%\VignetteIndexEntry{The forectADAPT Package}
%\VignetteDepends{}
%\VignetteKeywords{one-step ahaed forecasting, exponential smoothing, tuning parameter selection}
%\VignettePackage{forecastADAPT}

\documentclass[nojss]{jss}
\usepackage{orcidlink,thumbpdf,lmodern}

\newcommand{\class}[1]{`\code{#1}'}
\newcommand{\fct}[1]{\code{#1()}}

%% my packages
\usepackage{amsmath}
\usepackage{graphicx}
\usepackage{subfig}
\usepackage[onehalfspacing]{setspace}
\usepackage{geometry}
\usepackage{enumitem}
\usepackage{xcolor}
\geometry{left=0.9in,right=0.9in,top=1.1in,bottom=0.8in}
\setlist[itemize]{labelsep=0.1cm,leftmargin=0.9cm,parsep=0cm}

\author{Violetta Dalla\\National and Kapodistrian\\University of Athens \And Liudas Giraitis\\Queen Mary\\University of London \And George Kapetanios\\King's College London}
\Plainauthor{Violetta Dalla, Liudas Giraitis, George Kapetanios}

\title{Adaptive forecasting: Implementation (\proglang{R} package \pkg{forecastADAPT})}
\Plaintitle{The forecastADAPT Package}
\Shorttitle{The \pkg{forecastADAPT} Package}

\Abstract{Forecasting strategies that are robust to structural breaks and other forms of structural change have attracted renewed attention in the literature. These simple approaches are particularly appealing to applied econometricians, as they are based on weighted averages that downweight past information and can adapt in real time. They are easy to implement and remain effective under a wide range of non-stationary environments.
Our aim is to introduce the \proglang{R} package \pkg{forecastADAPT}, which implements such robust forecasting strategies through a range of illustrative examples and empirical applications. The methodology underlying the package, which covers forecast construction, estimation of forecast errors, and data-driven selection of tuning parameters, is based on the framework developed by Giraitis, Kapetanios, and Price (2013).}

\Keywords{one-step ahead forecasting, exponential smoothing, tuning parameter selection, \proglang{R}}
\Plainkeywords{one-step ahead forecasting, exponential smoothing, tuning parameter selection, R}

\Address{Violetta Dalla\\
Department of Economics\\
National and Kapodistrian University of Athens\\
Athens 10559, Greece\\
E-mail: \email{vidalla@econ.uoa.gr}\\
URL: \url{https://en.econ.uoa.gr/index.php?id=31606}\\
~\\
Liudas Giraitis\\
School of Economics and Finance\\
Queen Mary University of London\\
London E1 4NS, UK\\
E-mail: \email{l.giraitis@qmul.ac.uk}\\
URL: \url{https://www.qmul.ac.uk/sef/staff/liudasgiraitis.html}\\
~\\
George Kapetanios\\
Business School\\
King's College London\\
London WC2R 2LS, UK\\
E-mail: \email{george.kapetanios@kcl.ac.uk}\\
URL: \url{https://www.kcl.ac.uk/people/george-kapetanios}
}

\usepackage{Sweave}
\shortcites{forecast}
\begin{document}
\SweaveOpts{concordance=TRUE}
\input{forecastADAPT-concordance}

\section{Introduction} \label{sec:intro}

Exponential smoothing is a widely used method for forecasting univariate time series. The method relies on a tuning parameter, which can either be specified by the user or selected optimally. For instance, the function \code{HoltWinters} in the \pkg{stats} \citep{R} package implements exponential smoothing, with the tuning parameter chosen by minimizing the sum of squared forecast errors. Similarly, the function \code{ets} in the \pkg{forecast} \citep{forecast, forecast_jss} package and the function \code{es} in the \pkg{smooth} \citep{smooth} package perform exponential smoothing within a state space framework, allowing for optimal selection of the tuning parameter.

In these approaches, the full history of the data is typically used to estimate the tuning parameter. However, when the data exhibit structural changes or other forms of non-stationarity, relying on the entire sample may be inappropriate. In contrast, the method proposed in \cite{GKP2013} provides a theoretical justification for selecting the tuning parameter based on the recent past, thereby allowing the procedure to adapt to changes in the data without requiring prior knowledge of its underlying structure; see also \cite{GKMP2015}.

The \proglang{R} \citep{R} package \pkg{forecastADAPT} implements this approach through the function \code{forAD}, which performs adaptive exponential smoothing with data-driven selection of the tuning parameter. Moreover, the method can be combined with AR forecasting to account for remaining dependence in the data. Section \ref{sec:forAD} describes the forecasting procedure in detail and provides illustrative examples.

\section[Adaptive forecasting: forAD]{Adaptive forecasting: \code{forAD}} \label{sec:forAD}

In this section we outline the basic theoretical ideas behind the practical implementation of the adaptive forecasting method for a univariate time series $\{x_t\}$, as developed in \cite{GKP2013}. The time series is assumed to have an unspecified and a priori unknown structure: it may be stationary or non-stationary, may include deterministic or stochastic trends, and may undergo structural changes.

The forecasting method consists of two steps. First, the adaptive forecast is used to predict the persistent part of the data and to decompose the
data series  into a persistent component (signal) and a “stationary” remainder. Second, an AR forecasting method is applied to the stationary remainder by fitting an AR model.

The resulting procedure thus combines adaptive forecasting for the persistent (non-stationary) component   with
AR forecasting for the component exhibiting stationary dynamics.


Suppose  we are  given  a   sample  $x_1,...,x_{N}$, and we are interested  in the   forecasting of  $x_{N+1}$.

The  adaptive  forecast of $x_{t}$ given $x_1, ..., x_{t-1}$, as developed in \cite{GKP2013},
   accounts  for  the potential structural  change  by  extracting  the   persistent component of the  data. It
is computed as  the  weighted average:
\begin{equation}\label{for_1}
\hat x_{t|t-1,\rho}=w_{t1,\rho}x_{t-1}+...+w_{t,t-1,\rho}x_1,
\end{equation}
  where the  non-negative  weights  $w_{tj,\rho}$, $j=1, ..., t-1$ are  parameterized  by a  tuning  parameter $\rho$ and  sum up to  $1$:
$$ w_{t1,\rho}+w_{t2,\rho}+...+w_{t,t-1,\rho}=1.$$
We use  the  exponentially  decaying weights
$$
w_{tj,\rho}=\frac{\rho^j}{\sum_{s=1}^{t-1}\rho^s}, \quad j=1,...,t-1
$$
with   $0<\rho\le 1$.
A key element  in the  adaptive  forecast  (\ref{for_1}) of $x_t$ is a data-based  selection of the  tuning parameter $\hat\rho_{t}$,
which is  used  to  compute the  forecast $\hat x_{t|t-1,\hat\rho_t}$.
 The optimal tuning parameter $\hat\rho_{t}$
 as shown in \cite{GKP2013} can be found by minimizing the mean squared forecast error, $Q_{t,\rho}$, over a training period $t-t_0,...,t-1$ of length $t_0$:
\begin{equation*}
\hat\rho_{t}={\rm  argmin}_{ \rho\in (0,1]}Q_{t,\rho}, \quad
Q_{t,\rho}=\frac{1}{t_0}\sum_{s=t-t_0}^{t-1}\big(x_s-\hat x_{s|s-1,\rho}\big)^2.
\end{equation*}

The package utilizes the \code{optimize} function from the \pkg{stats} \citep{R} package, where we set for  $\rho$  \code{lower = 0.001} and \code{upper = 1} in the bounded optimization.

\subsection{One-step  adaptive  forecasting procedure: Adapt }
The  typical adaptive forecast analysis    proceeds  as  follows. It  involves an out-of-sample forecast  of  $x_{N+1}$ and the evaluation of adaptive forecasts over  an  evaluation period $(N-E, ..., N]$, where
 $E$ is a pre-selected length of  the evaluation period.

1. Out-of-sample forecast of $x_{N+1}$:

For a    given sample  $x_1, ...., x_N$,  the  adaptive   out-of-sample    forecast  $\hat x_{N+1|N,\hat \rho_{N+1}}$  is    computed, and the tuning parameter  $\hat \rho_{N+1}$ is recorded.

2. Forecasts over the evaluation period:

For each $t$  in  the  evaluation period $ t\in (N-E,\, N]$,  the   adaptive   forecasts   $\hat x_{t|t-1,\hat \rho_{t}}$ are computed, the tuning parameters  $\hat \rho_{t}$ recorded, and the  errors
\begin{equation}\label{err_in_1}
e_t=x_t-\hat x_{t|t-1,\hat\rho_t}
\end{equation}
are   obtained.

3. Evaluation of forecast accuracy:

 To evaluate  forecast quality over  $ t\in (N-E,\, N+1]$,
the mean squared forecast error (MSFE) for the adaptive forecast  $\hat x_{t|t-1,\hat\rho_t}$
 is  computed as
\begin{equation}\label{msfe_in_1}
{MSFE}_t=\frac{1}{n_v}\sum_{s=t-n_v}^{t-1} e_s^2,
\end{equation}
where the computation for  each $t$ is
based on the   past  $n_v$  forecast errors $e_s$, with  $n_v$   fixed and preselected. This   produces   a  series   of  MSFEs:
$$
{ MSFE}_t, \quad  t\in (N-E,\, N+1].
$$
The package  provides both  numerical and  graphical outputs  of $MSFE_t$, which allow comparison and evaluation of different forecasting methods over the evaluation period.

In the package,   the methodology described above is referred to as ``Adapt'', denoted by the  superscript $^{(ad)}$, i.e. $\hat x_{t|t-1,\hat \rho_{t}}=\hat x_{t|t-1}^{(ad)}$.

In sum, given the  sample  $x_1, ...,  x_N$,
the   adaptive  forecast   produces:
\begin{itemize}
   \item The forecast $\hat x_{N+1|N}^{(ad)}$,
    \item The tuning parameter $\hat \rho_{N+1}$,  used  to  compute  this  forecast,
 \item The    MSFE$_{N+1}$.
\end{itemize}
In addition, the  package provides outputs (plots  and numerical  values)  of  these   quantities   and   forecast  errors $e_t$,
 for  the entire evaluation  period:
\begin{eqnarray*}
\hat x_{t|t-1}^{(ad)}, \,\hat \rho_t,\,\,e_t, \,\,MSFE_t, \quad  \mbox{for \,\,$t\in (N-E, N].$}
\end{eqnarray*}

The adaptive forecast $\hat x_{t|t-1}^{(ad)}$
  can be interpreted  as {\it  signal  extraction}. It
decomposes    $x_t$ into  a persistent component (signal)  and a noise component:
\begin{eqnarray}\label{e:for1}
x_t &=& \hat x_{t|t-1}^{(ad)}+e_t,
 \quad t\in (N-E,\, N],\\
&=:&Adapt_t+u_t,\label{e:for1L}
\end{eqnarray}
where $Adapt_t = \hat x_{t|t-1}^{(ad)}$ is  the signal,  and the noise  component $u_t = e_t$ is the forecast error.

