Generate ROI For Your Mobile Game

There is a lot to consider when predicting ROAS. The first is how far in the future the ROAS be predicted with certain confidence (3 months, 1-2 years, etc.). Of course, the longer the timespan, the more difficult it becomes to predict with certainty. 

The level of granularity needed for a useful prediction is also important. In some cases, daily cohort analysis may well be okay but in most cases, you will need at least campaign or user-level data – the likes of which can be hard to attain for mobile games post-IDFA. However, if there is any historical data available, the predictive models will be more accurate and therefore, more useful. 

Standard modeling approaches to predict ROAS

Option 1 – Take the last available ROAS number (e.g. day 14) and assume that there is a stable ratio between 6-months ROAS and 14-day ROAS. Looking at historical data, you can estimate this ratio and use it to predict 6-month ROAS based on the 14-day ROAS.

Option 2 – Estimate the retention to calculate the lifetime of a cohort and multiply it by average revenue per daily active user (ARPDAU).

Whilst these models are useful, they only work on an aggregated level. On top of this, they are not very helpful when working with smaller campaigns because they require large user samples. So, if you were trying to predict ROAS for a single, smaller campaign, these models wouldn’t give an accurate prediction. 

What’s different when SuperScale predicts ROAS?

At SuperScale, we take into account more than just the last available ROAS number. For example, daily growth over previous days is an important signal that is completely ignored if your model takes into account only the last available revenue metric.

We can identify cohorts that stopped growing sooner than the average cohort (because of payer churn) which helps the model to not overestimate the predictions. It is always better to have a conservative prediction that may underestimate the ROAS, than an overly optimistic one that motivates UA managers to spend more than is profitable.

Recognizing the different monetization behaviors from different user segments is also key. If you group players into daily cohorts, you mix all segments together and lose crucial pieces of information. We can calculate separate predictions for different segments giving us better overall accuracy.

Our machine learning (ML) prediction pipeline is built on experience from analyzing dozens of games across different genres. It’s one of the fundamental building blocks that enable us to have confidence in our industry-leading ROAS predictions.