Forecasts and Baselines
Build simple forecasts, measure their error, and learn why every model must beat a baseline.
Objective
By the end of this lesson, you can build two baseline forecasts, measure them with mean absolute error, and explain why a model must beat a baseline.
Start with a baseline
A baseline is the simplest reasonable forecast. If your fancy model cannot beat it, the model is not worth using.
Two common baselines:
- Naive: tomorrow will be the same as today.
- Seasonal naive: this Monday will be the same as last Monday.
Measuring error
Mean absolute error (MAE) is the average size of your misses, ignoring whether you guessed high or low. Smaller is better. It is in the same units as your data, which makes it easy to explain.
The seasonal baseline wins easily because it knows weekends are different. Any real model for this data must beat an MAE of 3.0 to be worth it.
Beyond baselines
Real forecasting tools, like the statsmodels library, model trend and season together. Run this in Colab once your data is a daily series:
import pandas as pd
from statsmodels.tsa.holtwinters import ExponentialSmoothing
series = pd.read_csv("daily_demand.csv", parse_dates=["date"], index_col="date")["demand"]
train, test = series[:-14], series[-14:] # hold out the last two weeks
model = ExponentialSmoothing(train, trend="add", seasonal="add", seasonal_periods=7).fit()
forecast = model.forecast(14)
print("Model MAE:", (forecast - test).abs().mean())
Notice the split: train on the past, test on the most recent days. Never shuffle time series data before splitting, or the model gets to peek at the future.
A national example
Judgment check
Forecasts fail on unusual days: storms, holidays, emergencies. If a forecast drives decisions that affect people's safety, like staffing a cooling center, plan for what happens when it is wrong. A forecast should come with a range, not just one number.
Quiz: Forecast check
Discussion
Ask. Answer. Build together.
Loading…