In most  of  the  cases,
the noise   component
$u_t$ is
  approximately  stationary  and weakly    dependent.
This  series $u_t$   can be   further modelled and   predicted   using  an AR  process, see  Section \ref{e:comb}.

When the  data  series  $x_t$
   exhibits  abrupt  changes (e.g., breaks in the  mean),
   the persistency and  non-stationarity
 might be not fully captured by $ Adapt_t$.
 In such cases, it may be necessary to perform an additional adaptive forecast on the forecast errors $e_t$ in (\ref{e:for1}), effectively applying a second-level adapt procedure.


   \subsection{Two-step  adaptive  forecasting procedure:  Adapt$^2$}\label{e:2step}
   To  assess  whether a  two-step   adaptive  forecasting procedure is required,
    the user applies a  second  adaptive  forecast step
   to
the  adaptive forecast   errors $e_t$ in   (\ref{e:for1}), with the  aims  to  extract  any remaining persistency and non-stationarity.

For each $t\in (N-E,N]$, this procedure  produces  the following   outputs:
\begin{itemize}
\item the adaptive forecast of the  errors, $\hat e_{t|t-1}^{(ad)}$,
\item the resulting residual  $u_t=e_t-\hat e_{t|t-1}^{(ad)}$,
\item  the $MSFE_t$,
\item and the tuning  parameter $\hat \rho_{t}$.
\end{itemize}
For  $t=N+1$, the procedure provides  the   out-of-sample forecast $\hat e_{N+1|N}^{(ad)}$, along with  $\hat \rho_{N+1}$
and   $MSFE_{N+1}$.

\vskip.2cm
\noindent Two step  adaptive  forecasting procedure  allows  the decomposition of  the data  $x_t$  into  a persistent signal  and a noise component:
\begin{eqnarray}\label{e:for4}
x_t &=& \hat x_{t|t-1}^{(ad)}+\hat e_{t|t-1}^{(ad)}+u_t,  \quad t\in (N-E,\, N]\\
&=&Adapt_t^2+u_t,\nonumber
\end{eqnarray}
where the  two-stage  forecast adapt$^2$, denoted by $$Adapt_t^2=
\hat x_{t|t-1}^{(ad2)},\quad
\hat x_{t|t-1}^{(ad2)}=\hat x_{t|t-1}^{(ad)}+\hat e_{t|t-1}^{(ad)},$$
represents
the  signal, extracted by the two-step adaptive procedure, and $u_t$ is the corresponding forecast error.

The package also provides an out-of-sample adaptive forecast of  $x_{N+1}$  given by   $$Adapt_{N+1}^2= \hat x_{N+1|N}^{(ad)}+\hat e_{N+1|N}^{(ad)},$$
together with the tuning parameter
$\hat  \rho_{N+1}$ and  MSFE$_{N+1}$.

In most   cases,  the two-step adaptive procedure transforms  data $x_t$ into:
\begin{itemize}
\item a persistent (potentially non-stationary) signal $Adapt_t^2$, and
\item a noise component $u_t$ that is approximately stationary and weakly dependent.
\end{itemize}
This noise component $u_t$ can then be further modeled using a stationary AR process; see Section \ref{e:comb2}.

\vspace{0.2cm}

In the package,   the methodology described in (\ref{e:for4}) is referred to as ``Adapt$^2$''.
\vspace{0.2cm}
To determine whether the second adaptive step is necessary, one compares the  MSFEs from the one-step and two-step procedures over the evaluation period $(N-E, N]$:
\begin{itemize}
\item If, in general,$$MSFE_t^{(ad2)}\ge MSFE_t^{(ad)},$$
where $MSFE_t^{(ad2)}$ and $MSFE_t^{(ad)}$ denote the MSFEs of the adapt$^2$ and adapt methods, respectively, then no additional persistence is extracted from $e_t$, and the second adaptive step is not needed.
\item If, on the other hand,
$$MSFE_t^{(ad2)}< MSFE_t^{(ad)},$$
particularly for the out-of-sample forecast at $t = N+1$, then the two-step adaptive forecast should be preferred.
\end{itemize}


\subsection{ Combined forecasting procedure: Adapt+AR forecast }\label{e:comb}
This   method   combines   the
Adapt and  AR  forecasting procedures.
First,   over the evaluation period  the adaptive forecasts and corresponding forecast  errors, denoted by Adapt$_t$   and  $u_t$, for  $t\in (N-E,  \, N]$, are  computed as  in  (\ref{e:for1L}).

Next, for  each $t\in (N-E,  \, N+1]$,    an   AR($p$), $p=1,2,...,p_{max}$, forecast of   $u_t$ is  obtained:
\begin{eqnarray}\label{e:fore1-}
u^{(ar)}_{t|t-1}&=&\hat\phi_{0}+\hat\phi_{1}u_{t-1}+...+\hat\phi_{p}u_{t-p}\\
&=:&\hat\phi_{0,t}+\hat\phi_{1,t}u_{t-1}+...+\hat\phi_{p,t}u_{t-p}  \notag
\end{eqnarray}
by  fitting the   AR($p$) model
\begin{equation}\label{e:fore1--}
u_j=\phi_{0}+\phi_{1}u_{j-1}+...+\phi_{p}u_{j-p}+\xi_{j}
\end{equation}
to the subsample $u_j, \,\, j\in [t-n_{ar}+p, \, t-1]$,   where the   subsample  size $n_{ar}$  is  fixed.

The  estimates $\hat\phi_{0,t}, ...,\hat\phi_{p,t}$  may  vary with  $t$.

 This leads    to the decomposition
\begin{equation}\label{e:fore1}
u_t=u^{(ar)}_{t|t-1}+\varepsilon_t, \quad t\in (N-E,N],
\end{equation}
where $u^{(ar)}_{t|t-1}$  is the AR  forecast
and     $\varepsilon_t$ is the  associated forecast   error.

The   combined   forecast of  $x_t$  is then  given by
$$\hat x^{(1, ar)}_{t|t-1}=Adapt_t+u^{(ar)}_{t|t-1}, \quad  t \in (N-E, N+1].
$$
Accordingly, over  the  evaluation period, the data can be decomposed as
\begin{eqnarray}\label{e:for6}
x_t &=&
\hat x_{t|t-1}^{(1,ar)}+\varepsilon_t,  \quad t\in (N-E,\, N],
\end{eqnarray}
where $\varepsilon_t$ is the forecast error of the combined method, as defined in (\ref{e:fore1}).

  The mean squared forecast error (MSFE) of  the combined  forecast
  is computed as
\begin{equation}\label{msfe_in_2}
MSFE^{}_t=\frac{1}{n_v}\sum_{s=t-n_v}^{t-1} \varepsilon_s^2, \quad t\in (N-E, \,N+1],
\end{equation}
allowing comparison of forecasting performance across methods over the evaluation period.


The package provides  outputs  of:
\begin{itemize}
\item the  combined forecast $\hat x^{(1, ar)}_{t|t-1}$,
\item the forecast  error $\varepsilon_t$,
\item
and the  corresponding  MSFEs $MSFE_t$,
\end{itemize}
for $t\in (N-E,N]$.

In addition, the package reports the out-of-sample forecast $\hat x^{(1, ar)}_{N+1|N}$ for $x_{N+1}$, together with its
  MSFE$_{N+1}$.

In the package, the methodology described through (\ref{e:for6}) is referred to as ``Adapt+AR''.

\subsection{Combined forecasting procedure: Adapt$^2$+AR forecast }\label{e:comb2}
This   method   combines   the
Adapt$^2$ and  AR  forecasting procedures.


First, the two-step adaptive forecasts $Adapt^2_t$   and  the corresponding forecast  errors $u_t$ are  computed over the evaluation period $t\in (N-E,  \, N]$, as  in  (\ref{e:for4}).

Next, for each  $t\in (N-E,  \, N]$, an
 AR($p$), $p=1,2,..., p_{max}$ forecast $u^{(ar)}_{t|t-1}$ of   $u_t$ is obtained in the same way as in (\ref{e:fore1-})–(\ref{e:fore1--}):
\begin{eqnarray*}
u^{(ar)}_{t|t-1}=\hat\phi_{0,t}+\hat\phi_{1,t}u_{t-1}+...+\hat\phi_{p,t}u_{t-p}.
\end{eqnarray*}
This yields  decomposition
\begin{equation}\label{e:fore1+}
u_t=u^{(ar)}_{t|t-1}+\varepsilon_t, \quad t\in (N-E,...,N],
\end{equation}
where $u^{(ar)}_{t|t-1}$ is    the AR forecast  of  $u_t$ and    $\varepsilon_t$  is the corresponding forecast error.

The   combined Adapt$^2$+AR  forecast of  $x_t$  is then   given by
$$\hat x^{(2, ar)}_{t|t-1}=Adapt_t^2+u^{(ar)}_{t|t-1}, \quad  t \in (N-E, N+1].
$$
Accordingly, over the  evaluation period, the data can be decomposed as
\begin{eqnarray}\label{e:for6+}
x_t &=&
\hat x_{t|t-1}^{(2,ar)}+\varepsilon_t,  \quad t\in (N-E,\, N],
\end{eqnarray}
where $\varepsilon_t$ is the forecast error of the combined method, as defined in (\ref{e:fore1+}).

The mean squared forecast error (MSFE) of the combined forecast is computed in the same way as in (\ref{msfe_in_2}) for $t \in (N-E, N+1]$.

The package provides  outputs  of:
\begin{itemize}
\item the combined forecasts $\hat x^{(2, ar)}_{t|t-1}$,
\item the forecast  errors $\varepsilon_t$,
\item and the  corresponding MSFEs  $MSFE_t$,
\end{itemize}
over the evaluation period $t\in (N-E,N]$.

In addition, for $t = N+1$, the package reports the out-of-sample forecast $\hat x^{(2,ar)}_{N+1|N}$ together with its MSFE$_{N+1}$.




In the package, the methodology described in (\ref{e:for6+}) is referred to as ``Adapt$^2$+AR''.


