A fitted model earns its place when later forecasts improve a decision under a stated loss. This lesson compares squared and absolute forecast errors and builds an evaluation design that respects observation timing. It uses the ARMA forecast recursion, then explains a Diebold–Mariano comparison and its limits. Lower test loss is evidence about that sample and decision, not a guarantee of trading value.
Learning objectives. By the end of this note you should be able to:
- compute one-step and multi-step ARMA forecasts from the recursion rules;
- define MSE and MAE and choose a loss aligned with the decision;
- design a rolling or expanding out-of-sample experiment without leakage;
- state the Diebold–Mariano null for equal predictive accuracy at textbook level.
In this lesson
- Forecast generator (from ARMA)
- Loss functions
- In-sample vs out-of-sample
- Comparing two forecasts: Diebold–Mariano (textbook level)
Before you start
The full lesson, with worked examples and exercise solutions, is for members.
Read the full lessonNew to SelfMadeQuant.com? See membership options or browse the Econometrics and Financial Time Series syllabus.