Forecasting the customer lifetime value (CLV) is essential to allow User Acquisition managers to make the best possible decisions when optimizing campaigns. This CLV gives you the final revenue produced by users during their lifetime on the app, helping you to optimize your return on ad spend.
A deep dive into user data is essential in the mobile industry to understand how the customer will behave within the app. The Lumos Forecast aims to predict the CLV using only the first few days of data of the user. Following extensive data exploration, we have built a Machine Learning model that uses a mix of marketing data and in-app user anonymized as well as non-personal data.
One time users, revenue whales, market changes and app updates are several pitfalls that make the challenge more difficult and hence, exciting.
Lumos' strength is to team up with our partners to build the most accurate in-app tagging plan to allow data exploration to secure utmost precision. We also use our experience in marketing campaigns to extract the maximum number of insights from past operations and revenues. This is why our apps have specific projection periods that prevent those disturbances.
Once the feature engineering is done, the data goes through multiple steps: training a model, scaling and selecting the best features using multiple state of the art algorithms and also tuning the hyperparameters using Bayesian Optimization. The end goal is to fit the data to the model in the best way possible to produce forecasts for any situation. The results are then studied for each relevant cohort for each network.
Over the past few months, our model is becoming more and more accurate. To calculate the accuracy, we compare the real ARPU of the users to the forecasted ARPU provided by our algorithm. In July we reached an accuracy of 75%, then 82% in August and finally 88% in September 2021.
This new level of accuracy allowed Lumos to better optimize the User Acquisition campaigns and push the app to the top 20 US iOS in the Lifestyle category.