Essay

Why approximate future dev/test data before a product launch?

Question: Before releasing a new budgeting app, no real post-launch user data is available yet for dev/test splits. Explain why it is still important to approximate the data you expect after launch, and describe the main limitation of that approximation.

Sample answer: You still need to approximate future user data because dev/test sets should reflect the distribution the model must handle after launch. If the data used for evaluation does not resemble real future usage, model improvements may not transfer to the live product. In practice, you can build a proxy set using early beta users, pilot traffic, or other data that resembles the intended audience. The drawback is that this proxy may still differ from actual post-launch behavior, so it is only an estimate, not a perfect substitute.

Key points:

  • Dev/test sets should match the future target distribution as closely as possible.
  • Before launch, true future user data is unavailable.
  • Proxy data can come from beta users, pilot studies, or similar sources.
  • The approximation may not perfectly match real post-launch data.

Rubric: The answer should explain why a pre-launch proxy is needed to evaluate against the intended future distribution, and note that the proxy can still differ from actual future user data.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Machine Learning Strategy

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Yearning @ DeepLearning.AI

Related