diff --git a/00_Introduction/00_Introduction.ipynb b/00_Introduction/00_Introduction.ipynb index 3c30845..e37d020 100644 --- a/00_Introduction/00_Introduction.ipynb +++ b/00_Introduction/00_Introduction.ipynb @@ -789,7 +789,7 @@ "
with the characteristic polynomials as defined above.
+In order to properly measure the performance of a forecasting model, we first need to establish several baselines. This section introduces several methods that will serve as benchmarks. Obviously, any forecasting method we develop must beat these benchmarks. Otherwise, the new method is not even worth considering.
+In the notation below, \(T\) refers to the length of the time series and \(h\) refers to the prediction horizon.
+Forecasts of all future values are equal to the last observation.
+Forecasts are equal to the last observed value from the same season of the year (e.g. the same month of the previous year).
+where \(m\) is the seasonal period and \(k\) is the integer part of \((h-1)/m\) (i.e. the number of complete years in the forecast period prior to time \(T+h\)).
+As an example, if we were forecasting a monthly time series, the forecast for all future February values is simply equal to the last observed February value. With weekly data, the forecast of all future Friday values is equal to the last observed Friday value. And so on.
+Forecasting is one of the most common inference tasks in time series analysis. In order to properly gauge the performance of a time series model, it is common practice to divide the dataset into two parts: training and test data. Model parameters are estimated using training data, then the models are used to generate forecasts that are evaluated against the test data.
Error statistics come in different flavors, each with their own advantages and disadvantages.
The handbook goes over several time series forecasting methods and compares performance of said models on the Jena Climate Dataset. Specifically, each method attempts to forecast the temperature variable (in Celsius). A summary of the forecast accuracy for each model is shown below.
+Method | +Average MAE (Celsius) | +
---|---|
Naive | +3.18 | +
Seasonal Naive | +2.61 | +
Linear Regression | +2.86 | +
ARIMA | +3.19 | +
VAR | +2.54 | +
Simplex Method (pending validation) | +1.53 | +
LightGBM | +2.08 | +