\chapter{Forecasting}
\label{chap-forecast}
\section{Introduction}
\label{sec:fcast-intro}
In some econometric contexts forecasting is the prime objective: one
wants estimates of the future values of certain variables to reduce
the uncertainty attaching to current decision making. In other
contexts where real-time forecasting is not the focus prediction
may nonetheless be an important moment in the analysis. For example,
out-of-sample prediction can provide a useful check on the validity of
an econometric model. In other cases we are interested in questions
of ``what if'': for example, how might macroeconomic outcomes have
differed over a certain period if a different policy had been pursued?
In the latter cases ``prediction'' need not be a matter of actually
projecting into the future but in any case it involves generating
fitted values from a given model. The term ``postdiction'' might be
more accurate but it is not commonly used; we tend to talk of
prediction even when there is no true forecast in view.
This chapter offers an overview of the methods available within
\app{gretl} for forecasting or prediction (whether forward in time or
not) and explicates some of the finer points of the relevant commands.
\section{Saving and inspecting fitted values}
\label{sec:fcast-fitted}
In the simplest case, the ``predictions'' of interest are just the
(within sample) fitted values from an econometric model. For the
single-equation linear model, $y_t = X_t \beta + u_t$, these are
$\hat{y}_t = X_t \hat{\beta}$.
In command-line mode, the $\hat{y}$ series can be retrieved, after
estimating a model, using the accessor \verb|$yhat|, as in
%
\begin{code}
series yh = $yhat
\end{code}
%
If the model in question takes the form of a system of equations,
\verb|$yhat| returns a matrix, each column of which contains the
fitted values for a particular dependent variable. To extract
the fitted series for, e.g., the dependent variable in the second
equation, do
%
\begin{code}
matrix Yh = $yhat
series yh2 = Yh[,2]
\end{code}
Having obtained a series of fitted values, you can use the
\texttt{fcstats} function to produce a vector of statistics that
characterize the accuracy of the predictions (see
section~\ref{sec:fcast-stats} below).
The \app{gretl} GUI offers several ways of accessing and examining
within-sample predictions. In the model display window the
\textsf{Save} menu contains an item for saving fitted values, the
\textsf{Graphs} menu allows plotting of fitted versus actual values,
and the \textsf{Analysis} menu offers a display of actual, fitted and
residual values.
\section{The \texttt{fcast} command}
\label{sec:fcast-fcast}
The \texttt{fcast} command generates predictions based on the last
estimated model. Several questions arise here: How to control the
range over which predictions are generated? How to control the
forecasting method (where a choice is available)? How to control the
printing and/or saving of the results? Basic answers can be found in
the \GCR; we add some more details here.
\subsection{The forecast range}
The range defaults to the currently defined sample range. If this
remains unchanged following estimation of the model in question, the
forecast will be ``within sample'' and (with some qualifications noted
below) it will essentially duplicate the information available via the
retrieval of fitted values (see section~\ref{sec:fcast-fitted} above).
A common situation is that a model is estimated over a given sample
and then forecasts are wanted for a subsequent out-of-sample range. The
simplest way to accomplish this is via the \verb|--out-of-sample|
option to \texttt{fcast}. For example, assuming we have a quarterly
time-series dataset containing observations from 1980:1 to 2008:4,
four of which are to be reserved for forecasting:
%
\begin{code}
# reserve the last 4 observations
smpl 1980:1 2007:4
ols y 0 xlist
fcast --out-of-sample
\end{code}
%
This will generate a forecast from 2008:1 to 2008:4.
There are two other ways of adjusting the forecast range, offering
finer control:
%
\begin{itemize}
\item Use the \texttt{smpl} command to adjust the sample range
prior to invoking \texttt{fcast}.
\item Use the optional \textsl{startobs} and \textsl{endobs} arguments
to \texttt{fcast} (which should come right after the command word).
These values set the forecast range independently of the
sample range.
