Top Considerations for Spending on Social

In this article, Panoramic’s data science team presents a simple way to use data to better inform spend allocation in digital marketing.

The Optimization Problem 

How to divvy up spend across platforms or channels is a common problem at the start of any marketing campaign. Ideally, marketers want to maximize the impact that each platform will have toward their ultimate goal (i.e awareness, retention, etc).

Digital marketers commonly partition their budget based on gut-feeling and general platform knowledge without incorporating any concrete guideposts from historical data, such as past performance and spend. Platform features, like demographic skews (i.e. millennials on Snapchat), can often be the primary decision to spend X on Y platform. Taking another example, Facebook skews female relative to Twitter, so a marketer targeting females might allocate more spend on Facebook. Also, Facebook often gets larger spend allocation relative to other digital budgets due to its low cost per impression, large user base, and the ability to micro-target. Based on these general features of one platform, it might make quick sense to allocate more spend toward Facebook over other platforms like Twitter, but is that the best strategy? Using historical performance data integrated with digital spend data can provide help address this spend optimization problem.

Below is a breakdown of the method used to optimize a theatrical marketing campaign, using previous performance data. For simplicity, we only apply this method to two social platforms, Facebook and Twitter, but the same approach could apply to other social media platforms. 

Optimization Method

The method deployed here consists of two main components (shown below in Figure 1):

  1. Mapping social spend to social engagement (organic and paid)
  2. Mapping social engagement to opening weekend box office revenue
Figure 1. Schematic of the spend-to-engagement-to-box office estimates. 

The method described here does not account for spend on some digital platforms (e.g., YouTube and Snapchat) or traditional mediums (e.g., television, print, and radio). A detailed description of the two mappings is given below. 

Social Spend to Social Engagement

There is a correlation between social spend data and the level of engagement across all posts on both Facebook and Twitter. For the sample of movie spend data that was available to us, the correlation between Facebook spend and Facebook reactions and likes was approximately 0.49 and 0.47, respectively. Of course, this correlation varies depending on the quality of the creative used in a Facebook post (better creatives typically lead to higher engagement) and which movies are used in the calculation. For example, when calculating correlations and slopes, it might be ideal to only use horror films, only use wide release films, or only use movies from a given studio. However, using a smaller sample with less data points could make the spend-engagement relationship less robust, which is also why it is not always ideal to directly map spend to box office revenue. 

Similar results were also found for Twitter engagement. The correlation between Twitter spend and overall retweets and favorites on all film-specific Twitter accounts was 0.49 and 0.46, respectively. We use these correlations to calculate the univariate linear regression slopes between social spend and each social engagement metric. In other words, we compute the regression slope m:

ys = mx x

In the equation above, ys is the spend for a given social platform (Facebook or Twitter in this case) and x is the total accumulated engagement metric (e.g., likes or retweets). Based on the values of mx, we can estimate how much social engagement a marketing campaign will generate for both Facebook and Twitter.  

Social Engagement to Box Office Revenue

The final step is to use estimates of social engagement to predict opening weekend box revenue. This particular method is useful when a movie studio has a certain box office goal in mind prior to allocating any marketing spend on social media. After gathering social media data for over 350 films in the past four years, we found a strong relationship (R > 0.8 in some cases) between the major social engagement metrics and opening weekend box office revenue, which will be a topic discussed in a future post.

Random Forest models are used to estimate opening weekend box office revenue. These models use two estimators per tree with a max depth of five. In addition to the total engagement estimates, the number of movie theaters on opening weekend, which is usually known by the movie studio well in advance, is also used as a potential predictor in the Random Forest models. For each spend and engagement level, the Random Forest model generates box office revenues 5,000 different times with different random seeds and train/testing splits (the train/test split used here is 92:8). Because we do multiple box office revenue predictions for a given spend (and thus engagement) level, we will have a distribution of estimates and can thus calculate the mean, median, and ranges of our predictions. This model will also allow us to assess the probability of our box office goals with given spend levels. Learn more about Random Forest models here.

