Predictions

Once an AI model instance is active and has completed its first run, BPP makes its prediction scores available as filter fields throughout the platform — in the audience builder and in signal value formulas.

Prediction refresh schedule

Models run daily by default (at 2:00 AM UTC). The latest scores are always from the most recent run. Historical scores accumulate over time so you can track how customer predictions evolve.

If a daily run fails, the previous scores remain available — BPP does not clear existing data on a failed run.

Output fields by model type

Action Prediction

Field	What it is
action_prediction_score	Probability (0–1) that this customer will perform the target action. Higher = more likely.
score_bucket	The probability split into deciles (1–10). Bucket 10 = top 10% most likely.

Training quality metrics (feature importance, model accuracy) are visible on the model instance detail page — not in the audience builder.

RFM Clustering

Field	What it is
recency_score	How recently the customer transacted (relative score 1–5, higher = more recent).
frequency_score	How often the customer transacts (relative score 1–5, higher = more frequent).
monetary_score	How much the customer spends (relative score 1–5, higher = more valuable).
rfm_cluster	Segment label: Champions, Loyal Customers, At Risk, Hibernating, etc.

pcLTV — Predicted Customer Lifetime Value

Field	What it is
pcltv_<horizon>	Predicted revenue from this customer over the configured time horizon (e.g., pcltv_6m = next 6 months).

Interest

Field	What it is
taxonomy_name	The type of interest model (product, iab, or custom).
value	A ranked list of this customer's top interests.

Using predictions in audiences

Once a model reaches Complete status, its output fields appear in the audience filter builder.

Example filter conditions you can build:

Model	Example condition	What it targets
Action Prediction	action_prediction_score >= 0.80	Top 20% most likely to perform the action
RFM	rfm_cluster IN (Champions, Loyal Customers)	Your highest-value segments
pcLTV	pcltv_6m >= 200	Customers expected to generate over €200 in 6 months
Interest	interest_iab contains Sports	Customers with a Sports interest signal

You can combine model output conditions with any other filter — for example: country = Italy AND action_prediction_score >= 0.75.

Using predictions in signals

Prediction scores can also be used as conversion values in signals. For example:

• ActionPredictionScore × 500 — send a value-weighted conversion to Google Ads, so Smart Bidding optimises for high-propensity users.
• pcltv_6m × 1 — send the predicted LTV as the conversion value so bidding optimises for long-term customer value.

When predictions become visible

Prediction fields are hidden in the audience builder until the model completes its first successful run. This prevents incomplete or placeholder data from appearing in your filter options. After the first successful run, the fields become available automatically.