Blueshift Help Center

Set Up Predictive Models

Updated today

Within Blueshift’s Predictive studio, you can define and customize a predictive score based on your marketing objective and goal.

Predictive model lifecycle

Setting up a predictive model consists of several steps. It takes time for a model to learn from the data and even after a model is set up, the learning and fine-tuning continues and the predictive model becomes more refined.

View predictive models

You can view all the predictive models by going to Predictive Scores in the left navigation. Click the predictive model to view and edit the details.

Add a predictive model

You can set up the following predictive models in the Predictive studio:

Channel engagement model: This model predicts engagement on specific channels such as email, push, SMS, or in-app, based on past behavior.
Event engagement model: This model predicts engagement on specific events in the future, based on past behaviors.

Add a channel engagement model

To create a model that predicts engagement for a channel, click + Predictive Model. Select one of the following:

Email
Push
SMS
In-App

Add the following details for the model and Save.

Model Name	A unique name that identifies the model.
Goal Actions	The actions for which you are creating a predictive score. When a customer completes the action on the message that is sent using Blueshift, it is counted as conversion. You can select the following actions for the specific channels. Email: Open, Click Push: Click SMS: Click In-App: Open, Click
Goal Look-Forward	Predictions will be made and remain valid for this time period. The default time period is 7 days. For example, predict channel engagements in the next 7 days. Shorter time windows could lead to more accurate predictions, while longer windows could introduce noise/irregularities, which leads to lower predictive accuracy. Consider setting this value based on the historical conversion rate. For example, n number of conversions in the prediction window/number of active customers in lookback days * 100 is at least 1%.
Activity Look Back	Audience for the predictive scores will include those customers who had at least one action on the channel in the lookback period. The default value is 30 days (monthly active customers). Consider setting this value based on the historical conversion rate. For example, n number of conversions in the prediction window/number of active customers in lookback days * 100) is at least 1%.

Add an event engagement model

To create a model that predicts engagement on an event, click + Predictive Model and then select Event.

Add the following details for the model and Save.

Model Name	A unique name that identifies the model.
Goal Event	Select the events for which you are creating a predictive model. You can choose any combination of events that you are tracking with Blueshift. The presence of at least one of these events will be counted as a conversion. For example: purchase, subscription_renewal, etc. For more information about events, see Events in Blueshift.
Goal Look-Forward	Predictions will be made and remain valid for this time period. The default time period is 7 days. For example: predict purchase intent in the next 7 days or subscription renewal in the next 30 days. Shorter time windows could lead to more accurate predictions, while longer windows could introduce noise/irregularities, which leads to lower predictive accuracy. Consider setting this value based on the historical conversion rate. For example, n number of conversions in the prediction window/number of active customers in lookback days * 100 is at least 1%.
Activity Look Back	Audience for the predictive scores will include those customers who had at least one impression(event or campaign action) in the lookback period. The default value is 30 days (monthly active customers). This input should be equal to the length of the conversion funnel. In most use cases, the conversion is heavily correlated with the recent engagement, so 30 days (monthly) or 90 days (quarterly) is usually sufficient. E-commerce funnels are usually short, so you can set it to 7 to 14 days for most e-commerce purchase intent use cases. For Subscription businesses, you can set this input based on the billing cycles to predict subscription churns. Consider setting this value based on the historical conversion rate. For example, n number of conversions in the prediction window/number of active customers in lookback days * 100) is at least 1%.
Funnel Events (Advanced Options)	Add funnel events for the goal event. Funnel events are a set of events the customer has completed and which help to predict the likelihood of the customer completing the goal event.
User attributes (Advanced Options)	Include attributes like customer demographics (age, location), preferences (attributes of interest from the catalog), memberships (subscription type), lifetime attributes (lifetime orders, revenues), and so on. The latest value of the customer attributes is considered. For more information, see Customer attributes and types. Warning: Do not include a customer attribute that is updated as a result of the conversion. Otherwise, this would cause leakage (overfitting) and the model will not generalize well on unseen data. An example of one such customer attribute is “total_orders” for predicting purchase conversions. If we increment total_orders after conversion, for one-time purchasers, the model could confidently learn a simple rule that customers who convert must have total_orders > 0, which is not what we want.

Filter audiences

To filter audiences based on event counts and event attributes, contact [email protected]. Blueshift can set up the following customized filters:

Based on event count (for example, number of purchases > 0)
Based on event recency (for example, last purchase < 7 days)
Based on product attributes(for example, category, brand, sku) if product_sku is available on the event.

Related Articles