Archive

July 23, 2020

Regular readers will know I’m a big fan of the “square root rule” which is a quick rule of thumb to calculate the diminishing returns from increasing media spend. The square root rule says
$$ Y \propto \sqrt{X} $$
where Y is the output (e.g. conversions) and X is the input (e.g. media spend). This tells you that, as a rough rule of thumb, doubling your media spend will increase conversions by a factor of
$$\sqrt{2} \approx 1.4$$
or 40%. The rule is never perfectly accurate but it is a pretty good approximation when you don’t know very much about a channel. I first heard about it from Kevin Hillstrom and the PPC nerds at the Rimm Kaufmann group (RIP to what was once a great blog - I can’t find the old posts since they got taken over). It is also a special case of a Cobb Douglas production function because it uses a power law to model diminishing returns. In this post I want to expand on other aspects of Cobb-Douglas with a potential application to media/agency life. The Cobb Douglas model is

May 11, 2020

A lot of my headspace has been taken up with coronavirus related thinking of late. I’m sure it is the same for many of you too! At the end of March I was worried about the number of cases overwhelming the ICU capacity; it seems we have done social distancing well enough to avoid this although I don’t want to call it a success when the number of deaths in the UK is so much higher than other nearby countries. These days I worry about the following:

    Does our higher death rate now mean we are less likely to have a bad second wave? I think it does although “less likely” is doing a lot of work in that question; the precise amount of likeliness is what’s important. How long will the post lockdown recession last? There are a lot of people on full pay who haven’t been spending much money, but I don’t think this is enough to balance out those who have been furloughed or lost their jobs. When can I start doing sport again? I know I’m allowed to exercise (and I’m very grateful for this - and Italian style lockdown would be much harder

April 29, 2020

This post looks at a few different things you might see when using the Covid Forecasting Tool and explains how to interpret them. I’ve been getting a few questions about output from the tool and I want to consolidate and expand on my answers here. Example 1: When it mostly works first time First let’s look at a nice example where everything works and things are easy to interpret. Here is a chart of the raw data:

Data since 1st October 2019
First we look at the impact plot which uses the data up to the date of lockdown to make a forecast. If the actual data since lockdown is very different to the forecast then it is reasonable to conclude that lockdown has had an impact.
Impact Plot
You can see that what actually happened (the black line) is very different from what was predicted to happen if lockdown did not occur (the blue line). It is far outside the green zone of “stuff we could reasonably expect to happen”; there is less than a 5% chance of seeing what you actually saw if the forecasting model is correct. Indeed, Google’s CausalImpact method says there

April 24, 2020

E-Analytica Forecasting launched publicly on Monday. It has been a moderate success, at least compared to most of my efforts, with over 150 analyses being conducted since than (not all of them by me!). Initial Motivation When doing a forecasting project the value I add comes from customising machine learning models to take into account specifics of the client’s business. But to quantify how much value I’m adding it is necessary to have a baseline to compare with. Often I would pick a very simple model and run it over the training data to get baseline performance. But there are more complicated models that are just as easy to run as the simple ones so I worried that I was picking a baseline that was, in some way, too easy for me to beat. I wanted a way of running lots of forecasting models at once and to be able to compare the output in a standard way. This is quite easy for forecasting methods that are part of the forecast package but once you add things like bsts and prophet into the mix this gets more complicated. I had some functions written to help me do this when (as

April 6, 2020

March, particularly early March, feels like a different time. Anyway, here are some links I saved (only one is coronavirus related). The Mandalorian’s Guide to Search On and off, I have worked with Andrew McGarry for over five years now so I’ve seen a bit of what is going on and what he is doing with The McGarry Agency. In this piece for Search Engine Land Andrew talks about some of the personal challenges he’s faced in growing his company; particularly how the traits and aspects of his personality the he thought were strengths were actually holding him back. I came to realize in my own career journey that the success we can achieve in our careers is influenced by our chosen identity and the work we do on a daily basis Andrew and I do not have the same personality, but we do have some things in common so I wonder how many of these lessons apply to me and how important learning them will be for my continued growth over the next few years. Why I regret inventing the innocent smoothie brand The faux folksy “brand voice” of innocent [it doesn’t have a capital “i” and it is