In part 1 of this article, I set up my reasons for investigating Coinstar (CSTR) stock and set up what I see to be the key question that will determine Coinstar returns in the coming quarters: whether Netflix's (NASDAQ:NFLX) lost customers are opting for Redbox. In this part, I look at what state trends in Google searches might have to say about this question. In the third installment that follows, I try to put it all together and reflect on how this analysis has informed my position with respect to Coinstar stock.
Putting all state level data together
In this portion of the analysis, I pool together all Google searches for "Netflix," "Redbox" and "IMDB" from the 50 states from 2004-present. I weigh each state by its population size in a given year to account for the fact that what is happening in California is most likely far more important for a company's revenues than what is happening in Vermont, due to the huge discrepancy in population size between the states.
I am attempting to model the relationship between Netflix searches in a given state in a given week and Redbox searches in that state and week, and, specifically, whether the declines in Netflix searches seen in every state since they announced their price increase is associated with an increase in Redbox searches in those states at that point in time. I expect that this decline will be seen.
I want a model with essentially two components. First, I want to capture the average association between Netflix and Redbox searches prior to the announcement. Second, I want the average association of Netflix and Redbox searches after the announcement. Remember, I expect a negative correlation between Redbox searches and Netflix searches after the price increase, reflecting the fact that as searches for Netflix decline searches for Redbox increase, which I would take to mean that Netflix's lost customers are going to Redbox.
Table 1 attempts to explore this expectation descriptively. I divided the state-week searches for Netflix that have occurred since 2008 into four quartiles, each reflecting 25% of the state-week search volume. (Note, I used state populations in each year as weights, so the observation numbers in the table will not divide into quartiles; the true division is based on the weighted state search volumes). This is a simple way to categorize whether search volumes were high (Quartile 4) or low (Quartile 1). I cross-tabulated this measure of Netflix search volume with the timing of price increase announcement (before and after), and calculated the mean level of Redbox searches in each cell (again, using population size weights). I have highlighted the cells of interest.
What table 1 shows is that higher search volumes for Netflix appear to be associated with higher search volumes for Redbox, both before and after the announcement. This can be seen looking across quartiles. Before the price increase, in states and weeks where Netflix was searched infrequently (quartile 1), average searches for Redbox were also low, while in states and weeks where Netflix was searched more frequently (quartile 4), Redbox was also searched more frequently. After the price increase the numbers are a bit messier, but, ignoring quartile 1, the same association appears to hold: More Netflix searches correspond to more Redbox searches. Because the values in quartile 1 are very high and driven by states with very small populations (e.g., Montana, Vermont, Maine, etc.) which have greater uncertainty in their estimated search volumes, I think it is reasonable to ignore the values in that quartile.
Table 1. Average Redbox searches by quartile of Netflix searches before and after price increase announcement, 2008-present.
Weighted quartile of Netflix searches | ||||||
Quartile 1 | Quartile 2 | Quartile 3 | Quartile 4 | Total | ||
Prior to announcement | ||||||
Avg. Redbox searches | 0.35 | 1.04 | 1.92 | 3.23 | 1.50 | |
Std. Dev. Redbox searches | 1.18 | 0.75 | 1.15 | 2.24 | 1.71 | |
Obs. | 4155 | 1928 | 1658 | 995 | 8736 | |
After announcement | ||||||
Avg. Redbox searches | 13.63 | 2.90 | 3.41 | 4.97 | 4.84 | |
Std. Dev. Redbox searches | 9.81 | 3.07 | 4.95 | 3.49 | 4.08 | |
Obs. | 71 | 73 | 229 | 442 | 815 | |
Total | ||||||
Avg. Redbox searches | 0.42 | 1.06 | 1.99 | 3.73 | 1.79 | |
Std. Dev. Redbox searches | 1.67 | 0.83 | 1.60 | 2.77 | 2.23 | |
Obs. | 4226 | 2001 | 1887 | 1437 | 9551 |
Note: Quartiles computed using state population weights. Quartile 1 is fewest Netflix searches, Quartile 4 is most. Using only data from 2008 forward because Redbox was not commonly searched prior to that point.
