Feature store
A feature store computes, versions and serves the model inputs used in training and in production, guaranteeing that both see the same definition of a feature. It removes training-serving skew, the most common silent cause of models that degrade after launch.
Why it matters for ARPU
Decisions are only as trustworthy as the features behind them, and a decision layer that cannot reproduce its own inputs cannot explain its own outcomes.
Related terms
Propensity modelA propensity model estimates the probability that a customer takes an action, such as buying, upgrading or cancelling.SignalA signal is an observed change in customer behaviour, product state, payment health or market context that carries information about future revenue.Real-time decisioningReal-time decisioning evaluates a customer's current context and returns an action within the latency budget of the moment, typically tens of milliseconds inside an app or call.Hypothesis provenanceHypothesis provenance is the complete, inspectable chain behind a decision: which signals raised it, which analysis sized it, who or what wrote the hypothesis, which experiment tested it, and what the causal read-out was.