Product/Predictive attributesBeta

Score how likely each customer is to act, in real time.

Name any event as the outcome, and Signals builds, hosts and serves a model that scores each customer when your app asks.

The real engine · no card · no sales call
The problem

Most propensity scores are a day out of date.

They are computed overnight, so the score never reflects what a customer is doing right now, at the moment they decide.

How it works

The score moves during the visit.

Name any event as the outcome and pick the attributes to learn from. Signals builds the training set from your event history in Snowflake, trains and hosts the model, and scores the customer each time your app asks, from attributes current within <1s of their last event.

Capabilities

A real-time propensity score, without a data science project.

01

Scored on what they just did

The model runs on the customer's attributes when your app asks, so a third paywall hit this morning moves the score before the reader leaves the page.

02

Any event can be the outcome

Pick the event and filter it by its properties if you need to, such as an order_completed over $100 or a plan_upgraded to the annual plan. Signals builds the training set, trains the model and hosts it, so there are no features to engineer and no serving code to write.

03

Checked before it goes live

Each training run shows how the model did on data it held back and which inputs it relied on, and you can compare runs side by side. Publish the one you want, and retrain when your product changes.

Predictive attributes

Frequently asked questions.

Our analytics tool already gives us a propensity score. What is different?

When the score is worked out. Many analytics and marketing tools recompute propensity on a schedule, so the score your app reads was set before the visit began. Signals runs the model when your app asks, on attributes current within <1s of the customer's last event, and returns the score to your app.

What can it predict?

Whether a customer will do something later in the same visit, where that something is any event you track, such as subscription_started, order_completed or plan_upgraded, filtered by its properties if you need to. Outcomes that play out over weeks, such as churn, are not supported yet, and an outcome with no event to learn from cannot be learned.

Where does training run?

Signals builds the training set in your Snowflake account from your event history, then trains the model on Snowplow's infrastructure and hosts the version you publish. Snowflake is the only warehouse supported for now.

Is it generally available?

Predictive attributes are in beta. You can define, train, publish and score them today, with Snowflake as the warehouse. Your app reads the score from its own endpoint, and you retrain when you want a newer model.

Is the score a probability?

Treat it as a ranking from 0 to 1, where a higher score means the customer is more likely to reach the goal event. Set the threshold your app acts on by checking it against your own results.

Do we need a data scientist?

No. You define the attribute, train it and publish it from Console, with no feature pipeline or serving code to build. Each run reports how the model did on data it did not train on and which inputs it relied on, so a data scientist on your team can check the work.

Start with one attribute.

Add the SDK, define one attribute, and watch it update as you browse and search your own product. The free tier runs on the real engine: no card, no sales call.