Digital Marketing Is All You Need
How well-targeted advertising can make your profits skyrocket
The main flaw of the baby entrepreneur, especially if he or she comes from a technical or technological background and is full of ideas but with little in-depth knowledge of business management, is to underestimate the phenomenal impact marketing can have on a very young company.
Coming myself from a background where technical mastery and innovation are essential, I’ve always been bathed in the idea that, as long as your project makes sense, is creative and groundbreaking, then customers will come on their own, and that it’s better to focus on developing product performance rather than setting up a sound and efficient marketing policy.
So what made me change my mind in such a radical way ?
I recently came across Silvr, a young French start-up focusing on providing funding to e-businesses for their marketing campaigns. I had to analyze, for some of their clients, the impact an investment in ad campaigns could have on their business. And the result was just mind-blowing.
Purpose of this post
For the sake of this demonstration, I will proceed step-by-step with analyzing the impact marketing funding had on weekly revenue for an example company, let’s call it XX. This will be an in-depth econometrics analysis of financial and accounting data, accompanied by Python code snippets and results if you want to reproduce the study on your own startup business.
How does it work ?
The idea is quite simple : you want to invest a certain amount in a marketing campaign on Google and/or Facebook, but don’t have the required funding.
Silvr can provide for the funds on-the-go (just what you need and when you need it), and your business will pay back the loan progressively, without endangering its solvency with tight deadlines : you just refund a given percentage of your weekly business revenues, let’s say 20%, until the original amount augmented by the commission is totally paid back.
The great thing with it is that you don’t start paying back immediately, but only after an offset of around 9 weeks, which leaves you plenty of time to enjoy the first rewards of the marketing campaign.
Why is this advertising so insanely effective ?
Google and Facebook, as well as all other social media platforms (e. g. Instagram, Twitter…), have totally disrupted the world of advertising. The era of TV ads targeted on a fraction only of the listeners is far behind us. Now is the time of the ultra-profiling of users, the more precise you get about his or her interests, the best chance you have that your ad will catch his/her eye, and in the end to get a purchase. Social Media are providing businesses with a wide-open door on the world of their potential customers.
Supporting data & methodology
XX is a young business specialized in online sales. I have a series of accounting data, before and after they have started funding advertising campaigns.
The case of XX is particularly interesting because we have data on weekly business revenues :
- before starting digital marketing (until the 12th of April),
- during the first advertising campaign (from 13th April to 7th June),
- during a non-advertising period afterward (8th June to 2nd August) where we will seek the longlasting revenue effects due to the newly-gained reputation induced by the first advertising campaign,
- and during a second advertising period from 3rd August to 20th September.
Econometrical Analysis
Descriptive Statistics
Let’s first take a look at the raw data, to get a global impression.
It seems obvious, at first sight, that the first digital marketing campaign has had a huge impact on weekly revenues, and very quickly. It also seems to have had long-lasting effects, since revenues were plateauing higher after the end of the first campaign than they were before. Let’s take a deeper look.
Weekly revenue and weekly online advertisement spendings are highly correlated (72%). Therefore, if we want to model weekly revenues, the variable marketing expenses will have to play a central part.
It seems obvious, looking at this histogram, that the weekly revenues before starting any digital marketing campaign were significantly lower than they would ever be at any point in the future.
It is also striking that the first marketing campaign made revenues literally skyrocket : in 3 weeks, they were multiplied by more than 6 ! However, whether the second advertising campaign has had an effect significantly different from the long-lasting effects of the first one is harder to tell.
Statistical Modeling
In this part, based on the exploratory data analysis I have conducted so far, I will make several assumptions in order to build a statistical model.
- The first assumption is that the introduction of online advertising has sufficiently disrupted the behavior of the revenue series that we can model the 2 periods, before and after, more or less separately.
- The second is proper to part 1 : I will assume that, before the launching of social media marketing, the revenue was a gaussian process of given mean (1471€ to be precise), without any growth trend, which seems coherent with the graphs where the revenue appears to have been stationary and quite stable in time.
- The third is that, in the second part, the revenue must be somehow proportional to advertisement spending, what seems intuitive, but also somehow inversely proportional to the time elapsed since the initial launching of digital marketing. The argument for this is two-sided : on the one hand, it seems noticeable on both graphs that the introduction of marketing has profoundly changed the behavior of the revenue process (see assumption 1), but also that this social media campaign has had a long-lasting effect even after the end of the first campaign, raising consumer awareness on the brand, whereas the second marketing campaign didn’t appear to have the same exploding effect.
Therefore, putting things in a nutshell, I will assume that revenue at time t (t will be a week number since we are working with weekly revenues), called R, is a function of :
- time itself, in so far as I will allow my regression to have different constant terms for each of the two periods before and after beginning advertising, to account for the durable disruptive effect of brand-awareness ;
- the advertisement spending at time t, called S,
- the inverse square root of the time-lapse since the beginning of online advertising, in order to take into account lasting though declining long-term effects of digital marketing campaigns.
- To this, I have added a Gaussian noise of mean 0, which is accountable for the stochastic variations around the trend, observed on a weekly basis.
Let’s first analyze the meaning of this hypothesis : you can notice that only the first term is non-null before the launching of the advertisement campaign. It encompasses the initial behavior of revenue, a stationary Gaussian process whose mean is simply the average weekly revenue in this time period.
After the beginning of digital marketing, the three other terms embody respectively a constant (the mean revenue that XX will reach in an infinite period of time if advertisement spending stops), the effect of ads expenses on revenue, and of the time-lapse since the first ads were displayed.
Explaining the data
The next step is to perform a linear regression on these 4 features (2 variables and 2 constants), in order to check the viability of our model (significance of coefficients, explainability, and predictability power).
I will perform the regression using Python’s scikit-learn tool, overwriting it so that it also calculates some statistical entities (p-values, t-statistics) that are very useful in probing the model’s validity.
The R2 of this regression, that is to say, the part of the total variation in revenue that is explained by my features (2 variables and 2 constants), is of more than three quarters (77%), which is quite a good result. We have to take a look at the coefficients and their respective p-values, in order to be sure that they all significantly contribute to the model :
Reassuringly enough, all p-values are very low, thus for all coefficients we can reject the null hypothesis that they are not significantly different from 0.
Our model seems quite robust !
Predicting the data
If we use our regression model to predict revenue, reassuringly enough, we find a trend that seems highly similar to the actually observed revenue. To confirm this view, we can just take a quick look at the correlation coefficient :
But what use is it really ?
Now, the question you’re all asking : is it really worth it ? I will make a simple test : looking at the actual revenue after the beginning of social media advertisement and trying to predict the revenue on the same period had no change occurred in the initial trend, I will check whether the investment (total ads spendings, augmented by the Silvr commission of 6%) pays back.
To detail the procedure, you remember that, when explaining the econometrical model, we saw that the model on the initial period was equivalent to a simple Gaussian process of mean 1471€, completely stationary. Propagating this modelization to the second period is therefore quite easy, the forecasted revenue is simply equal to the constant multiplied by the number of weeks we englobe.
Analyzing the return on investment simply requires to subtract this forecasted revenue from the actual revenue in the second period, to get the net revenue increase obtained through the digital marketing operations, and to further subtract the overall cost of the campaign (initial cost augmented by the commission), in order to end with the final net benefit on the period.
Wow, impressive ! In little more than 5 months, the average revenue has been multiplied by nearly 6 compared to what it was before the business launched its digital marketing campaign…
