Predictive Modeling + Causal Modeling = Better Performance

Using a combination of predictive and causal modeling techniques to deliver higher ROI

First, let’s get on the same page for the definition of predictive modeling and causal modeling. Causal modeling is used to understand what events or actions influence others. It is an estimation approach based on the assumption that the future value of a variable is a mathematical function of the values of other variables.  On the other hand, predictive modeling is the process by which a model that best predicts the probability of an outcome is created. However, when it comes to predicting human behavior such as clicks and conversions, predictive model has its limitations. At RevX, we use the combination of both causal and predictive modeling practices to overcome the limitations and best optimize a campaign. We call it “Predictive Causal Modeling”.


We use predictive modeling to target advertisements to consumers and we observe that targeted consumers purchase at a higher rate subsequent to having been targeted. However, we cannot say with certainty whether the advertisements influenced the consumers to purchase or predictive models did a good job at identifying those consumers who would have purchased anyway. So it’s reasonable to assume that both are true and accordingly optimize campaigns.  We use site map analysis of advertiser's web sites to identify audience segments. We then determine their purchase intent intensity using various user signals, through randomized controlled experiments. Based on these observational data we typically create two kinds of campaign strategies. One strategy that targets high intent segments which largely consists of consumers who would have purchased anyway. While the second strategy targets relatively low intent consumers to whom we then show advertisements multiple times, converting them to high intent consumers over time.


When it comes to user behavior, historical data can be unreliable to predict the future. Any model based on historical data implicitly assumes that there are certain steady state conditions or constants in the system. This cannot always be true when involving people. For example, purchase intents change drastically when a flash sale is announced. What makes predicting accuracy difficult is that with more data there is more probability of noise. To overcome this problem, we use an in-house developed casual modeling technique of min/max bid price analysis. This technique removes tipping points from the predicted value of conversion that pushed the bid price over the edge leading to wastage of impressions. After filtering outliers or tipping points we significantly reduce the wastage of impression. This helps us improve margins across campaigns and deliver significant value to advertisers as we can now target a wider audience for the same budgets.

By combining the best practices of predictive modeling driven by prediction of outcome and casual modeling driven by observational data, we have significantly improved our capability to accurately predict the purchase intent of a consumer. It has helped us improve our margin and also reach out to a wider consumer set for the same budget. It has also helped to target relatively lower intent consumer and convert them to purchasers. After applying the casual modeling technique to predictive modeling we have observed that the number of new consumer’s purchase has doubled, providing evidence of the success of Predictive Causal Modeling techniques. 

Sandip Acharyya