AI Glossary / AI and machine learning / Intermediate
Precomputed Features
Definition
Precomputed features are input variables calculated in advance and stored, so that training or serving reads them ready-made instead of recomputing them from raw data every time they are needed.
What it means
The trade they embody is storage for speed. A feature like a customer's ninety-day spend or a text's embedding can cost real computation; computing it once and storing the result turns every later use into a cheap lookup.
They shine when features are expensive, reused often, or must be consistent: one computation, many consumers, and no risk that two pipelines compute the same feature two ways.
The risks are staleness and skew. A precomputed value reflects the moment it was computed, so a feature refreshed nightly is wrong for anything that happened today, and a feature computed by one pipeline must match what the serving path expects, down to the last default.
For a business, precomputed features are why some AI systems answer in milliseconds, and why a question worth asking is always: how fresh is this number, really?
How it works
Identify features that are expensive to compute or consumed by many models, and compute them on a schedule into a shared store.
Record each feature's computation time and refresh cadence, so consumers know the staleness they are accepting.
Serve training and inference from the same store, which removes a whole class of online-offline skew by construction.
Choose the refresh cadence from the decision's tempo: fraud scoring wants minutes, churn modeling tolerates days.
Monitor the gap between feature freshness and event time, and alert when refreshes lag their schedule.
Three practical examples
- A retailer's recommendations read precomputed product embeddings, answering in milliseconds what live computation would take seconds to produce.
- A bank refreshes customer aggregate features nightly; fraud features, which cannot wait, are computed on demand instead.
- A data platform stores one canonical ninety-day-spend feature, ending a era where three teams computed it three ways.
Think of it this way
Think of a restaurant's prep kitchen: sauces and stocks are made in the afternoon so the evening service is assembly, not cooking from scratch. Tonight's special, though, is cooked to order, because freshness beats speed for it.
Worked example
the stale scoreboard
Suppose a credit offers engine reads a precomputed income-band feature refreshed monthly. The figures are illustrative.
A customer's circumstances change mid-month, and for three weeks the engine offers limits based on the old band, until the refresh catches up.
The fix is tiered freshness: slow-moving features stay monthly, fast-moving ones refresh daily or compute on demand, and every feature carries its timestamp into the decision log.
The general lesson: precomputation buys speed with staleness, and the honest price list says which features pay it and how much.
Illustrative case study
A streaming service rebuilt its homepage ranking around precomputed features after live computation made page loads crawl at peak hours. Moving to a feature store cut response times by an order of magnitude and, unexpectedly, improved consistency: training and serving now read identical values, closing a skew that had quietly cost ranking quality. The trade-off arrived as a incident review when a cache of stale features survived a catalog relaunch for six hours. The fix was a freshness dashboard on the wall of the operations room: every feature's age, green to red, with expiry rules now treated as seriously as the features themselves.
Common misconceptions and mistakes
- Precomputed means computed once, forever. Values are refreshed on a cadence; the skill is choosing the cadence per feature.
- Freshness does not matter for features. Every precomputed value is a snapshot, and decisions built on stale snapshots are stale decisions.
- It is only a performance trick. It is also a consistency mechanism: one computation, consumed identically by training and serving, kills a class of skew.
Business use
Classify features by how fast their truth changes, and assign refresh cadences or on-demand computation accordingly.
Serve training and serving from the same precomputed store wherever possible, removing skew by construction.
Log feature timestamps into decisions, so stale-input incidents are diagnosable after the fact.
Put feature freshness on the operations dashboard; a stalled refresh should page someone, not surprise someone.
Plan cold-start behavior explicitly: new entities have no precomputed history, and the fallback is part of the design.
People also ask
What are precomputed features?
Input variables calculated ahead of time and stored, so models read them as lookups instead of recomputing them.
Why precompute?
Speed at serving time, consistency between training and serving, and savings on expensive repeated computation.
What is the cost?
Staleness: each value reflects its computation moment, so refresh cadence must match how fast the truth changes.
How do they relate to feature stores?
A feature store is the usual home for precomputed features, managing their computation, freshness and shared access.
What happens on a cold start?
New customers or products have no history to precompute from, so systems fall back to on-demand defaults until enough data accumulates for the store.
Check your understanding
What do precomputed features trade, and what question keeps them honest?
Answer: They trade freshness for speed and consistency: compute once, store, and let training and serving read identical values cheaply. The question that keeps them honest is how fresh is this number: every precomputed feature is a snapshot, and the refresh cadence must match how fast the underlying truth moves, or the system serves fast, consistent, stale answers.
Related terms
Learn next
AI Agents and Automation for Small Business covers freshness rules and shared stores for automated data work: https://aiforbusinessandlife.com/academy/ai-agents-automation-small-business [C4]
More in AI and machine learning
- Plate notation - Plate notation is a compact way to draw probabilistic graphical models with repeated structure.
- Polysemanticity - Polysemanticity is the property of a neural-network component, often an individual neuron, responding to several distinct features or concepts rather than...
- Positional encoding - Positional encoding is the information added to a transformer's inputs so it can tell where each token sits in the sequence, because the architecture itself...
- Prediction - A prediction is a model's answer for a case it was not told the answer to: the estimated price of a house, the class of an email, the demand for next month...
- Predictive estimation - Predictive estimation is judging a model or a parameter choice by how well it predicts unseen data, rather than by how well it explains or fits the data...
- Predictive ML - Predictive ML is the branch of machine learning that learns from historical examples to predict unknown outcomes: what will happen, which class a case...
Sources
[S7] Hopsworks, MLOps Dictionary. Full URLs are in the source ledger. Verified 11 October 2026.
Previous: Precomputation Next: Precondition