Boost Programmatic Campaign Performance using Data Science
In recent years, programmatic advertising is taking over the online advertising industry. Programmatic advertising involves using technology for automatic selling and purchasing of ad impressions between advertisers and publishers through real-time auctions, Real-Time Bidding (RTB) is quickly becoming the leading method. Programmatic removes the human factor out of the process and allows for skipping tasks such as negotiations, tenders, and manual insertion of orders, which makes buying display space faster and cheaper.
In contrast to the traditional online ad market, where a certain amount of impressions is sold at a fixed rate, RTB allows advertisers to bid each impression individually in real-time at a cost based on impression-level features. Thus, RTB allows Addressable Advertising, i.e., the ability to serve ads to consumers directly based on their demographic, psychographic, or behavioral attributes.
By 2020, programmatic ad spend is expected to hit $68.87 billion.
Because programmatic advertising is automated, you don’t have to stop what you are doing to change an ad. This could make the personalization of ads easier than ever before.Instead of paying to advertise their product during shows with the most viewers, programmatic advertising allows companies to send targeted ads to very specific audiences.
Programmatic voice search ads are in their infancy, but this shift could have an impact on strategies moving forward.
75% of all video ads are programmatic. This means if you want access to the vast video audience and haven’t jumped on the programmatic bandwagon, you’re behind the times.
Many marketing firms act as agents for the advertisers and take part in the real-time auction on behalf of them. To achieve cost-efficient real-time bidding, quality inventory, or strategies and provide the advertisers with the desired performance (clicks, on-site traffic, etc.) at the lowest price possible, firms develop their own machine learning algorithms using techniques such as xgboost, random forests, neural nets, etc.
Need for Optimum Bid Prices and Quality inventory
As a marketing firm, it becomes a key responsibility of yours to execute the campaign on behalf of your client as efficiently as possible. Spending hundreds, thousands, or even millions of dollars a month on marketing, you better do it right!
As a campaign manager, you would want to keep spending money on strategies that are driving performance. Pausing spends on inefficient programmatic inventories. You are all set to get a positive return on your investment (ROI).But that’s not where this optimization should end. When you find something that is working and getting you customers, you can optimize it even further by leveraging the capabilities of machine learning algorithms and use that to tweak your campaigns to see if you can increase your conversion rates.
Programmatic marketing campaigns generally have different strategies specifically to optimize different KPI’s (Key Performance Indicators). These KPIs are specified by the advertisers and signify the business goal that they want to achieve like driving customers towards the advertiser’s website and contribute to different kinds of conversions (Home Page, Product Page, or Confirmation Page). The strategies mainly focus on :
- Branding of the client towards an unreached audience
- Identifying quality inventory (keywords, site domains, publishers, etc.)
- Getting incremental conversions to client
- Optimum bid prices for RTB of above-mentioned inventories
Optimizing these strategies is an ongoing practice that you or someone in your marketing team will continue to do on a daily, weekly, or monthly basis.
Four simple steps :
- Collect data
- Analyze data for insights
- Take action within your marketing campaigns
- Repeat daily, weekly, monthly
Marketing optimization can be frustrating, confusing and at times costly. And as ad tech systems become ever-increasingly complex and spread out, marketing optimization only promises to become more and more tedious.
Thus, to achieve the marketing goal you have to :
- Continuously identify the quality inventory or strategies
- Action back the learnings into the campaign at frequent intervals
An easy solution to this can be implementing some of the popular industry practiced machine learning algorithms into your data pipelines to identify the good strategies and action them back into the campaign at regular and frequent intervals.
Using ML/AI ensures minimum wastage of ad spend and less turnaround time between learning and activation of inventories into programmatic campaign, so as to drive overall marketing goal.
How to leverage ML + Data Engineering within Programmatic ecosystem
- Business and Data Understanding
Programmatic marketing campaigns generally have different strategies specifically to optimize different KPI’s (Key Performance Indicators). These KPIs are client-specific and as each KPI has different meanings, it’s evident that the data preparation process and underlying features affecting our target variable, i.e., KPI will be different. Below you can see the variety of features we deal in our ad tech ecosystem. Each auction corresponds to a user, which actually tells about the region, time, site domain, device, etc. when the ad impression was served. Based on different modeling experiments we will be identifying important features among these which directly affect the KPI and in the end will be targeting them through real-time auctions.
Given below is the snapshot of data prepared for one of the test cases. Do see different types of categorical features that will be used for modeling and evaluation.
It is also good to know about what is the distribution of the data you get through the existing budget spend across strategies. For example, let’s have a look at the distribution of ad impressions on the height feature of the ad banner shown.
We can see that the majority of impressions go towards height 250 yet it’s event rate or Target percentage is lower than of height 50. Thus, if we can add this type of data that can differentiate for our KPI, the model built will be robust. A similar analogy can be applied towards the width of the ad shown and other features shown above.
- Modeling and evaluation
Modeling is the heart of data science. It is performed in the following manner:
- Selection of modeling technique is the very first step to take
- It is followed by the generation of test scenarios for validating the model’s quality
- After that, a few more models are generated
- All the models are then assessed to make sure that they fall in line with the business goals
Below is a snapshot of model comparisons keeping Recall score as the comparison criteria.
Choosing the best model and then seeing the importance of different features gives us a rough idea of what combination of features affects the goal most and what to target. For example, past the modeling exercise, you can see Publisher Id, geo region, and site domain are some features that mainly direct towards the KPI success.
Like we discussed above, width 320 and height 50 are coming important compared to other versions but not at that scale. So our intuition was somewhat correct. Great!!
- Deployment
The important model features, here, the combination of publisher, geo, browser, and site domain, can then be used to target through DSPs and optimize towards KPI on any live campaign.
Let’s take an example. For one of the clients, we tried to follow the approach and activated important feature combinations to achieve optimum CPA (Cost incurred per acquisition) as KPI.
Over the weeks the model learned new feature combinations affecting the KPI and tried to achieve the optimum CPA at the lowest possible bid prices. The diagram below shows the performance of different strategies compared to our ML-based strategy, a good validation to know the ground truth.
The red line shows the performance of the ML strategy. Week over week it learned to activate on quality inventories and made the CPA less. Of course, some strategies are doing better but at higher bid prices. Let’s leave the bid prediction for the next blog.
Takeaways
- Overview of Programmatic Ad Tech and Real-Time bidding (RTB)
- Background on Programmatic marketing campaign goals and KPIs
- Leveraging ML algorithms to drive performance of programmatic advertising campaigns
