OneTrace.pro Help center

Predictions

Predictions score each customer based on their purchase history and behavior: RFM group, churn risk, purchase probability, predicted revenue, engagement, and cart abandonment risk, as well as the best send hour and preferred channel. The results are written to the profile as regular traits, so you can use them right away in segments, journeys, and message templates.

The section is in the menu under Data → Predictions. All roles with access to analytics can view the page; Administrator and Owner can change settings and start a recalculation.

Turn on predictions

Purchase predictions start calculating once a purchase event is selected. Best send time and channel are calculated even without it.

  1. Open Predictions. If OneTrace.pro finds a likely event (for example, order_completed with an amount property), a hint appears at the top: click Use it.
  2. In the What counts as a purchase block, choose:
    • Purchase event: the event for a placed order;
    • Purchase amount: a numeric property with the order total. Without it, total spend, the monetary RFM score, and predicted revenue aren't calculated;
    • Purchase probability horizon: 7, 14, 30, 60, or 90 days, the period to predict a purchase for.
  3. In the Behaviour for the activity model block, set the Product view event and the Add to cart event, if you have them. They make the prediction more accurate, and cart abandonment risk is only calculated when a cart event is selected.
  4. Leave the Probability model set to Automatic: OneTrace.pro picks the more accurate model for your project.
  5. Click Save. The first calculation starts right away and usually takes a few minutes.

Prediction settings and model summary

When predictions are recalculated

Predictions are recalculated every night: purchases are taken from the project's entire history, and activity from the last 90 days. After you change the settings, a calculation starts right away, and the Recompute button starts one manually, for example after an import or a migration of order history.

Traits are assigned to customers and visitors who were active in the last 90 days. Only changed values are written to the profile, and each change can trigger a journey with the Trait changed trigger.

Which traits appear

Trait What it means
rfm_recency, rfm_frequency, rfm_monetary 1–5 scores for how recent the last purchase was, the number of purchases, and the total spent; 5 is best
rfm_segment RFM group: champions, loyal, at_risk, lost, and others
purchase_count, total_revenue, last_purchase_at number of purchases, their total, and the time of the last one
churn_risk churn risk, from 0 to 1
purchase_probability probability of a purchase within the selected horizon, from 0 to 1
predicted_revenue expected revenue from the customer over the horizon (an estimate of customer lifetime value, CLV)
next_purchase_days in how many days the customer is likely to buy again
engagement_score engagement from 0 to 100 based on activity over 30 days; 0 means no visits in 30 days
last_activity_at, sessions_30d, product_views_30d last activity, number of sessions, and product views over 30 days
cart_abandon_risk probability that an item added to the cart in the last 7 days won't be bought; cleared after a purchase
best_send_hour best send hour, 0–23, in the customer's time zone; see below
preferred_channel the channel the customer responds to best: email, sms, web_push…

The full list with explanations is in the Profile traits block at the bottom of the page.

RFM groups

Group in the interface rfm_segment value Who they are
Champions champions buy often and recently
Loyal loyal regular customers
Potential loyalists potential_loyalists recent, bought a second time
New customers new_customers one recent purchase
Promising promising one fairly recent purchase
Need attention need_attention average recency and frequency
About to sleep about_to_sleep one purchase, haven't come back in a while
Can’t lose them cant_lose used to buy very often but disappeared long ago
At risk at_risk bought a lot, but long ago
Hibernating hibernating rare, old purchases
Lost lost one old purchase

What the page shows

  • Model: calculation status, the selected model, the number of customers and of visitors without purchases, predicted revenue over the horizon, the number of open carts, how many profiles changed, and the calculation time.
  • RFM segments: how many customers are in each group.
  • Probability distribution: how many customers fall into each range of churn risk and purchase probability.
  • Engagement: the distribution of engagement_score.

Model comparison and probability calibration

Model quality

OneTrace.pro checks the prediction against the past: it makes a prediction as of one horizon ago and compares it with what actually happened.

  • Model comparison: metrics for the two models on the same customers. AUC shows how well the model separates buyers from non-buyers (0.5 is guessing; closer to 1 is better). Log-loss and Brier show how accurate the probabilities themselves are; lower is better. In Automatic mode, the model with the lower log-loss is chosen, and it's shown in bold.
  • Probability calibration: the predicted and actual share of purchases in customer groups. The closer the Predicted and Actual bars, the more you can trust the probabilities in segment rules.
  • What drives purchases: which signals raise or lower the purchase probability: active days, an open cart, time since the last purchase, and others.

Best send time and channel

Every night, OneTrace.pro looks at when each customer opens and clicks messages and writes two traits to the profile:

  • best_send_hour: the hour (0–23) in the customer's time zone when they most often responded to messages over the last 180 days. If there are few responses, the hour is based on website activity over 90 days. The time zone comes from the timezone trait; if it's missing, from the project settings.
  • preferred_channel: the channel in which the customer responds to messages noticeably more than the project average. Each channel is compared with its own average: emails count opens, while SMS and push notifications count only clicks, so their shares can't be compared directly.

These traits aren't written for new customers or those with little history. No purchase event is needed for them.

The Best send time and channel block shows how many customers fall into each hour, the Project best hour (the most common best hour among customers), and how many customers prefer each channel.

Best send time and channel: distribution by hour and preferred channels

Where it works:

  • in campaigns: At each recipient’s best time, each recipient gets the message at their best hour within a day;
  • in journeys: the Wait for best hour node before sending;
  • in the cascade: Preferred channel of the profile first;
  • in segments and conditions, as regular traits, for example preferred_channel = web_push.

How to use predictions

In the segment builder, add a Trait condition and choose a prediction trait:

  • "Valuable customers slipping away": rfm_segment is one of at_risk, cant_lose, for a win-back journey with a personal discount;
  • "About to buy": purchase_probability greater than 0.6; don't give them a discount, a reminder is enough;
  • "Churn risk": churn_risk greater than 0.7 and engagement_score less than 20;
  • "Abandoned cart": cart_abandon_risk greater than 0.5.

Prediction traits are also available in journeys (conditions and the Trait changed trigger) and in message templates, for example {{ traits.purchase_probability }}. The Decision journey node uses prediction traits to choose on its own who gets a discount, who gets a product selection, and who gets nothing, and it learns from the outcomes.

Good to know

  • Probabilities need history: at least 50 customers, 10 of them repeat buyers, for the purchase model, and at least 200 profiles for the activity model. Until there's enough data, the status is Not enough data, and only RFM scores, counters, and engagement are calculated.
  • Several purchases on the same day count as one.
  • The model doesn't account for seasonality or promotions: after a sale, probabilities lag behind reality for a while. Use them to compare customers with each other rather than as an exact forecast.
  • Returns aren't subtracted from the number of purchases.
  • The best hour builds up around the times when messages were already sent: if campaigns always went out at 10:00, the calculation won't see any other hour until messages start going out at different times.
  • Don't change prediction traits through the API or imports: they'll be overwritten at the next calculation.