첫 예측 기준선

Before fancy models, you need a baseline: a dead-simple forecast that anything fancier must beat to earn its place. 🪜

The simplest baseline is the naive forecast: predict that tomorrow equals today. Surprisingly hard to beat for steady data.

A naive forecast just repeats the most recent observation forward. No trend, no season, just the latest known point.

For cyclical data, predict the value from one full season ago. This seasonal naive respects repeating patterns cheaply.

Another baseline averages the last few points. A moving average smooths out noise and projects that level ahead.

Keep the last chunk of time as a test set. Never shuffle it: time order must be preserved so the future stays unseen.

Compare predictions to the held-out truth with an error metric. MAE reports the average size of your mistakes in real units.

Your baseline's error becomes the benchmark. Any model that scores worse is not worth its extra complexity.

Always plot the prediction over the real test values. Your eyes catch lag and bias that a single number can hide.

If a complex model cannot beat naive, trust the simple one. Simplicity that performs is a feature, not a failure.

With a baseline set, you are ready for real models like ARIMA or Prophet, each judged against this honest starting point.

What does a simple naive forecast predict for the next time step?

You built naive, seasonal, and moving-average baselines, then measured error so future models must prove they are better.