\end{itemize}
What if one wants to generate a true forecast that goes beyond the
available data? In that case one can use the \texttt{dataset} command
with the \texttt{addobs} parameter to add extra observations before
forecasting. For example:
%
\begin{code}
# use the entire dataset, which ends in 2008:4
ols y 0 xlist
dataset addobs 4
fcast 2009:1 2009:4
\end{code}
%
But this will work as stated only if the set of regressors in
\texttt{xlist} does not contain any stochastic regressors other than
lags of \texttt{y}. The \texttt{dataset addobs} command attempts to detect
and extrapolate certain common deterministic variables (e.g., time
trend, periodic dummy variables). In addition, lagged values of the
dependent variable can be supported via a dynamic forecast (see below
for discussion of the static/dynamic distinction). But ``future''
values of any other included regressors must be supplied before such a
forecast is possible. Note that specific values in a series can be
set directly by date, for example: \texttt{x1[2009:1] = 120.5}. Or,
if the assumption of no change in the regressors is warranted, one can
do something like this:
%
\begin{code}
loop t=2009:1..2009:4
loop foreach i xlist
$i[t] = $i[2008:4]
endloop
endloop
\end{code}
\subsection{Static, dynamic and rolling forecasts}
The distinction between static and dynamic forecasts applies only to
dynamic models, i.e., those that feature one or more lags of the
dependent variable. The simplest case is the AR(1) model,
%
\begin{equation}
\label{eq:ar1}
y_t = \alpha_0 + \alpha_1 y_{t-1} + \epsilon_t
\end{equation}
%
In some cases the presence of a lagged dependent variable is implicit
in the dynamics of the error term, for example
%
\begin{align*}
y_t &= \beta + u_t \\
u_t &= \rho u_{t-1} + \epsilon_t
\end{align*}
%
which implies that
%
\[
y_t = (1-\rho) \beta + \rho y_{t-1} + \epsilon_t
\]
Suppose we want to forecast $y$ for period $s$ using a dynamic model,
say (\ref{eq:ar1}) for example. If we have data on $y$ available for
period $s-1$ we could form a fitted value in the usual way: $\hat{y}_s
= \hat{\alpha}_0 + \hat{\alpha}_1 y_{s-1}$. But suppose that data are
available only up to $s-2$. In that case we can apply the chain rule
of forecasting:
%
\begin{align*}
\hat{y}_{s-1} &= \hat{\alpha}_0 + \hat{\alpha}_1 y_{s-2} \\
\hat{y}_{s} &= \hat{\alpha}_0 + \hat{\alpha}_1 \hat{y}_{s-1}
\end{align*}
%
This is what is called a dynamic forecast. A static forecast, on the
other hand, is simply a fitted value (even if it happens to be computed
out-of-sample).
\subsection{Printing and saving forecasts}
To be written.
\section{Univariate forecast evaluation statistics}
\label{sec:fcast-stats}
Let $y_t$ be the value of a variable of interest at time $t$ and let
$f_t$ be a forecast of $y_t$. We define the forecast error as $e_t =
y_t - f_t$. Given a series of $T$ observations and associated
forecasts we can construct several measures of the overall accuracy of
the forecasts. Some commonly used measures are the Mean Error (ME),
Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean
Absolute Error (MAE), Mean Percentage Error (MPE) and Mean Absolute
Percentage Error (MAPE). These are defined as follows.
