Along with facing limitations for measuring marketing results, global giants like Google and Apple, have begun to look for solutions that can give them comprehensive, clear, and actionable data on the efficiency of their budget. Divante Innovation Lab took up this topic and combined various models and analytical tools to create the most optimal DataOps on the market.
In order to quickly and efficiently create and evaluate new hypotheses, we created Predicty based on the idea of composable architecture. Our core algorithm is Robyn, made by Facebook. By adding other modules, we’ve expanded the functionalities of the whole product. It’s a tool that determines the optimal marketing investment.
Results: What’s in it for retailers?
We set up a framework for scaling marketing mix modeling (MMM) for future customers. One of the biggest challenges in modeling is data preparation and validation. Due to colossal amounts of data, it’s hard to analyze long time spans and nearly impossible to generate real-time analysis and predict future results. The innovative technology stack that we use allows for the reduction in the processing of large data volumes from dozens of hours to several minutes.
With Predicty, customers are able to calculate the impact on the key commerce key performance indicators (KPIs) that are usually used to track performance. There’s a north star parameter in the form of revenue that is then broken down into elements like:
- ROI, margin, frequency of purchase, channel saturation, and statistical attribution models for performance managers.
- Traffic and conversion analysis for eCommerce managers that provides much more valuable insights than last-click analysis.
Digital teams can easily check and simulate various marketing and eCommerce parameters, even day-to-day, as opposed to a one-time agency service.
These components include:
Budget allocator: By using the allocator, LPP can determine the impact of eCommerce components on the company's overall revenue. The allocator automatically calculates the most optimal budget distribution per channel or campaign.
Paid channels attribution: This process consists of looking for correlation between the output and input data, and calendar data. It provides more accuracy than last-click analysis for determining the impact of each channel on revenue, traffic, and more.
Halo effect: The halo effect is a natural phenomenon of the interaction of channels with each other and is impossible to measure with last-click analysis. Predicty shows, for example, how a given campaign on Instagram influenced the organic traffic from Google on the website.
Saturation curves: They show the growth potential of each channel. This provides knowledge about which channels are worth investing in and to what extent. By simply simulating various scenarios by increasing or potentially reducing the budget for various marketing campaigns, it can be seen when it pays off and when it brings a loss.
Return on impressions: This measures the number of people who actually see the ad or other marketing materials with their own eyes. It allows for determining what the average value generated by one impression on a given channel is.
Revenue simulation: Simulations allow for the quick consideration of various development scenarios of selected KPIs and predict future revenues based on data. It’s made possible by modeling the data in various ways like, for example, by reducing or increasing the expenditure on advertising.
The technology behind it
We enhanced Robyn with new simulation capabilities. These opened up completely new possibilities in planning and allocation.
The API-first approach made this process easier and quicker. Integration with BigQuery and being able to import from many formats allows flexibility in the selection of input data for prediction models. The queuing system allows for controlling time-consuming tasks, such as imports or fitting of a statistical model.
Elasticsearch and Kibana, included as one of the components, allow for effective time-series validation. The use of statistical models other than Robyn allows for cross-validation and quick hypothesis testing. The multi-layer cache system allows for convenient and quick access to data. Thanks to the smart use of DataFrame in the data processing process, we also maintain interoperability between Python and R environments.
A unique feature of Predicty is the reusability of trends between models along with combining models into larger volumes. The result of one experiment can also be a regressor for the next one. We create end-points so that we can easily return or use ready-made data without having to repeat it all over again.
At Divante, we have experience and awareness of the need for custom front ends and dashboards. This is why we’ve created an interface that adds a new quality along with the ability to be operated by non-technical people. By simply moving the buttons, it’s possible to simulate changes in the budget or the intensity of the campaign.
If you’re interested in MMM and prediction-based analysis, follow our entries because we’ll share the results of the next stages on an ongoing basis. If you’d like to see a demo of the product, you can subscribe here.