\vskip.3cm
\subsection{Selection of   the optimal  forecast  method }\label{e:comb2+}

The    Adapt +AR and   Adapt$^2$+AR procedures
 for forecasting  $x_{N+1}$,  given the sample  $x_{N}, ..., x_1$,  provide a comprehensive framework for both prediction and model evaluation.

In particular, the package allows for:
\begin{itemize}
  \item forecasting the next observation $x_{N+1}$,

  \item conducting a historical analysis of forecasts over the evaluation period,

     \item
    producing numerical and graphical outputs of forecasts, forecast errors, tuning parameters, and MSFEs over the evaluation period $t \in (N-E, N]$,

     \item extracting and visualising the persistent component of the data, captured by the Adapt$_t$ and Adapt$^2_t$ forecasts,


   \item identifying the forecasting method that yields the lowest MSFE over the evaluation period and for the out-of-sample forecast at $t = N+1$,

  \item
 assessing the adequacy of the selected model by testing for the absence of autocorrelation in the forecast errors (i.e. whether they resemble zero-mean, possibly heteroskedastic, white noise).
\ \end{itemize}
The package permits  selection of the optimal   forecasting  method based on the   smallest  MSFE,
 both over the evaluation period $(N-E, N]$ and for the out-of-sample forecast at $t = N+1$. This selection is further supported by diagnostic checks on the forecast errors, ensuring that they exhibit no significant autocorrelation over the  period  $(N-n_{ar}, N]$.

\vskip.3cm

In summary, the package implements four adaptive forecasting methodologies: Adapt, Adapt$^2$, Adapt+AR, and Adapt$^2$+AR. The parameters  $t_0$, $n_v$, and $n_{ar}$ are preset but can also be specified by the user. When  the sample size is limited, these parameters are automatically adjusted to their maximum feasible values.

The evaluation period of length $E$ is chosen to be as large as possible given the data and parameter settings,
that calculations are feasible adjusting the parameters $t_0$, $n_v$ and $n_{ar}$ when necessary.
The user may visualise either the full evaluation period or selected subperiods, depending on the focus of the analysis.

The function used to compute adaptive forecasts  for the time series $\{x_t\}$ is called \code{forAD}, and is of the form
\begin{Code}
forAD(x, p_max = 3, T0 = 50, n_v = 100, n_ar = 400,
      plots = TRUE, P = 50, PL = 500, p = 1, date_1 = NULL)
\end{Code}

The input arguments are defined as follows.

The 1st argument, \code{x}, specifies the data $x_1,...,x_N$. It can be a univariate numeric time series object (\code{ts}, \code{xts}, \code{zoo}), a numeric vector, or a numeric column of a data frame.

 The 2nd argument, \code{p_max}, determines the maximum order of the AR($p$) model, with $p=1,...,p_{\max}$.
  The default is \code{p_max = 3}, but it can be specified by the user. Setting \code{p_max = 0} disables AR forecasting.


 The 3rd argument, \code{T0}, is the length $t_0$ of the training period used for selecting the tuning parameter. The default value is \code{T0 = 50},  or it can be specified by the user.


 The 4th argument, \code{n_v}, defines the number $n_v$ of past forecast errors used in the computation of the MSFE. The default is \code{n_v = 100},  or it can be specified by the user.


 The 5th argument, \code{n_ar}, determines the number $n_{ar}$ of observations used for estimating the AR($p$) models and computing correlograms. The default value is \code{n_ar = 400},
  or it can be specified by the user.


The 6th argument, \code{plots}, is a logical indicator controlling whether plots are produced (default: \code{TRUE}). When enabled, the following graphical outputs are generated over a period of length $P$:

\begin{enumerate}
\item
 The  MSFEs, $MSFE_t$, for $t\in (N-P, N+1]$, corresponding to the methods:

 adapt, adapt+AR($p$), adapt$^2$, adapt$^2$+AR($p$), for $p = 1, \ldots, p_{\max}$.


\item The tuning parameters $\hat\rho_t$, for $t\in (N-P, N+1]$ for the adapt and  adapt$^2$ methods.

\item Forecast plots over the period $t \in (N-P, N]$, including:
\begin{itemize}
\item adapt forecasts $\hat x_{t|t-1}^{(ad)}$,
\item adapt$^2$ forecasts $\hat x_{t|t-1}^{(ad2)}$,
\item adapt+AR($p$) forecasts $\hat x_{t|t-1}^{(1,ar)}$ for the selected $p$,
\item adapt$^2$+AR($p$) forecasts $\hat x_{t|t-1}^{(2,ar)}$ for the selected $p$,
\item data  $x_t$.
\end{itemize}
These plots are also produced over a longer period of length $PL$.


\item Forecast errors $e_t$, for $t \in (N-P, N]$, for the adapt and adapt$^2$ methods, as well as for adapt+AR($p$) and adapt$^2$+AR($p$) for the selected $p$.



\item Correlograms of the forecast errors $e_t$, for $t \in (N-n_{ar}, N]$, for all methods (adapt, adapt+AR($p$), adapt$^2$, adapt$^2$+AR($p$)), computed up to $\min(10, n_{ar}-1)$ lags and for $p = 1, \ldots, p_{\max}$. These correlograms are based on the package \pkg{testcorr} \citep{testcorr}.

\end{enumerate}

The argument \code{plots} can also be specified as a logical
vector of length 5, allowing the user to select which of the above plots to display. When \code{p_max = 0}, only plots corresponding to the adapt and adapt$^2$ methods are produced.

The 7th argument, \code{P}, determines the length of the period displayed in the plots (default: \code{P = 50},  or it can be specified by the user).

 The 8th argument, \code{PL}, specifies the length of the longer period used in the forecasting plots (default: \code{PL = 500}),  or it can be specified by the user.




  The 9th argument, \code{p}, sets the order of the AR model used in the plots. The default is \code{p = 1}.


The 10th argument, \code{date_1},
defines the time index corresponding to the out-of-sample forecast $t = N+1$. If \code{date_1 = NULL}, the index is set to \code{NROW(x) + 1}. For time series objects (\code{ts}, \code{xts}, \code{zoo}), the user should provide the appropriate date in the corresponding format.

\vskip.2cm

An object of class \class{forAD} is returned as a list with the following components:


\begin{table}[h]
\centering
\begin{tabular}{ll}
\code{for_1} & The one-step ahead  forecast $\hat x_{N+1|N}^{(ad)}$, $\hat x_{N+1|N}^{(1,ar)}$, $\hat x_{N+1|N}^{(ad2)}$, $\hat x_{N+1|N}^{(2,ar)}$\\&
for all  methods (adapt, adapt+AR($p$), adapt$^2$, adapt$^2$+AR($p$), $p=1,\ldots,p_{\max}$). \\[0.2cm]

\code{rho} & The tuning parameters $\hat\rho_{N+1}$ for (adapt, adapt$^2$) methods. \\[0.2cm]
\code{ar_coef_se} & The estimated coefficients $\hat\phi_{0,N},\hat\phi_{1,N},\ldots,\hat\phi_{p,N}$ \\
& and their standard errors for the adapt+AR($p$), $p=1,\ldots,p_{\max}$, methods. \\[0.2cm]

\code{ar_coef_se_sq} & The estimated coefficients $\hat\phi_{0,N},\hat\phi_{1,N},\ldots,\hat\phi_{p,N}$ \\
& and their standard errors for the adapt$^2$+AR($p$), $p=1,\ldots,p_{\max}$, methods. \\[0.2cm]

\code{MSFE} & The MSFE$_{N+1}$ for all methods.
 \\[0.2cm]

\code{for_in} & The  rolling forecasts $\hat x_{t|t-1}^{(ad)}$, $\hat x_{t|t-1}^{(1,ar)}$, $\hat x_{t|t-1}^{(ad2)}$, $\hat x_{t|t-1}^{(2,ar)}$ over the evaluation period \\
&  for all methods. \\[0.2cm]

\code{rho_in} & The rolling  tuning parameters $\hat\rho_t$  over the evaluation period for (adapt, adapt$^2$)  methods. \\[0.2cm]

\code{MSFE_in} & The rolling  MSFEs over the evaluation period for all methods. \\[0.2cm]

\code{err_in} & The rolling one-step ahead forecast errors  over the evaluation period  for all methods.\\[0.2cm]

\code{data} & The original data $x_1, \ldots, x_N$. \\
\end{tabular}
\end{table}




The output of the function \code{forAD} consists  of  a collection of tables and plots presenting the aforementioned results.

The package permits the computation of forecasts for all methods (adapt, adapt$^2$, adapt+AR, adapt$^2$+AR) when the sample size satisfies $N \ge 25 + 2p_{\max}$. When the user sets $p_{\max} = 0$, forecasts (adapt and adapt$^2$) can be computed for $N \ge 21$.

\vfill

\pagebreak

\subsection*{Examples}

We provide some examples to illustrate adaptive forecasting of a univariate time series $\{x_t\}$ using the function \code{forAD}.

\subsubsection*{Example 1}

We simulate a time series $\{x_t\}$ of sample size $N=500$ as $$x_t=2+|\sin(4t/N)|+0.5w_t,$$ where $\{w_t\}$ is an AR(1) series with parameter $\phi_1=-0.5$ and i.i.d. $N$(0,1) innovations.
This  time  series is characterized by a  time-varying mean and   dependent noise arising from  AR(1) process.

We simulate the data for $\{x_t\}$:
%
\begin{CodeChunk}
\begin{CodeInput}
R> N <- 500
R> set.seed(123)
R> w <- arima.sim(list(order = c(1, 0, 0), ar = -0.5), n = N)
R> x <- 2 + abs(sin(4 * seq.int(1, N) / N)) + 0.5 * w
\end{CodeInput}
\end{CodeChunk}
%
We use the function \code{forAD} to evaluate adaptive forecasting for $\{x_t\}$:
%
\begin{CodeChunk}
\begin{CodeInput}
R> library(forecastADAPT)
R> print(forAD(x))
\end{CodeInput}
\end{CodeChunk}

The output of the function \code{forAD} is comprised of 2 sections of tables and a set of 28 plots.


The {\bf 1st section} of tables \code{Adaptive+AR one-step ahead forecast output} contains 3 tables.

 The {\bf  1st table} presents the forecasts, MSFEs, and  relative MSFEs at $N+1$.