%
\[ {\rm ME} = \frac{1}{T} \sum_{t=1}^T e_t \qquad
{\rm MSE} = \frac{1}{T} \sum_{t=1}^T e_t^2 \qquad
{\rm RMSE} = \sqrt{\frac{1}{T} \sum_{t=1}^T e_t^2} \qquad
{\rm MAE} = \frac{1}{T} \sum_{t=1}^T |e_t|
\]
%
\[ {\rm MPE} = \frac{1}{T} \sum_{t=1}^T 100\, \frac{e_t}{y_t} \qquad
{\rm MAPE} = \frac{1}{T} \sum_{t=1}^T 100\, \frac{|e_t|}{y_t}
\]
%
A further relevant statistic is Theil's $U$ (Theil, 1966), defined as
the positive square root of
%
\[
U^2 = \frac{1}{T}
\sum_{t=1}^{T-1} \left(\frac{f_{t+1} - y_{t+1}}{y_t}\right)^2
\cdot \left[
\frac{1}{T} \sum_{t=1}^{T-1}
\left(\frac{y_{t+1} - y_t}{y_t}\right)^2 \right]^{-1}
\]
The more accurate the forecasts, the lower the value of Theil's $U$,
which has a minimum of 0.\footnote{This statistic is sometimes called
$U_2$, to distinguish it from a related but different $U$ defined in
an earlier work by Theil (1961). It seems to be generally accepted
that the later version of Theil's $U$ is a superior statistic, so we
ignore the earlier version here.} This measure can be interpreted as
the ratio of the RMSE of the proposed forecasting model to the RMSE of
a na\"ive model which simply predicts $y_{t+1} = y_t$ for all $t$.
The na\"ive model yields $U = 1$; values less than 1 indicate an
improvement relative to this benchmark and values greater than 1 a
deterioration.
In addition, Theil (1966, pp.\ 33--36) proposed a decomposition of the
MSE which can be useful in evaluating a set of forecasts. He showed
that the MSE could be broken down into three non-negative components
as follows
%
\[
{\rm MSE} = \left(\bar{f}-\bar{y}\right)^2 +
\left(s_f - rs_y\right)^2 +
\left(1-r^2\right) s_y^2
\]
%
where $\bar{f}$ and $\bar{y}$ are the sample means of the forecasts
and the observations, $s_f$ and $s_y$ are the respective standard
deviations (using $T$ in the denominator), and $r$ is the sample
correlation between $y$ and $f$. Dividing through by MSE we get
%
\begin{equation}
\label{eq:theil}
\frac{\left(\bar{f}-\bar{y}\right)^2}{\rm MSE} +
\frac{\left(s_f - rs_y\right)^2}{\rm MSE} +
\frac{\left(1-r^2\right) s_y^2}{\rm MSE} = 1
\end{equation}
%
Theil labeled the three terms on the left-hand side of
(\ref{eq:theil}) the bias proportion ($U^M$), regression proportion
($U^R$) and disturbance proportion ($U^D$), respectively. If $y$ and
$f$ represent the in-sample observations of the dependent variable and
the fitted values from a linear regression then the first two
components, $U^M$ and $U^R$, will be zero (apart from rounding error),
and the entire MSE will be accounted for by the unsystematic part,
$U^D$. In the case of out-of-sample prediction, however (or
``prediction'' over a sub-sample of the data used in the regression),
$U^M$ and $U^R$ are not necessarily close to zero, although this is a
desirable property for a forecast to have. $U^M$ differs from zero if
and only if the mean of the forecasts differs from the mean of the
realizations, and $U^R$ is non-zero if and only if the slope of a
simple regression of the realizations on the forecasts differs from
1.
The above-mentioned statistics are printed as part of the output of
the \texttt{fcast} command. They can also be retrieved in the form of
a column vector using the function \texttt{fcstats}, which takes two
series arguments corresponding to $y$ and $f$. The vector returned is
%
\[
\left(
\begin{array}{lllllllll}
{\rm ME} & {\rm MSE} & {\rm MAE} & {\rm MPE} & {\rm MAPE} &
U & U^M & U^R & U^D
\end{array}
\right)'
\]
%
(Note that the RMSE is not included since it can easily be obtained
given the MSE.) The series given as arguments to \texttt{fcstats}
must not contain any missing values in the currently defined sample
range; use the \texttt{smpl} command to adjust the range if needed.
\section{Forecasts based on VAR models}
\label{sec:fcast-VAR}
To be written.
\section{Forecasting from simultaneous systems}
\label{sec:fcast-system}
To be written.
%%% Local Variables:
%%% mode: latex
%%% TeX-master: "gretl-guide"
%%% End:
