Must a dynamic model update continuously?

No; frequent scheduled or triggered retraining fits this usage.

Is online inference the same thing?

No; online inference means producing predictions on demand, while online or dynamic training concerns learning updates.

Can a new version be worse?

Yes. New data, modeling choices, or pipeline changes can reduce quality, which is why evaluation and monitoring remain necessary.

AI Glossary / Statistics and probability / Intermediate

Dynamic model

Definition

A dynamic model, in the machine-learning usage of the supplied glossary, is a model that is retrained or updated frequently as new data become available. Its learned parameters can change over time so it can respond to evolving patterns. This differs from a static model whose learned parameters remain fixed between deliberate replacements. Dynamic updating is a training arrangement, not a guarantee of continuous improvement.

What it means

A demand predictor trained once may gradually become less useful as customer behavior changes. A dynamic model incorporates newer evidence through scheduled retraining, triggered retraining, or incremental updates. The schedule can be frequent without updating after every individual observation.
The term also has other meanings in statistics and engineering, where a dynamic model can describe a system evolving through time even if its parameters remain fixed. Here the focus is the supplied Google glossary's frequent-retraining meaning. A time-series model is therefore not automatically dynamic in this particular sense.

How it works

Define the prediction task, available feedback, and conditions under which learning should update. Identify when labels become reliable enough to use. A model cannot legitimately train on an outcome before that outcome is available, and recent data may arrive with delays or revisions.
Choose an updating strategy. Full retraining fits a new model from a selected dataset, while incremental learning modifies an existing model using additional observations. A rolling window or recency weighting emphasizes newer data, but can discard useful history or amplify short-lived fluctuations. The appropriate balance depends on the process and objective.
Keep evaluation separate from the update. In sequential evaluation, predict first, record the result, and update only after the corresponding outcome becomes available. Evaluate candidate versions on appropriate future or held-out observations and check relevant subgroups, calibration, and operational costs.
Maintain data validation, model versioning, and a way to return to a prior working version. Automated retraining can propagate faulty labels, pipeline changes, or feedback loops more quickly if they are not detected. Model outputs can influence which outcomes are observed, complicating later training data.
Distinguish training frequency from inference timing. A static model can generate predictions on demand, and a frequently retrained model can generate predictions in batches. Changing a prompt, retrieving fresh documents, or updating a feature value also does not necessarily change the learned model parameters.

Three practical examples

A retailer retrains a demand model regularly as validated recent sales become available.
An operations team incrementally updates a duration predictor after completed tasks provide usable outcomes.
An AI classification service periodically evaluates a newly trained version before replacing the current version in production.

Think of it this way

Think of a dynamic model as a reference guide that is revised as new observations accumulate. Frequent revisions can keep it relevant, but each revision still needs accurate information and review. A guide that is revised quickly from poor evidence can become less reliable, not more.

Worked example

Consider a deliberately simple constant predictor for task duration. Its initial prediction is ten minutes. After each observed duration y, it updates its prediction m using m_new = 0.75 m_old + 0.25 y. This is an illustrative learning rule, not a universal dynamic-model algorithm.
Three successive observed durations are fourteen, eighteen, and sixteen minutes. The first prediction is ten, with squared error (14 - 10)^2 = 16. After observing fourteen, the model updates to 0.75 x 10 + 0.25 x 14 = 11.
The second prediction is eleven, with squared error (18 - 11)^2 = 49. Its update becomes 0.75 x 11 + 0.25 x 18 = 12.75. The third prediction is 12.75, with squared error (16 - 12.75)^2 = 10.5625. After observing sixteen, the prediction becomes 13.5625.
The mean squared error of the three predictions is (16 + 49 + 10.5625)/3 = 25.1875. A fixed predictor of ten has mean squared error (16 + 64 + 36)/3, approximately 38.6667 on these same observations.
This example favors updating on this short sequence, but it does not establish general superiority. Scoring a prediction after updating on its own outcome would leak information and give a misleading evaluation. Other sequences or update weights could produce a different comparison.

Illustrative case study

Fictional case: Seabrook Dispatch uses a fixed model to estimate task duration. When work patterns change, analysts propose more frequent retraining. They first discover that completion labels arrive several days late and that a recent logging change altered some timestamps.
The team repairs the data issue, defines a reliable label window, and evaluates successive model versions using information available at each prediction time. It introduces monitoring and a rollback procedure before increasing update frequency. The resulting process supports adaptation while preserving the ability to explain which model produced a particular estimate.

Common misconceptions and mistakes

Dynamic does not mean the system learns from every conversation or prediction automatically. Retraining on new data is distinct from using new inputs at inference. Frequent updating can overfit recent noise, forget older patterns, or worsen performance after a data problem. A model that describes changing system states can still have fixed parameters. Finally, an improving training metric does not establish improved performance on future cases.

Business use

Dynamic models can be useful when relevant patterns change and reliable feedback arrives often enough to support updates. Organizations should compare the benefit of fresher learning with training cost, validation effort, and operational risk. Define ownership of the update policy, preserve versions and data lineage, and measure performance after deployment. A stable task with limited new evidence may not benefit from frequent retraining.

Check your understanding

What is the third prediction in the worked example?

Answer: 12.75, before observing the third outcome.

What is the prediction after the third update?

Answer: 13.5625.

Would using the third outcome before scoring its prediction be valid?

Answer: No; it would contaminate that evaluation with information not available at prediction time.

Related terms

Online learning
Retraining
Concept drift
Training-serving skew
Static model
incremental learning
prequential evaluation
model versioning

Learn next

Compare scheduled retraining and incremental updates on a time-ordered dataset. Evaluate each prediction before using its outcome for learning, then investigate how label delay, window length, and update frequency affect performance and reproducibility.

Retraining

Sources

Google Machine Learning Glossary, dynamic model: https://developers.google.com/machine-learning/glossary#dynamic-model Google Machine Learning Crash Course, Production ML systems: Static versus dynamic training: https://developers.google.com/machine-learning/crash-course/production-ml-systems/static-vs-dynamic-training

Previous: Dynamic mode decomposition Next: Dynamic perfect hashing