 The results for the adapt and adapt+AR methods (for $p=1,2,3$) are as follows:
 \begin{itemize}
 \item  {adapt} forecast  $\hat x_{N+1|N}^{(ad)}=2.62$, MSFE$_{N+1}=0.39$,
  \item  { adapt}+AR(1) forecast  $\hat x_{N+1|N}^{(1,ar)}=2.70$, MSFE$_{N+1}=0.29$,
  \item  { adapt}+AR(2) forecast  $\hat x_{N+1|N}^{(1,ar)}=2.69$, MSFE$_{N+1}=0.30$,
 \item   { adapt}+AR(3) forecast  $\hat x_{N+1|N}^{(1,ar)}=2.68$, MSFE$_{N+1}=0.30$.
 \end{itemize}
 The corresponding results for the { adapt}$^2$ and { adapt}$^2$+AR($p$) are:
 \begin{itemize}
 \item  { adapt}$^2$ forecast  $\hat x_{N+1|N}^{(ad2)}=2.68$, MSFE$_{N+1}=0.40$,
  \item  { adapt}$^2$+AR(1) forecast  $\hat x_{N+1|N}^{(2,ar)}=2.83$, MSFE$_{N+1}=0.31$,
  \item   { adapt}$^2$+AR(2) forecast  $\hat x_{N+1|N}^{(2,ar)}=2.80$, MSFE$_{N+1}=0.32$,
 \item  { adapt}$^2$+AR(3) forecast  $\hat x_{N+1|N}^{(2,ar)}=2.78$, MSFE$_{N+1}=0.32$.
 \end{itemize}
 The relative MSFEs suggest that adapt+AR methods outperform alternatives, with the adapt+AR(1) providing the best forecasting performance.


The {\bf 2nd table} reports the estimated  AR($p$) coefficients ($\hat \phi_{0,N}, ..., \hat \phi_{p,N}$)  and their  standard errors (s.e.) for the   adapt+AR($p$) forecasts $\hat x^{(1, ar)}_{N+1|N}$, for $p=1,2,3$:
\begin{itemize}
 \item  for  $p=1$: $\hat\phi_{0,N}=0.007$  (s.e.=0.026), $\hat\phi_{1,N}=-0.493$  (s.e.=0.043),
  \item  for  $p=2$: $\hat\phi_{0,N}=0.006$  (s.e.=0.026), $\hat\phi_{1,N}=-0.493$  (s.e.=0.050), $\hat\phi_{2,N}=-0.003$  (s.e.=0.050),
    \item  for  $p=3$: $\hat\phi_{0,N}=0.007$  (s.e.=0.026), $\hat\phi_{1,N}=-0.490$  (s.e.=0.050), $\hat\phi_{2,N}=0.021$  (s.e.=0.056), $\hat\phi_{3,N}=0.052$  (s.e.=0.050).
 \end{itemize}
 The stars indicate significance levels $^{***}$ (1\%), $^{**}$ (5\%), and $^*$ (10\%). The results show that only the AR(1) coefficient is statistically significant.
Based on the significance of the coefficients, the adapt+AR(1) is preferred among the adapt$^2$+AR methods.



 Based on the significance of the coefficients, the adapt+AR(1) is preferable among the adapt+AR methods.


The {\bf 3nd table} presents the estimated AR($p$) coefficients ($\hat\phi_{0,N}, ..., \hat\phi_{p,N}$)  and their  standard errors (s.e.) for the   adapt$^2$+AR($p$) forecasts $\hat x^{(2, ar)}_{N+1|N}$, for $p=1,2,3$:
\begin{itemize}
 \item  for  $p=1$: $\hat\phi_{0,N}=-0.031$  (s.e.=0.026), $\hat\phi_{1,N}=-0.515$ (s.e.=0.042),
  \item  for  $p=2$: $\hat\phi_{0,N}=-0.034$ (s.e.=0.026), $\hat\phi_{1,N}=-0.540$  (s.e.=0.050), $\hat\phi_{2,N}=-0.049$ (s.e.=0.050),
    \item  for  $p=3$: $\hat\phi_{0,N}=-0.032$  (s.e.=0.026), $\hat\phi_{1,N}=-0.537$  (s.e.=0.050), $\hat\phi_{2,N}=-0.049$  (s.e.=0.057), $\hat\phi_{3,N}=0.003$  (s.e.=0.050).
 \end{itemize}
  Significance levels are indicated by  $^{***}$ (1\%), $^{**}$ (5\%), and $^*$ (10\%).
The results once again show that only the AR(1) coefficient is statistically significant.
Based on the significance of the coefficients, the adapt$^2$+AR(1) is preferred among the adapt$^2$+AR methods.



\vskip.2cm

The {\bf 2nd section} of tables

\code{Adaptive+AR and adaptive²+AR($p$) one-step ahead forecast output x[N-k+1|N-k]}

contains 3 tables.

The {\bf 1st table} shows the relative MSFEs,
 $MSFE^{}_{t}/MSFE_{t}^{(ad)}$,   for $t=N-9,..., N, N+1 $  and  $p=1,2,3$. The first line consists of 1s, indicating  that all MSFEs are measured relative to the adapt method.
  For all $t=N-9,...,N,N+1$, the relative MSFEs  for the adapt+AR and adapt$^2$+AR methods are below 1,
 while those for adapt$^2$ are slightly above one, suggesting that this method is not preferable. Overall, the relative MSFEs of adaptive+AR(1) tend to be lower than those of the other adapt+AR specifications ($p=2,3$) and adapt$^2$+AR methods ($p=1,2,3$). Hence, adapt+AR(1) method is preferable over the most recent time period.

The {\bf 2nd table} provides the forecasts $\hat x_{t|t-1}^{(ad)}$, $\hat x^{(1,ar)}_{t|t-1}$,
$\hat x_{t|t-1}^{(ad2)}$, $\hat x^{(2,ar)}_{t|t-1}$ for
 $p=1,2,3$ and $t=N-9,...,N,N+1$. The first row includes the observed data $x_t$ for $t=N-9,...,N$ for  comparison.
  Given that the relative MSFEs of the adapt+AR methods are  similar across $p=1,2,3$, it is  expected that the  corresponding forecasts  $\hat x^{(1,ar)}_{t|t-1}$ are also very close across these specifications over the most recent time period.

The {\bf 3rd table} reports the tuning parameters $\hat\rho_t$, $t=N-9,...,N,N+1$ for the adapt and adapt$^2$ methods. The values lie  in  the range  0.91-0.93 for the adapt method and 0.96-0.97 for the adapt$^2$ method during the most recent time period. The values of the tuning parameters of the adapt and adapt$^2$ methods do not differ much, suggesting that the second stage adapt$^2$ provides only limited additional adjustment.



The {\bf plots} present graphically the results of the adaptive forecasting methods.

The first 8 plots in {\bf Figure \ref{ex1_msfe}}
 display  the MSFEs, $MSFE_t$, $t\in (N-P,\,N]$,
  for the methods adapt, adapt$^2$, adapt+AR, adapt$^2$+AR. The solid dot indicates the value at $t=N+1$:
  \begin{itemize}
 \item the blue line and dot (Adapt) correspond to the  MSFEs of the adapt method,
\item  the red line and dot (Adapt$^2$) correspond to  the  MSFEs of adapt$^2$ method,
\item the purple lines and  dots (Adapt+AR($p$)) show the MSFEs for adapt+AR  with $p=1,2,3$,
 \item the green lines and  dots (Adapt$^2$+AR($p$)) show the MSFEs for adapt$^2$+AR  with $p=1,2,3$.
  \end{itemize}


The next 8 plots in {\bf Figure \ref{ex1_for}} present the forecasts together with  the observed data over two plotting horizons: a  shorter  period  $t \in (N-P,N+1]=[451, \, 501]$,
and a  longer period   $t \in (N-PL,\, N+1]=[29, 501]$.
 In all plots, the dashed black line (Data) represents the observed series $x_t$, while the blue line correspond to
adapt, adapt+AR($p$), adapt$^2$, Adapt$^2$+AR($p$)  forecasts (with the selected $p=1$):
 $\hat x_{t|t-1}^{(ad)}$, $\hat x_{t|t-1}^{(1,ar)}$, $\hat x_{t|t-1}^{(ad2)}$, $\hat x_{t|t-1}^{(2,ar)}$.
  The red dot indicates the out-of-sample forecast  at $t=N+1$.

   The adapt forecasts  capture  the non-constant, sinusoidal mean and are very similar to the adapt$^2$ forecasts, suggesting that
   there is no need to apply  the adapt  procedure  twice.
   However, since the data include an AR(1) component with a negative coefficient, incorporating AR(1) dynamics (adapt+AR(1)) leads to a evident improvement in forecasting performance.

These plots reaffirm the conclusion drawn from the tables, indicating  that adapt+AR(1) is the preferred forecasting method.


{\bf Figure \ref{ex1_rho}}  displays the tuning parameters  $\hat\rho_t$, for $t\in (N-P, \, N]$, for   the adapt and adapt$^2$ methods, with the solid dot indicating  the  value at $t=N+1$.
The light blue line (and dot) correspond to the adapt method, while the pink line (and dot) correspond to the adapt$^2$ method.


The next 3 plots in {\bf Figure \ref{ex1_err}} present   the forecast errors for the methods adapt, adapt+AR($p$),  adapt$^2$, adapt$^2$+AR($p$) over the period  $t\in (N-P,\,N]=(451, \, 500]$, using the  selected value $p=1$.
The  first   panel  shows the   adapt forecast  errors $e_t$ (dashed black line) together with  their adapt   forecasts $e^{(ad)}_{t|t-1}$ (blue line). The adaptive forecasts of the errors
$e_t$ are rather close to zero,  indicating  that second  stage of  adapt is not  necessary. The second and third panels  display the  forecast errors for  the  adapt+AR(1), and adapt$^2$+AR(1) methods. The two plots are very similar suggesting that a single stage of adaptive forecasting is sufficient.


The final 8 plots in {\bf Figure \ref{ex1_acf}} show  the correlograms/autocorrelation functions (up to 10 lags) of the forecast errors for the  adapt, adapt+AR($p$), adapt$^2$, adapt$^2$+AR($p$) methods,
for $p=1,2,3$. The forecast errors from the adapt and adapt$^2$ methods
exhibit substantial autocorrelation at lag 1. However, when  adaptive methods are combined with AR forecasting, the autocorrelation becomes negligible,  providing further supporting
evidence that incorporating AR(1) dynamics improves forecasting performance.

Overall, the adaptive forecasting method combined with AR dynamics performs very well, tracking the non-constant sinusoidal mean and the negative autocorrelation structure of the data.


