What is a delta-delta feature?

The change of the change, used in speech features to capture how quickly the sound itself is accelerating.

Absolute or percentage change?

Both are used. Absolute keeps units clear; percentage compares across differently scaled accounts or products.

Do deltas help non-temporal models?

Yes, wherever a comparison is meaningful, such as treatment versus control or this region versus the average.

AI Glossary / AI and machine learning / Intermediate

Delta feature

Definition

A delta feature is an input built from the change between values, such as this month's sales minus last month's, so the model sees movement rather than only level.

What it means

Many business questions are really about change: is demand rising, is the customer cooling, is the machine drifting. A raw field such as monthly sales carries that information only implicitly, buried in the level.

A delta feature makes the change explicit. Subtract the previous value, or compare with a seasonal baseline, and the model receives the trend as a first-class input instead of having to reconstruct it.

In speech and signal work the term has a precise home: delta features are the frame-to-frame differences of the main features, capturing how the sound is moving, and they sit alongside the originals in the input.

Delta features are a form of feature engineering, judgment about what matters, expressed as arithmetic. They often lift a mediocre model more than a fancier algorithm would.

For a business, deltas are the cheapest upgrade to most forecasting and monitoring models, because the movement is usually what the decision actually turns on.

How it works

Choose a base field whose change matters, such as weekly active users, and a comparison point: the previous period, the same period last year, or a rolling average.

Compute the difference for every record: this period minus the comparison, sometimes expressed as a ratio or percentage instead of a subtraction.

Add the delta alongside the original level, not instead of it; level and movement answer different questions and the model benefits from both.

Handle the gaps honestly: the first record has no previous value, and missing periods distort the subtraction, so decide and document the fill rule.

In signal pipelines, compute deltas across consecutive frames with a fixed window, and optionally add delta-deltas, the change of the change, which capture acceleration.

Three practical examples

Churn modeling: change in a customer's order frequency over the last quarter, alongside the frequency itself.
Demand forecasting: week-on-week change in units sold, so the model separates a strong level from a rising trend.
Fraud monitoring: sudden change in transaction size or count for an account, compared with its own history.

Think of it this way

Think of a car's speedometer beside its odometer. The odometer is the level, the speedometer is the delta, and for deciding whether you will arrive on time the speedometer is the instrument you actually watch.

The limit: a speedometer is noisy on a bumpy road. Short-window deltas amplify jitter, so smoothing and window choice matter.

Worked example

adding deltas to a churn model

Suppose a subscription firm's churn model uses each customer's current monthly order count and plateaus at 74 per cent accuracy. The figures are illustrative.

The team adds two deltas: change in orders versus the previous month, and change versus the same month last year, leaving the original level in place.

Accuracy rises to 79 per cent, and the model's top early-warning list changes character: it now surfaces customers who are sliding, not only customers who are small.

The retention team prefers the new list, because a sliding big customer is worth a call and a stable small one is not.

Illustrative case study

A fictional equipment maker monitored machine vibration levels and caught failures only hours before they happened, too late to schedule repairs.

Adding the hour-on-hour change in vibration, rather than the level, surfaced a slow drift that preceded failures by days, because the drift, not the value, was the signature.

The invented lesson: when you monitor a level and get surprises, monitor its delta; many failures announce themselves as movement first.

Common misconceptions and mistakes

A delta replaces the original feature. Keep both; a customer can have a falling trend from a high level or a low one, and the distinction matters.
Deltas are only for time series. Any repeated measurement works: change between regions, between channels, before and after a treatment.
A bigger change is always a stronger signal. Deltas need normalizing by scale and noise; a 10 per cent move on a small account can be pure jitter.

Business use

List the decisions that depend on something rising or falling, and give the model a delta for each; do not expect it to compute change for itself.

Pair every delta with its window and missing-data rule in the documentation, because deltas are the features most likely to be silently miscomputed.

Watch delta features in drift monitoring: they are usually the first inputs to go strange when a data pipeline breaks.

In board packs, deltas are often the headline: the model that predicts the delta is usually the model the directors actually asked for.

Check your understanding

Why does adding a delta feature often improve a model that already has the raw level?

Answer: Because the level hides the movement, and the delta hands the model the change directly, sparing it from reconstructing a difference it cannot reliably infer.

Related terms

Feature engineering
Time series
Concept drift
trend
lag feature
delta-delta

Learn next

AI for Business Owners covers how to turn business questions into the inputs a model actually needs: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S8] Jurafsky and Martin, Speech and Language Processing (3rd ed. draft). Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Delta encoding Next: Delta modulation