression model using the LINEST function. This takes the known_data_y (the variable you're trying to predict), known_data_x (all the explanatory variables you're using to predict) and instructions to calculate_b (add an intercept, which is advisable) and whether you want the results to be verbose (return all of the stats for the model). You can also more easily get these answers by adding a trendline to a scatter chart.

The formula for the equation is telling us which of the following?

You need to acquire 950 conversions per day in the next quarter. What should your daily budget be?

Do spend and conversions have a relationship?

Fast forward to a few weeks in the future – turns out our prediction was wrong. How do we transform the data to account for a non-linear relationship in the data, i..e diminishing returns.

You ramped up spend according to your model... but didn't see the results you expected! As spend increased the CPA rose to intolerable levels. We didn't accurately predict the relationship between spend and conversions, and need to adjust.

Hey,
We're not hitting our daily conversion target – the model was too optimistic.
I can get us more budget from the board, but we have to be sure we'll hit that 850 per day figure.
Let me know when you figure it out.

Ah yeah
Sounds like you're hitting diminishing returns
You want to transform the data first to account for non-linearity

In the Diminishing Marginal Returns template provided, what do you need to spend to get at least 950 conversions per day?

How much did the extra conversions cost you, going from $5,000 per day to $5,250 per day?

The path to progress is rarely a straight line

You can't always predict tomorrow with yesterday's data. Luckily you can make the future happen today with a controlled experiment, and use that data for your forecasts.

After implementing a series of inspired creative tests, performance has improved. The decrease in CPA is the welcome result, however there is no longer a clear relationship between spend and conversions, so it's hard to know how to forecast...

Single variable linear regression models only work if that variable is responsible for the majority of changes. In reality performance is a factor of multiple variables – this is where multi-variable regression and marketing mix modeling usually comes in. You would include other variables in your LINEST function that could have affected conversions, for example what creative tests you were running.
There is a trick for estimating diminishing returns quickly, without having to build a complex model – scale testing. One thing most analysts don't think about, is that if you don't have the data you need to make a decision, you can create the data you want, by running an experiment! Scale testing involves simply increasing spend in intervals in a specific campaign or geo-region, and making no other changes to the accounts for 1-2 weeks. This gives you clean data on how spend affects conversions, in order to do your forecast (there will still be seasonal effects, but these should be minimal if you time the experiment well). The exercise file below shows the results of a scale test and how to interpret them.

Scale Testing

Hey there
Great job on the scale test, really clear and actionable results.
Good to see we have a framework for making these decisions going forward.

Setting realistic goals by making the future happen

Gustav Gieger

Head of Performance

Dushyant Dixit

Your CEO asks you to double marketing spend, but you have no idea what that will do to CPA... how do you forecast realistic goals?

Vexpower Learning Certificate of Completion

Congratulations German Kapnin 💪