Over the next couple of weeks, we're running content from our recently-published Programmatic Handbook. Here, Lewis Sherlock, head of demand platforms, international at Verizon Media Group/Oath, explains what to look for in your DSP’s machine learning systems.
Right now, every major DSP is talking about machine learning. Marketers are looking at machine learning as a way to make smarter buys and exceed campaign expectations. Machine learning can dynamically deliver relevant brand messages to the right people, in the right context and at the right moment. So in an industry where everyone is vying to tell their story as to how and why their machine learning is better or different, it is hard to determine how to cut through the marketing to investigate exactly how machine learning can benefit a marketer.
Here are four key questions to consider when evaluating how a DSP is using machine learning.
How accurate and diverse are the data points?
In a DSP, machine learning effectiveness is all about the data. Highly accurate data points and a diversity of data sources – whether it be email, search, apps, user registration, content consumption or others – are what make the engine run. And it shouldn’t come only from audience data, as is common in the RTB world. Instead, a DSP should also leverage deep-site segmentation data from both supply and demand. Only with a large amount of accurate and diverse data can a DSP create the best view of an advertiser’s most relevant audiences and reach them.
Fortunately, in the last year, data accuracy and diversity questions have become focus areas for marketers. 84 per cent of marketers say data accuracy is a critical concern, according to a recent Lotame study.
Where and how is machine learning used?
In assessing DSPs and their use of machine learning, it’s important to understand the areas and functions where it’s deployed. Every DSP does things differently. One might use machine learning in campaign optimisation and forecasting. Others might use it in their modelling of predictive audiences, where deep learning using neural networks analyses and scores relevant data sets to predict an audience’s probability to perform a specific action. And others do both. We are also using machine learning to build a recommendation planning engine into the DSP (powered by AdLearn) and enhancing our forecasting tool. Foundational machine learning use cases in a DSP can include:
Performance Prediction – Estimate the KPI rates (CTR, CVR, IVR etc.) per impression.
Control System – Maximise ROI while meeting pacing and performance constraints by computing campaign-level bid adjustments.
Forecasting – Predict properties of a campaign’s price– volume curve, which is then used to turbo-charge the control system to maximise efficiency.
Bidding – Combine performance predictions and information from the forecasting system to enable optimal bidding.
Given its complexity, it’s important for marketers to understand how machine learning is activated across a DSP. By demystifying use cases and gaining clarity, they can make smarter decisions and more targeted plans.
How flexible is the system?
How flexible are the DSP’s machine learning capabilities? Flexibility is key, because it speaks to the quality of the technology. For example, can the system optimise bidding for both first-price and second-price auction dynamics? Keep in mind, bidding for first-price inventory demands flexibility. It requires sophisticated prediction and forecasting of competing bids. Also, can the machine learning system optimise to brand, performance and multi-level goals? The rubric here can vary dramatically. For these reasons, a DSP with malleable machine learning capabilities is increasingly important today.
Is everything working together?
It’s not enough for a DSP to feature the right algorithms. It needs the right algorithms that are working together. There are standard machine learning algorithms out there which any DSP can leverage, but what really makes a machine learning engine stand apart is the ability to work in concert with other customised proprietary algorithms. This enables it to determine the best strategy and optimal bidding tactics to deliver against campaign goals. There must be connective tissue among systems so they can collaborate, learn from a campaign and create better performance. Believe it or not, many DSPs fail to deliver here.
Every DSP features machine learning technology today, but each one has different capabilities and degrees of sophistication. For advertisers to understand the best tools for their purposes, they need to ask questions and determine from the answers whether that DSP meets their needs. These four use cases are a good place to start and will help them move beyond the buzzwords.
You can read the entire Programmatic Handbook online here.