\vfill

\pagebreak

We have the following outputs:
%
\begin{CodeChunk}
\begin{CodeOutput}
-----------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output
-----------------------------------------------------------

Forecasts x[N+1|N], MSFE[N+1|N], RMSFE[N+1|N]

|             | Forecast| MSFE| Relative MSFE|
|:------------|--------:|----:|-------------:|
|Adapt        |     2.62| 0.39|             1|
|Adapt+AR(1)  |     2.70| 0.29|          0.75|
|Adapt+AR(2)  |     2.69| 0.30|          0.75|
|Adapt+AR(3)  |     2.68| 0.30|          0.75|
|------------ | --------| ----| -------------|
|Adapt²       |     2.68| 0.40|          1.02|
|Adapt²+AR(1) |     2.83| 0.31|          0.79|
|Adapt²+AR(2) |     2.80| 0.32|          0.80|
|Adapt²+AR(3) |     2.78| 0.32|          0.81|


AR coefficients and standard errors (adapt+AR)

|      |     |   const|       ar1|     ar2|     ar3|
|:-----|:----|-------:|---------:|-------:|-------:|
|AR(1) |     |   0.007| -0.493***|      NA|      NA|
|      |s.e. | (0.026)|   (0.043)|    (NA)|    (NA)|
|AR(2) |     |   0.006| -0.493***|  -0.003|      NA|
|      |s.e. | (0.026)|   (0.050)| (0.050)|    (NA)|
|AR(3) |     |   0.007| -0.490***|   0.021|   0.052|
|      |s.e. | (0.026)|   (0.050)| (0.056)| (0.050)|

AR coefficients and standard errors (adapt²+AR)

|      |     |   const|       ar1|     ar2|     ar3|
|:-----|:----|-------:|---------:|-------:|-------:|
|AR(1) |     |  -0.031| -0.515***|      NA|      NA|
|      |s.e. | (0.026)|   (0.042)|    (NA)|    (NA)|
|AR(2) |     |  -0.034| -0.540***|  -0.049|      NA|
|      |s.e. | (0.026)|   (0.050)| (0.050)|    (NA)|
|AR(3) |     |  -0.032| -0.537***|  -0.049|   0.003|
|      |s.e. | (0.026)|   (0.050)| (0.057)| (0.050)|


------------------------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output x[N-k+1|N-k]
------------------------------------------------------------------------

Relative mean squared forecast errors RMSFE[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt        |    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|
|Adapt+AR(1)  | 0.74| 0.74| 0.74| 0.73| 0.75| 0.74| 0.73| 0.74| 0.75| 0.76| 0.75|
|Adapt+AR(2)  | 0.74| 0.74| 0.74| 0.75| 0.74| 0.76| 0.73| 0.75| 0.75| 0.75| 0.75|
|Adapt+AR(3)  | 0.74| 0.73| 0.73| 0.74| 0.74| 0.76| 0.74| 0.74| 0.76| 0.75| 0.75|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 1.01| 1.00| 1.00| 1.03| 1.02| 1.04| 1.01| 1.01| 1.04| 1.04| 1.02|
|Adapt²+AR(1) | 0.77| 0.76| 0.76| 0.78| 0.78| 0.79| 0.77| 0.78| 0.80| 0.80| 0.79|
|Adapt²+AR(2) | 0.79| 0.79| 0.79| 0.79| 0.79| 0.81| 0.80| 0.80| 0.81| 0.81| 0.80|
|Adapt²+AR(3) | 0.78| 0.79| 0.80| 0.78| 0.79| 0.80| 0.80| 0.81| 0.81| 0.81| 0.81|


Forecasts x[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Data         | 2.32| 2.40| 3.38| 2.76| 1.95| 3.07| 2.12| 2.01| 3.18| 2.49|     |
|-----------  | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt        | 2.69| 2.66| 2.64| 2.69| 2.70| 2.64| 2.68| 2.63| 2.58| 2.63| 2.62|
|Adapt+AR(1)  | 2.61| 2.85| 2.77| 2.34| 2.68| 3.02| 2.48| 2.92| 2.89| 2.34| 2.70|
|Adapt+AR(2)  | 2.61| 2.85| 2.77| 2.34| 2.68| 3.02| 2.48| 2.92| 2.89| 2.34| 2.69|
|Adapt+AR(3)  | 2.61| 2.87| 2.77| 2.31| 2.68| 3.06| 2.46| 2.89| 2.90| 2.30| 2.68|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 2.80| 2.75| 2.72| 2.80| 2.80| 2.72| 2.76| 2.70| 2.62| 2.70| 2.68|
|Adapt²+AR(1) | 2.65| 3.15| 3.00| 2.07| 2.76| 3.50| 2.35| 3.30| 3.25| 2.08| 2.83|
|Adapt²+AR(2) | 2.64| 3.16| 3.03| 2.07| 2.74| 3.52| 2.38| 3.29| 3.30| 2.10| 2.80|
|Adapt²+AR(3) | 2.65| 3.18| 3.03| 2.04| 2.74| 3.56| 2.38| 3.27| 3.30| 2.05| 2.78|


Tuning parameters rho[N-k+1|N-k]

|       | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt  | 0.92| 0.92| 0.92| 0.92| 0.92| 0.93| 0.92| 0.92| 0.91| 0.92| 0.91|
|Adapt² | 0.96| 0.96| 0.96| 0.97| 0.97| 0.97| 0.96| 0.97| 0.96| 0.97| 0.96|
\end{CodeOutput}
\end{CodeChunk}

\pagebreak

\captionsetup{belowskip=7pt}
\captionsetup[subfloat]{farskip=10pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selection of the best forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_01.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_02.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_03.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_06.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_04.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_07.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_05.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_08.pdf}}
\caption{Comparison of the mean squared forecasts errors (MSFE) for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex1_msfe}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Forecasting results} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_10.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_14.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_11.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_15.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_12.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_16.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_13.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_17.pdf}}
\caption{Data and their forecasts for the forecasting methods Adapt, Adapt+AR(1), Adapt$^2$, Adapt$^2$+AR(1).}
\label{ex1_for}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selected tuning parameter} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_09.pdf}}
\caption{Selected tuning parameter for the forecasting methods Adapt, Adapt$^2$.}
\label{ex1_rho}
\end{figure}

\bigskip\bigskip

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Evaluation of forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_18.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_19.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex1_20.pdf}}
\caption{Forecasts errors for the forecasting methods Adapt+AR(1), Adapt$^2$+AR(1).}
\label{ex1_err}
\end{figure}

\clearpage

\captionsetup{belowskip=-2pt}
\captionsetup[subfloat]{farskip=18pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Testing for correlation in forecast errors} \vspace{6pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex1_21.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex1_25.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex1_22.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex1_26.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex1_23.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex1_27.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex1_24.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex1_28.pdf}}
\caption{Correlogram of forecasts errors for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex1_acf}
\end{figure}

\clearpage
\pagebreak

\subsubsection*{Example 2}

We simulate a time series $\{x_t\}$ of sample size $N=500$ as $$x_t=I(t>0.5N)+I(t>0.97N)+0.3u_t,$$ where $I(\cdot)$ is the indicator function and $\{u_t\}$ is an i.i.d. $N$(0,1) series.

This time  series  exhibits  two structural breaks,
 in the  mean, occurring at $t=0.5N$ and  $t=0.97N$, while  otherwise the observations are independent.

We simulate the data for $\{x_t\}$:
%
\begin{CodeChunk}
\begin{CodeInput}
R> N <- 500
R> set.seed(123)
R> u <- rnorm(N)
R> x <- c(rep(0, 0.5 * N), rep(1, 0.47 * N), rep(2, 0.03 * N)) + 0.3 * u
\end{CodeInput}
\end{CodeChunk}
%
We use the function \code{forAD} to evaluate adaptive forecasting for $\{x_t\}$.
%
\begin{CodeChunk}
\begin{CodeInput}
R> library(forecastADAPT)
R> print(forAD(x))
\end{CodeInput}
\end{CodeChunk}
We use this dataset to illustrate the performance of adaptive forecasting methods in the presence of abrupt changes in the mean of a series.

From the tables, we observe that the  relative MSFEs (RMSFE) for the adapt+AR, adapt$^2$, adapt$^2$+AR forecasts are generally around 1.
In some cases they are slightly  above 1, and in others slightly  below 1. From the plots, see Figure \ref{ex2_msfe}, it is  clear that the MSFEs of the various methods are not substantially different from each other.

An exception occurs shortly after the break, where the adapt$^2$ method exhibits a lower MSFE.
Nevertheless, the simple adapt method quickly adjusts, and after a few periods its MSFE becomes similar to that of the adapt$^2$ method.
This indicates  that  adapt forecasting suffices for  this series.

This observation is also evident from the tables and the plots, see Figure \ref{ex2_for}, where the forecast values are very close across the methods, except immediately after the breaks where adapt$^2$ performs slightly better. The impact of the break on the adapt and adapt$^2$ forecasts is  further reflected in the selected tuning parameters, see Figure \ref{ex2_rho}, which become unstable after the break points. Moreover,  applying adapt to the adapt forecast errors has no effect before the break (see top plot in Figure \ref{ex2_err}), but does show
some adjustment immediately after the break, highlighting why adapt$^2$ briefly outperforms adapt.


The correlograms, see Figure \ref{ex2_acf}, suggest that there is no autocorrelation in the forecast errors for any method. This confirms that  AR forecasting is not needed, which is also supported by the AR coefficient tables showing no significant parameters. User can set \code{p_max = 0}, i.e. run \code{print(forAD(x, p_max = 0))}, to view only the  adapt and adapt$^2$ outputs.

Overall, the adaptive forecasting performs very well, tracking the non-constant breaking mean and the lack of autocorrelation of the data. Around the break points, the forecasts take few periods to adjust, but
considering that the method does not use any break-point detection, the performance is highly satisfactory.

