Pattern: Slowly changing dimension, type 2
Keep every version of a row so "what was true in February" stays answerable after March changes it.
Tag
modelling
18 pieces tagged “modelling”, newest first.
When should I not build a model?
When a rule gets you most of the way, when you cannot get labels, when nobody will act on the prediction, or when being wrong is expensive and unexpla
Pattern: Baseline before model
Build the dumbest thing that could work and measure it properly. Your model must beat it on the same split, or it has done nothing.
Slowly changing dimension, type 2, in short
Keep every version of a row so "what was true in February" stays answerable after March changes it.
Pattern: Deterministic surrogate key
Derive the key from the content instead of an auto-increment, so reprocessing the same row produces the same key.
Forward-chaining split, in short
Validate the way you will deploy — train on the past, predict the future, and leave the gap your real prediction lag imposes.
The question that kills the most projects
If this were 100% accurate, what would change, and what is that worth?
As-of feature computation, in short
Every feature must be computable from information available at prediction time. Write down each one's timestamp, and the leakage disappears.
Pattern: As-of feature computation
Every feature must be computable from information available at prediction time. Write down each one's timestamp, and the leakage disappears.
The churn model that broke its own evaluation
A feedback loop, a holdout we should have kept, and the cost of finding out late.
Deterministic surrogate key, in short
Derive the key from the content instead of an auto-increment, so reprocessing the same row produces the same key.
We cut the modelling module in half
Four lessons became three, and the deleted one was about algorithms.
Pattern: Forward-chaining split
Validate the way you will deploy — train on the past, predict the future, and leave the gap your real prediction lag imposes.
Forward-chaining split, in short
Validate the way you will deploy — train on the past, predict the future, and leave the gap your real prediction lag imposes.
As-of feature computation, in short
Every feature must be computable from information available at prediction time. Write down each one's timestamp, and the leakage disappears.
Slowly changing dimension, type 2, in short
Keep every version of a row so "what was true in February" stays answerable after March changes it.
Baseline before model, in short
Build the dumbest thing that could work and measure it properly. Your model must beat it on the same split, or it has done nothing.
Baseline before model, in short
Build the dumbest thing that could work and measure it properly. Your model must beat it on the same split, or it has done nothing.