So, how exactly do these models work, and what problems do they solve for different businesses?
Predictive modeling at a glance
Predictive modeling is a type of intelligent data analysis where historical data is used to forecast future events. Machine learning technologies and statistical modeling methods allow you to make the most accurate predictions. And the more data you have for analysis, the more accurate the model will be — that’s why predictive modeling is most often used by large companies that handle massive volumes of data.
Such models help predict customer behavior, demand patterns for goods or services, fluctuations in company revenue, and even changes in the offline world.
How predictive models work in different areas
For years, predictive modeling has been aiding services used by millions of people every day. One example would be Yandex Maps, which determines travel time using predictive technologies. Yandex Maps not only analyzes the current traffic situation, but also considers how it’s going to change.
Another example is Yandex Weather. This service combines traditional weather models with machine learning technologies, including predictive modeling. The Meteum 2.0 technology crunches data from thousands of devices on Earth and in outer space, and leverages feedback from users who report the actual weather conditions.
There are many other notable areas where predictive modeling is used, including retail, to adjust the product range and set up upselling; banking, to assess client solvency; and large-scale manufacturing, to monitor the state of equipment. Finally, predictive models are also used in marketing to determine the impact of advertising campaigns on businesses and users. Let’s look at some examples.
Applying predictive models to marketing
In marketing and related fields, predictive models can help you:
- Determine the likelihood of an event. The more data you have on various factors that can affect an event, the more precisely you can predict the outcome.
- Estimate how a specific audience segment will perceive your advertising campaign.
- Determine the optimal cost of in-app purchases.
- Find the most loyal customers with the highest LTV, specific in-app time, or another key metric.
- Select the best optimization options, KPIs, and budget for your marketing campaign.
Why and how different companies make predictions
Building a predictive model requires a substantial amount of resources and expertise. That’s why marketers, data analysts, and product managers often opt for off-the-shelf solutions. However, some businesses choose to create custom models geared toward their specific needs.
We talked to a few companies that developed in-house predictive models and asked them to share their post-implementation results. Below, we also provide our own take on developing LTV prediction models.
Attracting and optimizing traffic
The digital agency Go Mobile regularly uses predictive models, going as far as developing a standalone, prediction-based product called Go Predicts. This service aims to engage and optimize traffic, taking into account audience behavior.
Based on the trained model, we simulate a specific event: link clicks, repeat purchases, or other user actions. After that, we optimize the ad platform for the given action.
This framework enables advertisers to immediately see and pay for users with the highest potential over their entire customer lifetime, not just the ones who made a single purchase. By optimizing the campaign for Smart Events, we reduced the CPO and CRR metrics for the client. As a result, the advertiser quadrupled their budget for predictive campaigns.
Optimizing the cost of relevant traffic
Rocket10 used predictive models to attract a target number of installs and registrations for the inDrive app in specific locations at the best price.
We began testing our sources with video creatives. During the process, we kept the most effective creatives, improved their visuals, and showcased the service via texts. To complement our video ads, we used static native banners that reinforced video messaging with positive statements. We also added occasional static interstitial banners and playable ads to keep our advertising fresh for the audience.
Once we gathered enough data, we enabled our predictive models. This helped optimize the cost of acquiring relevant traffic and significantly increase the likelihood of conversions.
As a result of our campaign, the Impression to Conversion (I2C) rate for app installs at this stage averaged 0.2%, which translated to one app install per 500 impressions. The I2C rate for registrations was 0.17%: one registration per 600 impressions. Consequently, from January to June, inDrive received 200,000 installs, with 120,000 of them falling in the period from May to June. According to average calculations, the number of installs and registrations per month tripled.
Attracting the most valuable audience
Tools based on machine learning — in particular, predictive models — help you fill in data gaps and scale up the patterns observed across the entire audience who are just starting to learn about your app through advertising.
The Crypta-based LTV prediction model is trained on anonymized Yandex Direct data for similar apps, as well as their revenue and user retention. This allows Yandex Direct to adjust bids in real time and attract users who would have a higher LTV in your app specifically.
We first started working with LTV predictions by running an experiment with Yandex Games. The results proved interesting, so we tested them on several more campaigns and then scaled the approach to all AppMetrica apps.
Currently, we determine the potential LTV of our users based on their first-day actions, using this data to estimate the revenue for the next 28 days. The obtained results can be scaled over a longer period to get the potential user value for the app over the entire lifetime.
When building our models, we take into account the fact that part of the revenue is already known at the time of applying the ML model, and that after the date of installation, revenue comes from a small slice of potentially long-term customers. The calculation formula looks as follows: revenue_before + p_revenue_after × revenue_after.
Note: “revenue_before” is the first day’s revenue known at the time of the LTV assessment, “p_revenue_after” is the model predicting the likelihood of a user becoming a long-term customer, and “revenue_after” is the model predicting how much revenue a long-term customer will bring after the first day.
For apps that don’t send revenue data to AppMetrica, a separate model is trained to predict LTV directly, without decomposition. The model is trained on anonymized data from apps that do share revenue data. This approach is slightly less effective, but it helps increase prediction coverage.
Conclusion
With the continuous advancement of technology, sophisticated predictive models are becoming increasingly affordable. Today, these models are available not only to large businesses with lots of audience data, but also to smaller teams aiming to make their ads more effective. LTV predictions in Yandex Direct are one example of how predictive models can be tailored to all sorts of marketing needs without limitations.