\vfill

\pagebreak

We have the following outputs:
%
\begin{CodeChunk}
\begin{CodeOutput}
-----------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output
-----------------------------------------------------------

Forecasts x[N+1|N], MSFE[N+1|N], RMSFE[N+1|N]

|             | Forecast| MSFE| Relative MSFE|
|:------------|--------:|----:|-------------:|
|Adapt        |     2.06| 0.12|             1|
|Adapt+AR(1)  |     2.08| 0.12|          0.99|
|Adapt+AR(2)  |     2.08| 0.12|          0.99|
|Adapt+AR(3)  |     2.08| 0.12|          1.00|
|------------ | --------| ----| -------------|
|Adapt²       |     2.18| 0.12|          1.00|
|Adapt²+AR(1) |     2.21| 0.12|          1.01|
|Adapt²+AR(2) |     2.21| 0.12|          1.02|
|Adapt²+AR(3) |     2.21| 0.13|          1.03|


AR coefficients and standard errors (adapt+AR)

|      |     |   const|     ar1|     ar2|     ar3|
|:-----|:----|-------:|-------:|-------:|-------:|
|AR(1) |     |   0.017|   0.023|      NA|      NA|
|      |s.e. | (0.016)| (0.050)|    (NA)|    (NA)|
|AR(2) |     |   0.016|   0.023|   0.018|      NA|
|      |s.e. | (0.016)| (0.050)| (0.050)|    (NA)|
|AR(3) |     |   0.017|   0.023|   0.018|   0.001|
|      |s.e. | (0.016)| (0.050)| (0.050)| (0.050)|

AR coefficients and standard errors (adapt²+AR)

|      |     |   const|     ar1|     ar2|     ar3|
|:-----|:----|-------:|-------:|-------:|-------:|
|AR(1) |     |   0.003|   0.011|      NA|      NA|
|      |s.e. | (0.016)| (0.050)|    (NA)|    (NA)|
|AR(2) |     |   0.003|   0.011|  -0.012|      NA|
|      |s.e. | (0.016)| (0.050)| (0.050)|    (NA)|
|AR(3) |     |   0.003|   0.011|  -0.013|  -0.044|
|      |s.e. | (0.017)| (0.050)| (0.050)| (0.050)|


------------------------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output x[N-k+1|N-k]
------------------------------------------------------------------------

Relative mean squared forecast errors RMSFE[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt        |    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|
|Adapt+AR(1)  | 0.99| 1.01| 0.97| 0.98| 1.01| 0.99| 0.99| 0.99| 0.98| 0.98| 0.99|
|Adapt+AR(2)  | 0.99| 0.99| 0.98| 0.97| 1.01| 1.00| 1.01| 1.01| 0.98| 0.98| 0.99|
|Adapt+AR(3)  | 1.02| 0.99| 0.99| 0.98| 1.00| 1.01| 1.00| 1.02| 1.00| 1.00| 1.00|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 0.97| 0.97| 0.95| 0.96| 1.00| 1.01| 1.01| 1.01| 1.02| 0.99| 1.00|
|Adapt²+AR(1) | 0.97| 0.97| 0.96| 0.96| 1.00| 1.02| 1.01| 1.03| 1.03| 1.00| 1.01|
|Adapt²+AR(2) | 1.00| 0.99| 0.96| 0.95| 1.01| 1.04| 1.03| 1.02| 1.01| 1.01| 1.02|
|Adapt²+AR(3) | 1.01| 1.00| 0.97| 0.95| 1.03| 1.05| 1.04| 1.03| 1.02| 1.03| 1.03|


Forecasts x[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Data         | 1.96| 2.42| 2.38| 1.67| 1.73| 1.59| 2.05| 2.04| 2.10| 2.16|     |
|-----------  | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt        | 1.97| 1.97| 2.14| 2.23| 2.03| 1.92| 1.80| 1.89| 1.94| 2.00| 2.06|
|Adapt+AR(1)  | 1.99| 1.99| 2.17| 2.26| 2.03| 1.93| 1.80| 1.91| 1.96| 2.02| 2.08|
|Adapt+AR(2)  | 2.01| 1.99| 2.17| 2.27| 2.04| 1.93| 1.80| 1.91| 1.97| 2.03| 2.08|
|Adapt+AR(3)  | 2.01| 1.98| 2.16| 2.27| 2.03| 1.93| 1.80| 1.91| 1.97| 2.03| 2.08|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 2.31| 2.24| 2.44| 2.53| 2.21| 2.07| 1.91| 2.01| 2.06| 2.13| 2.18|
|Adapt²+AR(1) | 2.34| 2.26| 2.47| 2.56| 2.22| 2.07| 1.91| 2.03| 2.09| 2.15| 2.21|
|Adapt²+AR(2) | 2.36| 2.27| 2.48| 2.56| 2.22| 2.09| 1.91| 2.04| 2.09| 2.16| 2.21|
|Adapt²+AR(3) | 2.34| 2.23| 2.51| 2.58| 2.21| 2.09| 1.95| 2.06| 2.11| 2.15| 2.21|


Tuning parameters rho[N-k+1|N-k]

|       | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt  | 0.66| 0.65| 0.63| 0.62| 0.65| 0.65| 0.63| 0.65| 0.65| 0.63| 0.64|
|Adapt² | 0.88| 0.91| 0.90| 0.90| 0.94| 0.95| 0.96| 0.96| 0.96| 0.95| 0.95|
\end{CodeOutput}
\end{CodeChunk}

\pagebreak

\captionsetup{belowskip=7pt}
\captionsetup[subfloat]{farskip=10pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selection of the best forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_01.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_02.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_03.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_06.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_04.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_07.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_05.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_08.pdf}}
\caption{Comparison of the mean squared forecasts errors (MSFE) for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex2_msfe}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Forecasting results} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_10.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_14.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_11.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_15.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_12.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_16.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_13.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_17.pdf}}
\caption{Data and their forecasts for the forecasting methods Adapt, Adapt+AR(1), Adapt$^2$, Adapt$^2$+AR(1).}
\label{ex2_for}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selected tuning parameter} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_09.pdf}}
\caption{Selected tuning parameter for the forecasting methods Adapt, Adapt$^2$.}
\label{ex2_rho}
\end{figure}

\bigskip\bigskip

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Evaluation of forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_18.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_19.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex2_20.pdf}}
\caption{Forecasts errors for the forecasting methods Adapt+AR(1), Adapt$^2$+AR(1).}
\label{ex2_err}
\end{figure}

\clearpage

\captionsetup{belowskip=-2pt}
\captionsetup[subfloat]{farskip=18pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Testing for correlation in forecast errors} \vspace{6pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex2_21.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex2_25.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex2_22.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex2_26.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex2_23.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex2_27.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex2_24.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex2_28.pdf}}
\caption{Correlogram of forecasts errors for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex2_acf}
\end{figure}

\clearpage
\pagebreak

\subsubsection*{Example 3}

We simulate a time series $\{x_t\}$ of sample size $N=500$ as $$x_t=0.1t+\sum_{s=1}^tu_s+w_t,$$ where $\{u_t\}$ is an i.i.d. $N$(0,1) sequence, $\{w_t\}$ is an AR(1) process with parameter $\phi_1=0.5$ and  i.i.d. $N$(0,1) innovations, and $\{u_t\},\{w_t\}$ are independent.
This  series  combines a linear trend, a unit root component,  and  a  stationary AR(1)  process.


We simulate the data for $\{x_t\}$:
%
\begin{CodeChunk}
\begin{CodeInput}
R> N <- 500
R> set.seed(1234)
R> u <- rnorm(N)
R> set.seed(123)
R> w <- arima.sim(list(order = c(1, 0, 0), ar = 0.5), n = N)
R> x <- 0.1 * seq.int(1, N) + cumsum(u) + w
\end{CodeInput}
\end{CodeChunk}
%
We use the function \code{forAD} to evaluate adaptive forecasting for $\{x_t\}$.
%
\begin{CodeChunk}
\begin{CodeInput}
R> library(forecastADAPT)
R> print(forAD(x))
\end{CodeInput}
\end{CodeChunk}

From the tables, we observe that the relative MSFEs for the adapt+AR, adapt$^2$, adapt$^2$+AR forecasts are generally around 1, in some cases slightly above or below. From the plots, see Figure \ref{ex3_msfe}, it is also clear that the MSFEs of the various methods are very similar. These results suggest that adapt forecasting alone suffices. This is further supported by the forecast  tables and plots, see Figure \ref{ex3_for}, where the forecasts across methods are nearly identical.

Moreover, the adapt$^2$ method is unnecessary. The plot of the selected tuning parameters, see Figure \ref{ex3_rho}, shows that the adapt$^2$ parameters are very close to 1, indicating that the second-level adapt is effectively just the mean. This is further confirmed when adapt is applied to the adapt forecast errors, see top plot in Figure \ref{ex3_err}.




The correlograms, see Figure \ref{ex3_acf}, indicate negligible autocorrelation in the forecast errors for  all methods,  confirming that the AR forecasting is not required.
A minor exception is a small autocorrelation at lag 2 in the adapt and adapt$^2$ forecast errors, corresponding to occasional significance of the AR(2) parameter. However, since the forecasts across methods are largely unchanged, this minor autocorrelation can be safely ignored.

 The user can also try setting \code{p_max = 0}, i.e. running \code{print(forAD(x, p_max = 0))},  to generate outputs only for the adapt and adapt$^2$ methods.

Overall, the adaptive forecasting performs very well, tracking the non-constant trending mean, the stochastic trend (unit root), and the positive of autocorrelation present in the data.
Although the data contains an AR(1) component with positive autocorrelation, the results suggest that combining adaptive forecasting with  AR methods is unnecessary, since the adaptive method adequately tracks both the deterministic trend,  stochastic trend and the autocorrelation structure.