Case Study

In this section, we present a case study with an anonymous theatrical client where the box office goal was $10 million given a distribution of their film across 2,000 movie theaters. We then vary the spend levels separately for both Facebook and Twitter to compute how the box office distribution changes. The case study concludes with an evaluation of both Facebook and Twitter simultaneously. 

Figure 2. The change in box office prediction distributions with varying levels of Facebook spend, given a 2,000 movie theater count.


The box office revenue distributions were computed for different levels of Facebook spend that range from $0 to $2.5M as shown in Figure 2. According to the model, it is clear that the $10M office mark isn’t attainable for campaigns that spend less than $0.7M, and even with that spend, the probability to reach that goal is low. In fact, the $10M box office mark is still unlikely and below the median value (of $7.2M) with a Facebook spend of $2.5M. 

The median box office estimate has the largest increase during the first $50,000 of Facebook spend. This implies that the first increment of Facebook spend is the most important. The minimum box office estimate levels off after this spend mark, and then remains flat after additional spend. This suggests that poor box office performance, perhaps due to bad movie reviews, will remain poor even with additional Facebook spend. However, this is not true for the maximum value, which steadily rises with increased spend. 

This particular model assumes no spend on other platforms. There is also a large dependency on movie theater distribution, meaning an increase in movie theater count would shift the curve upward in Figure 2

Figure 3. The change in box office prediction distributions with varying levels of Twitter spend, given a 2,000 movie theater count.


Like Facebook, the first $50,000 of spend on Twitter is the most important in terms of box office revenue predictions for the median, minimum, and maximum values as shown in Figure 3. The estimates flatten out between $70,000 and $180,000 spend, with very little return on investment (ROI) in that span. However, the Twitter spend-box office revenue curve begins to increase again after $180,000 spend for the median and maximum. The flat range for Twitter in Figure 3 is a result of the training data, which suggests a weaker relationship between Twitter engagement and Twitter spend over the mid-range engagement levels for movie theater counts close to 2,000. This weaker relationship did not exist in the Facebook data, and as a result, there is no flat range in Figure 2

The $10M mark for opening weekend box office revenue is never achieved by the median prediction, but can be achieved after a mere $70,000 of spend on Twitter. The main difference between Twitter and Facebook is the steepness of the slope between spend and box office revenue. Because Twitter’s slope is steeper than Facebook, it is clear that spending on Twitter will bring in more revenue on opening weekend relative to the same spend on Facebook. It is clear that the charts show the opposite of the conventional social wisdom of allocating more spend to Facebook.

Multi-platform Spend Allocation

Figure 4. Median box office predictions with varying levels of Facebook and Twitter spend with contour intervals at $0.8M.

When building a Random Forest model with both Facebook and Twitter engagement as predictors, we get another combined picture of how the spend allocation would affect box office revenue. Figure 4 shows how opening weekend box office revenue contours for a given Facebook-Twitter spend combination.

The results show that the goal of $10M in box office revenue is achievable with a $2.5M marketing spend on Facebook and a $300,000 marketing spend on Twitter; in fact, that particular spend combination yields a mean value above $10.4M. Based on these insights, the studio would be advised to spend at levels that are within the upper-right corner of the contour plot in Figure 4

Many of the features that appeared in the Facebook model and Twitter model results are also present in the combined platform results. For example, the sharp increase in box office revenue with initial Twitter spend is apparent by the tightly spaced contours at the bottom of Figure 4. Similarly, the lull in ROI increase from $70,000 and $180,000 can be seen by the widely-spaced contours in the middle of the plot. These results also suggest that if the studio were to lower their goal to about $7M, they could achieve their goal with no Facebook spend and a Twitter spend of less than $300,000. It is worth nothing, however, this outcome could be a result of external variables, including the number of movie theaters or number of platforms.

It is important to note that the model used here was trained on a limited dataset of social spend and with no non-social platform data (e.g., TV). The more data you can aggregate, the more robust your results will be. Nonetheless, this simple model of spend and opening weekend box office revenue can be useful in deciding how to allocate money on different social platforms during your marketing campaigns. Including more platforms can give a more complete version that would be akin to a media-mix model.