On this page
What this lesson covers
Machine learning uses algorithms to find patterns in known data and apply those patterns, rather than relying on rules written by a person.
The lesson follows a revenue forecast through training, validation and testing, using time-based splits and rolling-window validation. It explains overfitting, underfitting and neural-network training, then compares supervised, unsupervised, reinforcement and deep learning. It also shows why new datasets and human judgment matter.
Key points
- Training set: The training set is where the algorithm finds the relationships between inputs and outputs.
- Validation set: The validation set is used to check those relationships, so the team can refine or tune the model.
- Test set: The test set is used last, to test how well the model predicts data it has not seen.
- Overfitted model: An overfitted model fits the training data very well, but it cannot predict accurately on a different dataset.
- Underfitted model: An underfitted model treats true features of the data as if they were noise.
Common mistakes
- Randomly splitting a time series can leak future observations into training for earlier test periods.
- Judging only the training months can make an overfitted model look excellent.
Try a practice question
Choose an answer, then check it to read the explanation for every option.
Continue studying in Quizara
In Quizara the lesson comes with more practice questions on the topic, explanations for every option and an AI tutor for your follow-up questions.
Written by the Quizara Editorial Team under our editorial standards.