Facebook Marketing Simulation Models
For an online retailer whose only marketing spends are on Facebook, do the following-
- Figure out whether there is a correlation between Facebook Fan activity and sales trends
- If yes, build a statistical model that links Facebook Fan activity to sales trends so that given sales targets, they know what should be the Fan growth trajectory and align their marketing dollars accordingly.
Background and Challenges
Our client, a rapidly growing online apparel retailer from UK has a single marketing Channel – Facebook. Riding on the popularity of a couple of path breaking products, they have seen an unprecedented sales growth in last couple of years. To leverage this steep growth in sales, and scale up the business in an efficient manner, it is important that they can forecast their future sales effectively.
ScoreData Corporation helped them on this using our state of the art time series models as part of an earlier project. Now, since the only marketing dollars they spend are on Facebook, they wanted to find out whether there was any statistical correlation between Facebook Fan activity and their sales. If yes, they were interested to find out the statistical relationship between the two so that given the sales targets, they know what kind of Facebook Fan growth they should aim for.
End result of any analytics assignment is as good as the quality of data that goes into it as input. So as the first step, we asked for relevant datasets for this analysis keeping in mind the business objectives and constraints. We were provided sales data, website conversion rate data and Facebook Fan data for last 12 months time, at daily level.
It was now time for us to roll our sleeves up and deep dive into the data. To start with, a thorough data audit report was developed and shared with the client. This served two purposes:
- establish data sanity, i.e. to make sure we are working on correct datasets
- understand various trends in the data and plan the next steps accordingly (please refer to illustration 1) .
As a part of the data audit, all kinds of univariate and bivariate trends were studied. And as a result of this exercise, a hazy but distinct picture started to emerge about the nature of relationships between various business dimensions.
It also emerged that the spikes in sales and Fan growth were actually seasonal events and the trends were completely different in these extraordinary time periods. Since these occasions accounted for a substantial part of the annual sales, it was recommended that they be studied separately.
Now, the first question to answer was “Is there a strong correlation between Fans (cumulative/ incremental) and sales?” And if there is, the next question would be “sales of a particular period are driven by Fans of exactly which period?”.
To answer these, a correlation study was conducted between sales and various Facebook Fans related dimensions and it was found that for this apparel brand, sales of any given week are highly correlated with Fans that came aboard in the previous three weeks (illustrations 2 and 3).
It was concluded from this analysis that there indeed was a strong correlation between sales and Fan growth. Moreover Fan growth in a week drives sales for three succeeding weeks after which the effect tapers off gradually.
Same set of analyses were run on the “occasion” data at distribution center (DCs) level for last couple of years and results were shared with the client.
Now that the nature of the relationship between sales and Fan growth had been established, it was time to get under the skin of mathematics in that relationship. This would enable the client to successfully forecast one of them, fixing target on the other.
From the historical data review it was clear that the correlation between the two was positive but not linear. Additionally, the relationship between various dimensions changed from DC to DC. As an outcome of this, multiple non linear models were developed between sales; various Fan parameters and other factors that played a role (like website conversion rate) at DC level. Then based on model statistics, models where predicted numbers followed the actual numbers closest in the validation datasets were picked and shared with the client (for an “actual” to “predicted” comparison,please refer to illustration 4)
Since the behavioral patterns changed drastically around the ” occasion ” periods, separate non linear models were developed for these time periods.
Results and Implementation
The model equations were packaged into spreadsheet based easy to use simulation models. With the help of these models, the client was able to set challenging but realistically achievable sales targets. Once targets fixed, they knew exactly how many marketing dollars to spend and in what schedule.
This helped the client run their business in a scientific and efficient way, and scale up in a sustainable manner during the times of 300% year on year revenue growth.