You’re a Bayesian. You do it everyday. It’s a highly effective strategy for managing your life. It’s what these days is probably called a “life hack”, except it’s really more significant than how to get ketchup off a tie. I’ll endeavour to explain what it means and why we would want to apply it to marketing analysis, of all things
That just about sums it up neatly. Should I take an umbrella, should I bet on United beating City, will I make it to the next service station without running out of petrol? What connects all these situations is that we are instinctively blending current information with our prior knowledge. The forecast said it would rain but it looks sunny outside, although it IS February in the UK. City are doing way better than United and are 2/1 on, but they never do that well away from home against United. Yes, yes, I KNOW it’s on red, but they always set it like that, you never run out when it says empty, you’ve always got a few extra miles, honestly.
Bayesian in the workplace
As humans we process information dynamically. We sift it, give a weight to its reliability and blend it with everything else we know to come to a decision. We do that in our lives, and we do it at work too. A small survey conducted by a company you have never heard of says the total opposite to what accumulated evidence over many years has told you. Do you discard all of that because this is the LATEST data? No, of course not. You might raise an eyebrow and lodge it somewhere in the recesses of your brain. I mean they MIGHT be onto something, but am I going to the MD to suggest a change in strategy? Not really. Now had that study been amongst 5,000 people and conducted by the most brilliant minds in the field you might put your coffee down and utter a gentle “hmmm”. Still not ready to call the MD but I’m getting the team to look into it for sure.
But surely Econometrics doesn’t work like that? It’s about certainty, data, numbers.
The reality of any marketing analysis is that the data you have is almost always imperfect. This can happen in many ways. It can be incomplete, it can be sample based, it can be small and therefore erratic. Worse than that you may (shock horror) not have considered econometric modelling first and foremost when designing the timing of your marketing campaigns. You might have done them all at once which may make the Marketing Director go “ahhh” but makes the Marketing Analyst go “argh”. Doing things all together makes it very hard to strip out their individual effects in a model. Consider the situation where you did a price reduction at the same time as a marketing campaign. Crazy I know, but bear with me.
If we put all of that into a model it will probably say that all the effect was the marketing campaign and none of it was the price cut, or vice versa. It’s pretty arbitrary and will depend on the exact phasing. We may end up with a model that says the marketing campaign was (unbelievably) good and the millennia-old adage that dropping your price usually results in higher sales didn’t apply here.
So, we either (deep breath) present what the model says and show some nice t-stats and p-values in our defence or we shrug our shoulders and say, dunno. Tricky. Can’t really say, sorry. Both of which I can confidently say from experience go down like a lead balloon in the panelled boardroom and someone says after you have left “Why DO we have that guy in to present every month?”
Bayesian to the rescue (cue sound of hooves)
What we could do is to say: You have posed a tricky problem here. That being said, I know from previous price reductions that the pricing effect is likely to have been X and I’m 80% confident either way on that. Might be a bit higher, or a bit lower but it’s about that. I’ll set a Bayesian Prior* (*see our lovely new e-book for more explanation) on price to constrain the effect and see what that tells us about the marketing campaign. Oh look, it’s come out quite sensible. Or it’s rubbish or it’s STILL amazingly good. Regardless, we have learned something Quite Interesting, as Stephen Fry would say.
That all sounds wonderful, we should do that. A lot.
Well, probably not. It is not often the case that it’s needed. If we have good data, we’d rather have a robust regular model if possible. It’s also more complex to do, so needs more time, thought and experience. The good thing is that we have invested a lot of time and effort into this area because we have for many years believed in the benefits of it. That’s what prompted us to write our second e-book on this topic. If you’re interested in learning more, please do download our ‘Guide to Bayesian Econometrics for the modern marketer’ and let us know if you have any questions or comments.
Tom Lloyd, Expert Econometrician and Fellow Bayesian