\vfill

\pagebreak

We have the following outputs:
%
\begin{CodeChunk}
\begin{CodeOutput}
-----------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output
-----------------------------------------------------------

Forecasts x[N+1|N], MSFE[N+1|N], RMSFE[N+1|N]

|             | Forecast| MSFE| Relative MSFE|
|:------------|--------:|----:|-------------:|
|Adapt        |     49.6| 2.81|             1|
|Adapt+AR(1)  |     49.8| 2.81|          1.00|
|Adapt+AR(2)  |     49.5| 2.75|          0.97|
|Adapt+AR(3)  |     49.5| 2.78|          0.98|
|------------ | --------| ----| -------------|
|Adapt²       |     49.8| 2.83|          1.00|
|Adapt²+AR(1) |     49.9| 2.85|          1.01|
|Adapt²+AR(2) |     49.3| 2.78|          0.99|
|Adapt²+AR(3) |     49.4| 2.81|          1.00|


AR coefficients and standard errors (adapt+AR)

|      |     |   const|     ar1|       ar2|     ar3|
|:-----|:----|-------:|-------:|---------:|-------:|
|AR(1) |     | 0.170**|  -0.009|        NA|      NA|
|      |s.e. | (0.080)| (0.050)|      (NA)|    (NA)|
|AR(2) |     | 0.190**|  -0.011|  -0.129**|      NA|
|      |s.e. | (0.080)| (0.050)|   (0.050)|    (NA)|
|AR(3) |     | 0.197**|  -0.012| -0.129***|  -0.013|
|      |s.e. | (0.081)| (0.050)|   (0.050)| (0.050)|

AR coefficients and standard errors (adapt²+AR)

|      |     |  const|    ar1|      ar2|    ar3|
|:-----|:----|------:|------:|--------:|------:|
|AR(1) |     |   0.01|  -0.01|       NA|     NA|
|      |s.e. | (0.08)| (0.05)|     (NA)|   (NA)|
|AR(2) |     |   0.01|  -0.01| -0.12***|     NA|
|      |s.e. | (0.08)| (0.05)|   (0.05)|   (NA)|
|AR(3) |     |   0.02|  -0.01| -0.13***|  -0.01|
|      |s.e. | (0.08)| (0.05)|   (0.05)| (0.05)|


------------------------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output x[N-k+1|N-k]
------------------------------------------------------------------------

Relative mean squared forecast errors RMSFE[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt        |    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|
|Adapt+AR(1)  | 1.00| 1.00| 0.95| 0.98| 1.03| 1.00| 1.01| 1.02| 1.01| 1.00| 1.00|
|Adapt+AR(2)  | 0.99| 0.99| 0.92| 0.96| 0.97| 0.98| 0.97| 1.03| 0.97| 1.00| 0.97|
|Adapt+AR(3)  | 1.04| 0.98| 0.94| 0.99| 0.99| 0.98| 0.99| 1.00| 0.98| 0.99| 0.98|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 1.00| 1.01| 0.96| 1.04| 0.98| 1.01| 1.03| 1.02| 1.05| 1.01| 1.00|
|Adapt²+AR(1) | 1.03| 1.00| 0.95| 1.00| 0.98| 1.03| 1.01| 1.01| 1.02| 1.01| 1.01|
|Adapt²+AR(2) | 1.04| 0.98| 0.94| 0.99| 0.98| 1.04| 0.99| 0.99| 1.00| 1.01| 0.99|
|Adapt²+AR(3) | 1.01| 0.99| 0.96| 0.98| 0.97| 1.01| 0.99| 1.02| 0.99| 1.00| 1.00|


Forecasts x[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Data         | 46.6| 44.4| 48.4| 47.2| 44.2| 44.7| 46.4| 45.5| 47.9| 50.2|     |
|-----------  | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt        | 46.2| 46.6| 44.8| 47.5| 47.3| 45.0| 44.8| 46.0| 45.7| 47.3| 49.6|
|Adapt+AR(1)  | 46.4| 46.8| 44.9| 47.6| 47.5| 45.2| 45.0| 46.1| 45.8| 47.4| 49.8|
|Adapt+AR(2)  | 46.3| 46.5| 44.9| 47.9| 47.1| 45.3| 45.4| 46.2| 45.6| 47.5| 49.5|
|Adapt+AR(3)  | 46.3| 46.5| 44.8| 47.9| 47.1| 45.2| 45.4| 46.3| 45.6| 47.5| 49.5|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 46.3| 46.7| 44.9| 47.6| 47.4| 45.2| 44.9| 46.1| 45.8| 47.4| 49.8|
|Adapt²+AR(1) | 46.5| 46.9| 45.1| 47.7| 47.6| 45.4| 45.1| 46.3| 46.0| 47.5| 49.9|
|Adapt²+AR(2) | 46.3| 46.4| 45.0| 48.3| 46.8| 45.5| 45.9| 46.4| 45.6| 47.7| 49.3|
|Adapt²+AR(3) | 46.3| 46.3| 44.9| 48.3| 47.0| 45.4| 45.9| 46.5| 45.6| 47.6| 49.4|


Tuning parameters rho[N-k+1|N-k]

|       | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt  | 0.17| 0.08| 0.15| 0.24| 0.25| 0.26| 0.27| 0.25| 0.25| 0.25| 0.20|
|Adapt² | 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99|
\end{CodeOutput}
\end{CodeChunk}

\pagebreak

\captionsetup{belowskip=7pt}
\captionsetup[subfloat]{farskip=10pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selection of the best forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_01.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_02.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_03.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_06.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_04.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_07.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_05.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_08.pdf}}
\caption{Comparison of the mean squared forecasts errors (MSFE) for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex3_msfe}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Forecasting results} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_10.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_14.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_11.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_15.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_12.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_16.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_13.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_17.pdf}}
\caption{Data and their forecasts for the forecasting methods Adapt, Adapt+AR(1), Adapt$^2$, Adapt$^2$+AR(1).}
\label{ex3_for}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selected tuning parameter} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_09.pdf}}
\caption{Selected tuning parameter for the forecasting methods Adapt, Adapt$^2$.}
\label{ex3_rho}
\end{figure}

\bigskip\bigskip

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Evaluation of forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_18.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_19.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex3_20.pdf}}
\caption{Forecasts errors for the forecasting methods Adapt+AR(1), Adapt$^2$+AR(1).}
\label{ex3_err}
\end{figure}

\clearpage

\captionsetup{belowskip=-2pt}
\captionsetup[subfloat]{farskip=18pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Testing for correlation in forecast errors} \vspace{6pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex3_21.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex3_25.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex3_22.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex3_26.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex3_23.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex3_27.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex3_24.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex3_28.pdf}}
\caption{Correlogram of forecasts errors for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex3_acf}
\end{figure}

\clearpage
\pagebreak

\subsubsection*{Empirical example 1}

We use the time series of the US monthly year-on-year inflation rate $\{INFL_t\}$ (in \%), seasonally adjusted, for the period January 1950 -- December 2024 with a sample size $N=900$ (retrieved on 15/08/2025). We employ the package \pkg{quantmod} \citep{quantmod} to retrieve the data from FRED. This package depends on  \pkg{xts} package \citep{xts} for handling time series data, which we also use to construct lagged variables.

\begin{CodeChunk}
\begin{CodeInput}
R> library(quantmod)
R> getSymbols("CPIAUCSL", src = "FRED", return.class = "xts")
R> INFL <- 100 * (CPIAUCSL - lag(CPIAUCSL, k = 12)) / lag(CPIAUCSL, k = 12)
R> INFL <- window(INFL, start = "1950-01-01", end = "2024-12-01")
\end{CodeInput}
\end{CodeChunk}
%
We use the function \code{forAD} to evaluate adaptive forecasting for $\{INFL_t\}$.
%
\begin{CodeChunk}
\begin{CodeInput}
R> library(forecastADAPT)
R> print(forAD(INFL, date_1 = "2025-01-01"))
\end{CodeInput}
\end{CodeChunk}

From the tables, we see that the relative MSFE of the adapt+AR methods is well below 1.
From the MSFE plots (Figure \ref{ex4_msfe}) it is clear that adapt+AR outperforms the other methods. Among these, adapt+AR(2) performs marginally better toward the end of the sample compared to adapt+AR(1) and adapt+AR(3). The table of AR coefficients further suggests the use of adapt+AR(2). Nevertheless, the differences among the adapt+AR forecasts are marginal. Based on the adapt+AR(2) method, the forecast of inflation for January 2025 is 2.92\%, compared to the true value of 2.99\%.


The correlograms (Figure \ref{ex4_acf}) indicate almost no autocorrelation in the forecast errors of adapt+AR(2), while the  forecast errors of  the adapt method exhibit strong autocorrelation at lag 1, confirming the need for AR adjustment. The user may set \code{p = 2}, i.e. run \code{forAD(INFL, p = 2, date_1 = "2025-01-01")}, to visualise the forecasts from the adapt+AR(2) and adapt$^2$+AR(2) methods.



A second round of adaptive forecasting is not required, as seen from the MSFE table and plots (Figure \ref{ex4_msfe}). This is further supported by the forecast plots (Figure \ref{ex4_for}), where adapt$^2$ appears more volatile and does not improve upon the adapt forecasts. The tuning parameters for the adapt method (Figure \ref{ex4_rho}) reach the lower bound of the optimisation,
hence, the adapt forecast is the last observed value.  For the adapt$^2$ method, there is some deviation of the tuning parameter from the lower bound in the final two years.
However, this does not seem to improve  the forecast, as the second stage of adapt forecasting applied to the forecast errors (Figure \ref{ex4_err}) gives values very close to zero in the period of the last two years.

Overall, the adaptive forecasting method performs very well, tracking  both the non-constant mean and the autocorrelation structure of the data.

\vfill

\pagebreak

We have the following outputs:
%
\begin{CodeChunk}
\begin{CodeOutput}
-----------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output
-----------------------------------------------------------

Forecasts x[N+1|N], MSFE[N+1|N], RMSFE[N+1|N]

|             | Forecast| MSFE| Relative MSFE|
|:------------|--------:|----:|-------------:|
|Adapt        |     2.87| 0.16|             1|
|Adapt+AR(1)  |     2.93| 0.12|          0.75|
|Adapt+AR(2)  |     2.92| 0.12|          0.74|
|Adapt+AR(3)  |     2.92| 0.12|          0.74|
|------------ | --------| ----| -------------|
|Adapt²       |     2.93| 0.17|          1.01|
|Adapt²+AR(1) |     3.03| 0.17|          1.02|
|Adapt²+AR(2) |     2.95| 0.15|          0.90|
|Adapt²+AR(3) |     2.93| 0.15|          0.93|


AR coefficients and standard errors (adapt+AR)

|      |     |    const|       ar1|        ar2|      ar3|
|:-----|:----|--------:|---------:|----------:|--------:|
|AR(1) |     |  -0.0003| 0.4088***|         NA|       NA|
|      |s.e. | (0.0186)|  (0.0457)|       (NA)|     (NA)|
|AR(2) |     |   0.0001| 0.4891***| -0.2017***|       NA|
|      |s.e. | (0.0183)|  (0.0492)|   (0.0491)|     (NA)|
|AR(3) |     |  -0.0006| 0.5077***| -0.2386***|   0.0776|
|      |s.e. | (0.0182)|  (0.0502)|   (0.0549)| (0.0501)|

AR coefficients and standard errors (adapt²+AR)

|      |     |   const|      ar1|       ar2|      ar3|
|:-----|:----|-------:|--------:|---------:|--------:|
|AR(1) |     |   0.009|  0.127**|        NA|       NA|
|      |s.e. | (0.021)|  (0.049)|      (NA)|     (NA)|
|AR(2) |     |   0.012| 0.167***| -0.310***|       NA|
|      |s.e. | (0.020)|  (0.047)|   (0.047)|     (NA)|
|AR(3) |     |   0.011| 0.135***| -0.292***| -0.108**|
|      |s.e. | (0.020)|  (0.049)|   (0.048)|  (0.049)|


