Feature engineering examples

Feature engineering turns raw history (orders, sessions, transactions) into columns that describe each customer’s behavior in one row. The examples on this page are illustrative: the kinds of model-ready features segmentation benefits from, shown as recurring families with retail and e-commerce examples. The same families apply to subscriptions, gaming, travel, and most other verticals. Your actual feature set is designed with you during onboarding; this page is not a description of a fixed pipeline.

Feature names below are <angle-bracket> placeholders. The actual columns in your customer 360 are designed for your data and your goal.

Windowed spend and trends

Behavior is measured over a recent window (say, the last 90 days) and compared with the window before it. Lifetime totals hide direction; windows reveal it.

Feature	Plain-English example	Why it helps segmentation
<spend last 90 days>	”Jordan spent $420 in the last 90 days.”	Recent behavior describes who a customer is now, not who they were two years ago.
<orders last 90 days>	”Three orders this quarter.”	Separates frequent small buyers from rare big-basket buyers at the same spend level.
<spend trend vs prior period>	”Spend is up 30% versus the previous 90 days: a grower.”	Two customers with identical spend can be a grower and a decliner; trend is what tells them apart.

Recency and cadence

When a customer last acted, and the rhythm they act on.

Feature	Plain-English example	Why it helps segmentation
<days since last purchase>	”Last order was 12 days ago.”	The most direct freshness signal: active, cooling, or lapsed.
<typical gap between purchases>	”Usually buys every 3 weeks or so.”	Cadence is one of the strongest natural axes for grouping customers: weekly shoppers and twice-a-year shoppers behave like different species.
<overdue flag>	”Normally monthly, but silent for 10 weeks.”	Silence that exceeds a customer’s own rhythm is an early churn signal, long before an absolute cutoff would fire.

Engagement scores

A single signal that blends how recently a customer acted, how often they act, and how much they’re worth. The blend is a design decision made per engagement; the concept is “one number that summarizes engagement.”

Feature	Plain-English example	Why it helps segmentation
<engagement score>	”Bought last week, buys most weeks, spends meaningfully. Near the top of the range.”	Collapses three correlated behaviors into one interpretable axis, so segments separate cleanly into highly engaged, drifting, and dormant groups.
<engagement band>	”Sits in the High engagement band.”	A labeled band is easier to read in segment profiles than a raw score.

Basket and category diversity

What a customer buys, and how varied it is.

Feature	Plain-English example	Why it helps segmentation
<average basket size>	”Averages 4 items per order.”	Distinguishes stock-up shoppers from single-item buyers.
<categories shopped last 90 days>	”Bought from 6 different categories this quarter.”	Separates single-category loyalists from explorers, two very different cross-sell audiences.
<share of spend in top category>	”70% of spend goes to skincare.”	Concentration reveals what a customer is really here for, even when they dabble elsewhere.

Timing preferences

When customers shop, as a mix rather than a timestamp.

Feature	Plain-English example	Why it helps segmentation
<evening purchase share>	”Almost all of Jordan’s orders happen in the evening, hardly any in the morning.”	Day-part mix groups customers by routine and tells you when a message will land.
<weekend share>	”Shops almost entirely on weekends.”	Weekday-vs-weekend split often tracks lifestyle differences that other columns miss.
<holiday shopper flag>	”Active almost only around holiday periods.”	Isolates seasonal customers so they don’t blur into year-round segments.

KPI tiering

A continuous KPI can be banded into a small set of ordered tiers, typically Low / Med / High, with a separate band for customers with no activity at all.

Feature	Plain-English example	Why it helps segmentation
<annual spend tier>	”Falls in the High spend tier.”	Raw dollar columns are heavily skewed by a few big spenders; tiers give the model a balanced, interpretable target.
<visit frequency tier>	”Med visit tier: a regular but not a devotee.”	Tiered KPIs make segment comparisons readable: “80% of this segment is High spend” says more than a mean ever does.

Note. Good feature sets also handle edge cases deliberately. For example, a brand-new customer with no purchase history is marked as new rather than showing the same zeros as a lapsed customer, so the model never confuses the two.

Designed with you

You don’t build these yourself. During onboarding, Neuralift’s team designs the feature set with you, driven by your data, your vertical, and the goal in the use case definition, then runs preparation and hands you the result to review in the data dictionary. The features you add here become the levers your later tactics can pull, so if a play you want isn’t expressible downstream, this is the step to revisit; see From insight to action.

Next steps

• Make sure the raw inputs are in shape: Data requirements.
• See how the model uses these columns in Inside the Neuralift Segmentation Engine.
• See how segments are read against your KPIs in KPI lenses & lift.