Back to Insights

Optimized Bids in 4 Clicks: Our Journey Building an Automated Bidding System

Let’s face the problem: the average day of a growth or user acquisition manager isn’t super fun. The lion’s share of the day goes to mundane tasks like pulling data, applying the same Excel equation to each network’s data, and uploading csv after csv. That is, if you’re working smart and not adjusting GEO by GEO or source by source from each network’s dashboard. Nasty job, but has to be done.

And all the while you are thinking of the exciting part of the job: finding patterns, exchanging ideas, developing hypotheses and testing them out, and hopefully every once in a while seeing the green numbers grow, but you don’t have enough time for all that because the campaigns don’t adjust themselves. Every campaign deserves a certain amount of attention, analysis, and actions, and there are so many decisions one can make in a day.

Reasons Current Methods Fell Short: Limitations of Existing ROAS Optimizers

You might say “Wait, but ROAS optimizer campaigns have already been introduced by networks to alleviate this problem, and they can also bid on a user level”. Yes, that’s right, however, they are usually exclusively available if you use the network’s mediation solution, and for every other network, the problem still stands. And that is if the ROAS model of the network works for your product, and more often than not, they don’t. What will happen is we end up not dedicating the attention each geo/network/platform/game combination deserves and we end up applying simplified logics to big bundles of combinations, reducing our sensitivity to fluctuations (either by bidding too conservatively or tanking a certain amount of loss every now and then) and oftentimes ignore some granular levels of optimization just because we can’t afford to do so (source bidding on a tier 4 GEO anyone?).

Solution: Automated Bid Optimization 

Here at Sunday we enjoy spending as much time doing the fun stuff and as little time doing the dull part of the job. We envisioned the ideal work day: we come to the office, have all the new and optimal bids handed to us on a silver platter, and apply them with only a few clicks before our morning coffee is cold enough to be served.

So we did just that. (Except for the silver platter, finance didn’t approve that one.)

We developed a system that fetches all the data, understands each campaign’s behavior and predicts how much revenue each bundle of users will ultimately make us, and suggests a set of bids accounting for our business objectives. We take these bids and upload them with a total of 4 clicks, and the bids of the whole portfolio is up to date.

Getting here was not an easy ride though. We tried many different methods and models, learned from our mistakes, and applied the successes to the next iteration. Here are a few of them.

Models: How to Predict User Behavior and Bid Accordingly

There are many models one can use for predicting ARPI development, and we did try quite a few including logarithmic curve fitting models, machine learning regression (XGBoost), Exponential Moving Average (EMA) and a few more. To evaluate the accuracy of each model, we tested them all against the historical data of our campaigns, and what we ended up using in the end was arguably the most simple one: progression multiplier. In this model, we measure the ARPI of an early day (3 or even 1) and multiply it by the Progression Multiplier, which is obtained from looking at the historical progression of the source in that certain GEO. This way we stay on top of ECPM changes as we use the most recent D1 or D3 revenue, and apply the historical progression of that campaign which accounts for how well the game is doing in that certain source & GEO combination. It also makes sure that you are taking into account the most recent user progression in case the latest game update has had any impact on user behavior. 

Granularity: How Deep Can We Go with Source Bidding?

With a big enough sample size, most models deliver reliable results. And luckily for a hyper-casual game, in which the behavior of an outlier like a whale has a negligible impact on the campaign level, the sample size doesn’t need to be too big which allows us to optimize on a very granular level. The smaller the sample size by which the model can create reliable predictions, the more granular you can afford to optimize. This is one of the main benefits of the multiplier model as it requires the least amount of data points compared to the other ones. 

Accuracy Score: How Reliable Are the Predictions?

There is no model that gets everything right, or at least we haven’t found it yet. So each prediction comes with a reliability score, which consists of 1 error for D1 ARPI, and one for progression of the game based on the historical data of the corresponding campaign/GEO/Platform/source combination. We can filter out the bids with a low-reliability score to avoid an unintended spike or drop in the campaign’s performance. The initial target we had for the model was to provide reliable predictions for over %80 of our ad spend, now we have it at %95.

Applying Bids: How to Ease the Hassle of Applying Tens of Sheets per Bidding Session

This part had already been taken care of by our attribution partner platform, JustTrack. It has implemented bid adjustment APIs of most networks, so all we have to do is to apply the standard format by which JustTrack accepts the bids to the system, after that, it’s just uploading one file and the bids are distributed to all the integrated networks. That’s where all the four clicks are needed!

Results and Future Work

Our bidding system is fully deployed and we’ve been using it for quite some time now. It has proven to be quite reliable and reacts to changes in market conditions quickly. But we didn’t stop there and added extra functionalities to it including a Discovery Mode, which manages the campaign bids on a higher level temporarily and looks for potential new sources that could bring additional profit. Many more features are also under development to increase the use cases and reliability of the platform. 

The greatest impact has been the time saved, which can be invested in more analytical, creative, and rewarding tasks. It has opened the doors to some growth possibilities that we simply couldn’t afford to do before. It also ensures none of the games of the portfolio are neglected and that they are all being managed by the highest standard.