You are on page 1of 4


Trend and Seasonal Adjustments

Econ 132
Day 4

To detrend a variable
Regress variable against trend (which may be linear or non-linear) and
subtracting it out of the original observations. The residual is the
detrended variable
Differencing. Differencing will typically remove the trend.
Does not require estimating any parameters (parsimonious)
Allows trend component to change over time (not deterministic)

Other methods (Hodrick-Prescott filter, etc.)

Trend and Seasonal Adjustments

Forecasting Model Evaluation

To deseasonalize a variable
Do seasonal differencing. Difference using a lag period that is
equivalent to one season (e.g. for quarterly data 4 lags, for monthly
data 12 lags, etc.)
Take running average of data (one-season equivalent time periods)

Source: Montgomery et al.

To deseasonalize and detrend

Sequentially difference. First seasonally difference to remove the
seasonal component and then difference one or more times (using
regular difference operator) to remove the trend.

How to evaluate the performance of a forecasting

technique for a particular time series?
How to choose between competing forecasting
Should not be based on only on the fit of the
forecasting model to historical data (or data used to
come up with the model).


Average error or mean error

Usually evaluation is done using one-step-ahead forecast errors
(1)= 1
where 1 is forecast of yt made one period prior.

Suppose that there are n observations for which forecasts have been
made and n one-step-ahead forecast errors (1), t = 1,2,,n.
What follows are some standard measures of forecast accuracy.

Mean absolute deviation or mean absolute


=1 | (1) |

Measures variability in forecast error.

=1 (1)

Estimate of the expected value of forecast error.

Unbiased forecasts if mean error is zero.
Biased forecasts if mean error is appreciably different
from zero.

Mean squared error


=1[ (1)]

Measures variability in forecast errors

Can be interpreted as an estimator of the variance of the
one-step-ahead forecast errors
The square root of the MSE is the standard deviation of
forecast errors


Relative forecast error or percent forecast

ME, MAD, and MSE are scale-dependent measures of forecast
Values expressed in terms of original units of measurement.
Difficult to interpret whether errors are small or large.
Not very useful for comparisons across different time series or across
different time periods.

1 =

100 =


For better comparability, measure of relative forecast error is needed.

Only meaningful if time series does not contain zero


Mean percent forecast error

Mean absolute percent forecast error

=1 (1)

=1 | (1) |


Choosing between competing models

Selecting model that fits historical data best may not result in
forecasting method that provides best forecasts of new data.

Too many parameters in the model because they improve the model fit.

Better approach is to select the model that results in smallest error

(MAPE, MPE, MSE, ME, MAD) of the one-step-ahead forecasts when
the model is applied to data not used in the fitting process.
Out-of-sample forecast error

Can be done by utilizing data splitting.

Divide time series into two segments one for model fitting and the other for
performance testing.