How Swappie used marketing mix modeling powered by Supermetrics and Google Cloud to improve ROAS by 15%

As we’re transitioning to a privacy-first online environment, many companies are looking into marketing mix modeling (MMM) as an alternative to measurement based on third-party cookies.

Swappie is the leading platform for buying and selling refurbished phones. The company is on a mission to make refurbished mainstream and reduce electronic waste—one refurbished phone at a time.

As a data-driven company, they use MMM powered by Supermetrics to measure advertising’s impact and optimally allocate budget.

Selecting Supermetrics as a partner

Prior to Supermetrics, Swappie was using Funnel as its data pipeline provider. According to Lauri, the major reasons for switching to Supermetrics were pricing and a better developer user experience.

Lauri says, “As we were evaluating alternatives, we were also in the middle of designing a new data architecture for our marketing data. Due to this, we evaluated alternatives not only from an MMM standpoint, but we required an overall strong backbone for serving marketing data for any possible purpose in the future.”

Funnel also provided less visibility into trouble-shooting. According to Lauri, “With Funnel, anyone could edit the data transformations and we lacked a change log or version control. If anything broke down, this made drilling down to the source of the problem difficult.”

The transition from Funnel to Supermetrics went smoothly with no problems in data quality. “We handled the integration project with Supermetrics and actually ended up with better integrations than we had in place before. We felt there was a more technical view to the integrations with Supermetrics”, Lauri says.

Setting up marketing mix modeling

Lauri and his team now use Supermetrics to pull online advertising data into BigQuery. Since the Swappie team does offline advertising as well, they feed offline data into their data warehouse separately. Once the data is inside BigQuery, they can transform, organize, and prepare it for modeling.

Lauri says, "Supermetrics makes it easy for us to get the raw data automatically in the format that's suitable for our modeling purposes."

They set up their machine learning pipeline on Vertex AI with data coming from BigQuery and Orbit—an open-source package for Bayesian time series modeling. Since Orbit is a regression model, it can predict sales outcomes while also estimating the certainty of the outcome. Additionally, with the Bayesian approach, it’s possible for the model to incorporate data from experiments to ground the prediction to observed reality.

Marketing Mix Modeling at Swappie

Increasing marketing efficiency with marketing mix modeling

Naturally, at some point, every marketing campaign will reach a threshold where each additional euro you put in will drive a lower impact. This is called diminishing returns. Identifying your point of diminishing returns will help you improve your campaign to get the most optimal results.

Besides, when you're doing MMM, you'll want to separate your sales into two components:

• Base sales are sales that will happen independently of short-term marketing activities, for example, returning customers, etc.
• Incremental sales are those additional sales driven by marketing activities, for example, display ads, promotions, etc.

Using their marketing mix model, the Swappie team simulates different investment scenarios and results.

First, they simulate channel efficiency. For example, the "Channel efficiency" graph below shows them how a channel's performance evolves over time—when it's most and least efficient—and the certainty of the prediction. The larger the spectrum, the more uncertain the prediction is. With this insight, they can estimate incremental sales and estimate the diminishing returns of different channels.

Lauri says, "We compare expected sales of different channels based on various spend levels. For example, you can see two curves of diminishing returns in the graph. It's clear that one channel drives more incremental sales than the other. Based on that insight, we'd put more money into the blue channel."

A dynamic estimation of channel efficiency over time and based on various spend levels.

Once the team knows each channel's incremental sales and diminishing returns, they can estimate the optimal budget. Lauri explains, "By simulating channel efficiency, we can predict the sales at different budget levels and how much we should spend on each channel."

Forecasted sales based on simulations of various budget levels.

Additionally, the Swappie team simulates their cost curves, as shown in the image below. Lauri elaborates, "In the ‘Cost by spend’ graph if you look at the red line, you can say the incremental cost of acquiring new customers for this channel is quite high with a low budget. If we increase the budget, the cost will drop significantly. But then it also goes up again. So we know the most optimal spending level for each channel, but we also know how much it will cost if we need to start pushing and scaling the market."

Finally, the team connects the cost with the profitability to see how much they should spend in order to be profitable. Lauri explains, “With the ‘Profitability by spend’ graph, we can project the profitability at different budget levels. The top of the curve that's where we're the most profitable. If our goal is profitability, we probably wouldn't want to spend more than that. But then, for growth reasons, we might want to go another direction."

Cost curves reveal the optimal investment levels based on the cost of incremental sales and on the estimated profitability at different spending levels to forecast budget and ROI.

Using these simulations, the Swappie team can adapt their marketing strategy and make smarter investment decisions. Lauri shares, "Without the marketing mix model, we'd be severely limited in having an objective view over our marketing activities. With the model using Supermetrics and Google Cloud, we've improved our marketing efficiency by at least 15%.”

Learnings from doing marketing mix modeling: Data quality, automation, and iterations

According to Lauri, if you’re thinking about building your own MMM solution, you need to think about data quality, automation, and iterations.

First, data quality is crucially important—good data makes a good model. Lauri says, “If you don’t have good data, you won’t get a good model. Your data engineering side needs to be in place and all the data sources need to work correctly.” He adds, “For example, we have quite an extensive transformation code base setup, so we have well-defined rules for how all transformations should be working.”

Next is automation. Since you need to feed a lot of data into your model, you’ll want to automate the process as much as you can. Lauri says, “If I had to do a lot of manual processes for this, probably I would be a lot less motivated to do it. And this is why having a solution like Supermetrics is key.”

Last but definitely not least is iteration. Start small, then iterate. Lauri shares, “Start with something easy, take baby steps, and see if things make sense. We started with a very simple linear regression model. But we improved it little by little.”

The future plan

Moving forward, the Swappie team will focus on improving the current model. Lauri shares, “We’re looking at better ways to do ad stocking and getting a better understanding of the best time to impact sales with marketing activities.”

He adds, “Supermetrics can help us achieve this goal by making sure the data flows correctly and automatically into our model. Going forward, I feel confident about scaling all the services deployed around our MMM."

Last, they also want to go more granular with the model. Lauri says, “In the longer term, we want to be able to drill down to the campaign level. We’d probably look at a hierarchical model to make this come true.”

Want to see what Supermetrics can do for you? Speak to our team to find the best solution for your business.