When you deconstruct and simplify your complex data patterns, you can better understand the data and build better social media marketing programs. What are you waiting for?
With the holidays approaching, I’m hearing the familiar descriptors in campaign and interaction data charts: no doubt you’re hearing them, too. "This time of year, we typically see a dip (or rise — I hear both) in traffic as business slows (or ramps up.)" I started wondering: How much of this is accepted as belief, and how much is based on data? And more importantly, how can data deconstruction and modelling be used to extract really useful information, supporting better-performing social campaigns and programs?
A few years back (actually, quite a bit more than a few) I worked for NASA|JPL on the Voyager spacecraft. In order to prepare for planetary fly-bys, we’d simulate maneuvers, use of cameras, power up the various detectors, etc. As we did this, we captured data: my team was responsible for understanding how ambient temperature affected the spacecraft’s radio receivers. As we turned instruments on and off, and as we rolled the spacecraft’s receivers toward and away from the sun, we’d record the temperature of the receivers, gaining information that was critical to keeping them locked on the Earth.
So you’re wondering: what does this have to do with understanding the performance of my marketing campaigns on the social Web? A lot, actually. Like the temperature of the receiver, your campaigns are impacted by some things you do directly, and by some other things that occur and thereby impact your programs indirectly: external events, competitive announcements, etc. I wrote an article a while back about the importance of launching campaigns on specific days of the week on this same topic.
In the case of a metrics following marketing campaigns — tallying social data around the interaction of customers with each other in a discussion forum — there are of course direct drivers of behavior: if you change the offer, if you alter the forum navigation or change creative elements, then you’re likely to see a change in traffic, conversion, bounce…the set of metrics that depend most closely on the parameters or attributes you just changed. No news there.
But you’ll also see other changes: the effects of holidays, or an event in the news, a product recall, etc. It impacts your results, so you have to explain it. Even better — after all, if it happened once it will probably happen again — you can try to understand these factors and build them into your planning model. How? You can use the use the same techniques I used at JPL to pick apart the various contributions of each of these factors. The benefit? Rather than assuming, you’ll know if what you are seeing is because of the holidays or instead an indication of an underlying cause that you really need to pay attention to.
Here’s how it works: Take a look at the charts in Figure 1: on the left is a typical representation of weekly social interaction in a community: weekends are the low points, with visitors more active during the week. To the right are two different "trend" interpretations of the pattern in the data.
In the center chart you see a typical linear trend line: Excel will do this easily. The point here is that by choosing the start/end points it would be just as easy to redraw the trend line sloping downwards or flat as is was to draw it upwards. Now take a look at the chart on the far right. Instead of a liner trend, the chart is split into two parts, with a flat trend line fitted to the each of these. The insight is this: if you strongly believe there are no influences on your results other than the changes you are making, then the data likely falls into a single set and the center chart (the slope of the line notwithstanding) is the right analysis. But that’s not always the case, and splitting data set may help you spot external influences.
If there are also external events occurring, splitting your data may be better approach: deconstruct your data into sets of components and see. Artist Piet Mondrian made creative use of deconstruction: start with the complexity of real life and reduce it — deconstruct it — to more clearly show the fundamental elements that collectively contribute to the complexity observed. Check out the progression from left-to-right, as the artist’s early work slowly gave way to a radically deconstructed view of the world.
To deconstruct social marketing data, start by separating local periodic cycles (weekly traffic patterns, for examples) from the larger global (external) influences. Create a model for the local cycle — in this case the weekly up and down that repeats regularly and lay it over the observed data. In the chart shown on the left in Figure 2 I’ve created a curve (using the polynomial trend line option) that fits the weekly data. On the right, the cycle has been plotted (red) over the original data using the baselines from Figure 1, far-right chart. The fit looks good, especially when compared with the single linear trend.
A couple of notes: nothing about this analysis says it any more "correct" than a single trend (or any other tracking model). Rather, the idea is to look at your data in more than one way, and see if the alternative views produce insights that lead to a better understanding of what is actually happening. Ultimately, the decision is yours: it’s your "marketing gut" that has to check the validity of the results.
Take a look at the charts in Figure 3: I’ve removed the original data, leaving only the model (red, chart on the right) representing the curves that were created based on the model of the local cycle and suspected discontinuity (the separation from on data set to the next). On the right, the model itself is de-emphasized (now shown in gray) to highlight the chart feature of interest: the discontinuity in the data between the two baselines. The task now is to explain this, and if it’s material to predict its impact on future campaigns.
Why go to all this trouble? Rather than trying to make sense of a graph like the original chart containing all of the complexity of the real world, it’s often easier to look for a simplifying model and then deduce why the specific characteristics of that simplified model do (or do not!) apply to your actual results. In the case described in this article, a half-year of cyclical data has been reduced to a single repeating curve and one discontinuity — a step change — in the data. This helps to zero in on the data itself and ask the important question: "What happened right here?" If an answer is found, that provides a big clue into understanding external events that may impact future campaigns. This is information that can be used to make better (in the quantitative sense) statements about how campaigns actually perform, and therefore to make better decisions on how and when to run them.