Another, more complete way of looking at the results in table 1 can be seen in figure 6. This figure shows a scatterplot of the association between searches for Netflix and searches for Redbox before (blue dots) and after (red dots) the price increase. Each observation is a state-week combination, and the size of the dots represents the respective population size of that state in that year. To fit the values nicely on the graph, I used natural logs, but the same take-away could be reached with the raw values. The take away point from figure 6 is that, both before and after Netflix announced the price increase, searches for Netflix and Redbox are positively associated. This, like the results of table 1, runs counter to my expectation and seems to provide evidence that Netflix's lost subscribers are not opting for Redbox.
Figure 6. Scatterplot of time-space searches for Redbox and Netflix before and after price increase announcement, weighted by state population sizes.
Of course, there are a number of other things that may be going on and biasing the conclusions reached by observing this simple relationship. A reasonable person might wonder whether temporal variation in the likelihood of renting movies is driving this trend; perhaps people are less likely to rent movies in July-December. (I think the opposite is probably true because the summer blockbusters come out on DVD then and people have more free time, but it's possible.) Others may be concerned that there are just some differences between states that make them more or less likely to search for Netflix and Redbox.
I'm not going to go deep into the statistical logic of all of this, but I believe that the linear regression models presented in table 2 overcome both of these problems. To compute them, I computed each state-week observations' deviation from the average value in that state over time for the variables reflecting Netflix searches, Redbox searches and IMDB searches. If there are differences between the states, running the regression on these deviations should account for much of them (for more on fixed effects models see this page). If there are differences between time periods, controlling for IMDB searches, which I assume are highly correlated with temporal fluctuations in movie rental searching, should capture that variability.
Table 2 presents the results of this regression. In columns 1 and 3, I essentially model the same two factors as I did in table 1. In columns 2 and 4 I ran the same model controlling for IMDB searches, which did not make much of a difference. (Columns 1 and 2 are the raw results as in table 1 while columns 3 and 4 are the results computed when natural logged values of Redbox and Netflix searches are used as in figure 6). My expectation that Netflix customers are leaving for Redbox would predict that the highlighted column - Netflix searches in Q3&Q4 2011 - would have a negative association. In none of the models that I ran does it. This is further support for the basic descriptive result that Netflix customers do not appear to be leaving for Redbox. If you are lost, the basic result shown in figure 6 and table 1 is upheld; the descriptive graph of what occurred in California (figure 1) does not appear to be the national norm
Table 2. Fixed effects regression results.
Raw | Natural Log | |||
(1) | (2) | (3) | (4) | |
Netflix searches prior to Q3&Q4 of 2011 | 0.597*** | 0.613*** | 1.034*** | 1.050*** |
(0.00471) | (0.00565) | (0.00776) | (0.00783) | |
Netflix searches in Q3&Q4 2011 | 0.214*** | 0.211*** | 0.121*** | 0.138*** |
(0.00662) | (0.00664) | (0.00701) | (0.00712) | |
IMDB searches | -0.0828*** | -0.128*** | ||
(0.0157) | (0.0119) | |||
Average value of Redbox searches | -0.405** | -0.346* | 3.956*** | 4.065*** |
(0.178) | (0.178) | (0.334) | (0.331) | |
Observations | 19583 | 19583 | 6010 | 6010 |
R-squared | 0.634 | 0.635 | 0.870 | 0.873 |
Note: Observations weighted by state populations in year. State fixed effects not shown. Remember that each variable represents an independent effect; in order to get the expected value, one must add the constant and the variables together. The natural log results eliminate state-week combinations where either Netflix or Redbox searches were zero because both variables are logged; this generally means that more recent observations are included.
In the next section of this article, I try to put all of these results together with data from Coinstar 10Qs and reflect on how I am positioning myself in light of this analysis.
Disclosure: I am short NFLX, CSTR.
Additional disclosure: I am short both through puts.