------------------------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output x[N-k+1|N-k]
------------------------------------------------------------------------

Relative mean squared forecast errors RMSFE[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt        |    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|
|Adapt+AR(1)  | 0.77| 0.79| 0.79| 0.77| 0.74| 0.75| 0.75| 0.73| 0.76| 0.75| 0.75|
|Adapt+AR(2)  | 0.77| 0.79| 0.79| 0.76| 0.75| 0.75| 0.75| 0.73| 0.75| 0.75| 0.74|
|Adapt+AR(3)  | 0.77| 0.79| 0.78| 0.75| 0.75| 0.74| 0.73| 0.74| 0.74| 0.74| 0.74|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 1.02| 1.05| 1.05| 1.00| 1.01| 0.98| 1.01| 1.00| 0.99| 1.01| 1.01|
|Adapt²+AR(1) | 1.04| 1.07| 1.07| 1.05| 1.02| 1.00| 1.02| 1.02| 1.02| 1.02| 1.02|
|Adapt²+AR(2) | 0.93| 0.95| 0.93| 0.93| 0.90| 0.88| 0.91| 0.88| 0.91| 0.90| 0.90|
|Adapt²+AR(3) | 0.97| 0.97| 0.95| 0.95| 0.93| 0.92| 0.93| 0.90| 0.93| 0.93| 0.93|


Forecasts x[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Data         | 3.46| 3.35| 3.23| 2.97| 2.93| 2.61| 2.43| 2.57| 2.71| 2.87|     |
|-----------  | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt        | 3.16| 3.46| 3.35| 3.23| 2.97| 2.93| 2.61| 2.43| 2.57| 2.71| 2.87|
|Adapt+AR(1)  | 3.18| 3.58| 3.30| 3.18| 2.85| 2.92| 2.47| 2.35| 2.62| 2.77| 2.93|
|Adapt+AR(2)  | 3.23| 3.60| 3.23| 3.20| 2.85| 2.97| 2.45| 2.40| 2.67| 2.75| 2.92|
|Adapt+AR(3)  | 3.25| 3.59| 3.22| 3.22| 2.84| 2.97| 2.42| 2.41| 2.65| 2.73| 2.92|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 3.12| 3.55| 3.36| 3.20| 2.86| 2.85| 2.42| 2.25| 2.50| 2.72| 2.93|
|Adapt²+AR(1) | 3.17| 3.72| 3.29| 3.14| 2.73| 2.85| 2.26| 2.18| 2.60| 2.81| 3.03|
|Adapt²+AR(2) | 3.27| 3.71| 3.11| 3.22| 2.76| 2.98| 2.21| 2.31| 2.66| 2.71| 2.95|
|Adapt²+AR(3) | 3.25| 3.71| 3.10| 3.20| 2.78| 2.99| 2.23| 2.31| 2.67| 2.69| 2.93|


Tuning parameters rho[N-k+1|N-k]

|       |  k=10|   k=9|   k=8|   k=7|   k=6|   k=5|   k=4|   k=3|   k=2|   k=1|   k=0|
|:------|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|
|Adapt  | 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001|
|Adapt² |  0.63|  0.62|  0.61|  0.64|  0.72|  0.67|  0.61|  0.63|  0.65|  0.65|  0.63|
\end{CodeOutput}
\end{CodeChunk}

\pagebreak

\captionsetup{belowskip=7pt}
\captionsetup[subfloat]{farskip=10pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selection of the best forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_01.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_02.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_03.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_06.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_04.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_07.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_05.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_08.pdf}}
\caption{Comparison of the mean squared forecasts errors (MSFE) for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex4_msfe}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Forecasting results} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_10.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_14.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_11.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_15.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_12.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_16.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_13.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_17.pdf}}
\caption{Data and their forecasts for the forecasting methods Adapt, Adapt+AR(1), Adapt$^2$, Adapt$^2$+AR(1).}
\label{ex4_for}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selected tuning parameter} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_09.pdf}}
\caption{Selected tuning parameter for the forecasting methods Adapt, Adapt$^2$.}
\label{ex4_rho}
\end{figure}

\bigskip\bigskip

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Evaluation of forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_18.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_19.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex4_20.pdf}}
\caption{Forecasts errors for the forecasting methods Adapt+AR(1), Adapt$^2$+AR(1).}
\label{ex4_err}
\end{figure}

\clearpage

\captionsetup{belowskip=-2pt}
\captionsetup[subfloat]{farskip=18pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Testing for correlation in forecast errors} \vspace{6pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex4_21.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex4_25.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex4_22.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex4_26.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex4_23.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex4_27.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex4_24.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex4_28.pdf}}
\caption{Correlogram of forecasts errors for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex4_acf}
\end{figure}

\clearpage
\pagebreak

\subsubsection*{Empirical example 2}

We use the time series of the US monthly unemployment rate $\{UNRATE_t\}$ (in \%), seasonally adjusted, over the period January 1950 -- December 2024, with  sample size $N=900$ (retrieved on 01/09/2025). Note that, due to the COVID-19 pandemic, the time series exhibits a sharp spike in April 2020, which took more than a year to revert back to pre-pandemic levels. For exposition, we do not remove any observations. To see the effect of this  episode on the adaptive forecasting, we present plots for the  most recent 6 years, that is we set \code{P = 72}.
The data are retrieved from FRED using the \pkg{quantmod} package \citep{quantmod}, which depends on the \pkg{xts} package \citep{xts} for handling time series data.
\begin{CodeChunk}
\begin{CodeInput}
R> library(quantmod)
R> getSymbols("UNRATE", src = "FRED", return.class = "xts")
R> UNRATE <- window(UNRATE, start = "1950-01-01", end = "2024-12-01")
\end{CodeInput}
\end{CodeChunk}
%
We use the function \code{forAD} to evaluate adaptive forecasting for $\{UNRATE_t\}$.
%
\begin{CodeChunk}
\begin{CodeInput}
R> library(forecastADAPT)
R> print(forAD(UNRATE, P = 72, date_1 = "2025-01-01"))
\end{CodeInput}
\end{CodeChunk}

From the tables, we observe that the relative MSFE of the adapt+AR, adapt$^2$, adapt$^2$+AR forecasts are well above 1.
From the MSFEs plots  (Figure \ref{ex5_msfe}), it is also clear that the  adapt method yields the lowest forecast errors.

 This result is driven by the extreme values of  the unemployment rate during 2020.
  The resulting forecast errors for that period are large (Figure \ref{ex5_err}),
  which distorts the AR estimation and the second-stage adapt$^2$ forecasting, as well as the correlograms of the forecast errors (Figure \ref{ex5_acf}).

The adaptive method down-weights these large forecast errors and, in many cases, assigns substantial weight to the most recent observation. This is seen  from the tuning parameter values  for the adapt method (Figure \ref{ex5_rho}), which frequently attain the lower bound of the optimization range. Consequently, the adapt method behaves similarly to a last-observation-based forecast.

Based on the adapt method, the forecast of the unemployment rate for January 2025 is  4.1\%, whereas the observed value was 4\%.

The user could set
 \code{p_max = 0}, i.e. run \code{print(forAD(UNRATE, p_max = 0, P = 72, date_1 = "2025-01-01"))},
 to focus on the outputs of the adapt and adapt$^2$ methods only.

Overall, the adaptive forecasting performs rather well, keeping in mind that  no pre-treatment of the data was applied around the COVID-19  period.

\vfill

\pagebreak

We have the following outputs:
%
\begin{CodeChunk}
\begin{CodeOutput}
-----------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output
-----------------------------------------------------------

Forecasts x[N+1|N], MSFE[N+1|N], RMSFE[N+1|N]

|             | Forecast|  MSFE| Relative MSFE|
|:------------|--------:|-----:|-------------:|
|Adapt        |     4.10|  1.30|             1|
|Adapt+AR(1)  |     4.09|  3.05|          2.35|
|Adapt+AR(2)  |     4.07|  3.13|          2.40|
|Adapt+AR(3)  |     4.07|  3.34|          2.56|
|------------ | --------|  ----| -------------|
|Adapt²       |     4.11|  2.78|          2.14|
|Adapt²+AR(1) |     4.11|  8.65|          6.65|
|Adapt²+AR(2) |     4.08| 10.50|          8.07|
|Adapt²+AR(3) |     4.08| 14.14|         10.87|


AR coefficients and standard errors (adapt+AR)

|      |     |   const|     ar1|      ar2|     ar3|
|:-----|:----|-------:|-------:|--------:|-------:|
|AR(1) |     |  -0.006|   0.025|       NA|      NA|
|      |s.e. | (0.029)| (0.050)|     (NA)|    (NA)|
|AR(2) |     |  -0.007|   0.028| -0.116**|      NA|
|      |s.e. | (0.029)| (0.049)|  (0.049)|    (NA)|
|AR(3) |     |  -0.007|   0.025| -0.115**|  -0.020|
|      |s.e. | (0.029)| (0.050)|  (0.050)| (0.050)|

AR coefficients and standard errors (adapt²+AR)

|      |     |  const|      ar1|      ar2|    ar3|
|:-----|:----|------:|--------:|--------:|------:|
|AR(1) |     |  -0.02| -0.31***|       NA|     NA|
|      |s.e. | (0.04)|   (0.04)|     (NA)|   (NA)|
|AR(2) |     |  -0.03| -0.36***| -0.17***|     NA|
|      |s.e. | (0.03)|   (0.04)|   (0.04)|   (NA)|
|AR(3) |     |  -0.03| -0.36***| -0.18***|  -0.02|
|      |s.e. | (0.04)|   (0.05)|   (0.05)| (0.05)|


------------------------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output x[N-k+1|N-k]
------------------------------------------------------------------------

Relative mean squared forecast errors RMSFE[N-k+1|N-k]

|             |  k=10|   k=9|   k=8|   k=7|   k=6|   k=5|   k=4|   k=3|   k=2|   k=1|   k=0|
|:------------|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|
|Adapt        |     1|     1|     1|     1|     1|     1|     1|     1|     1|     1|     1|
|Adapt+AR(1)  |  2.35|  2.35|  2.35|  2.35|  2.35|  2.35|  2.35|  2.34|  2.34|  2.35|  2.35|
|Adapt+AR(2)  |  2.40|  2.40|  2.41|  2.40|  2.40|  2.40|  2.41|  2.40|  2.40|  2.40|  2.40|
|Adapt+AR(3)  |  2.56|  2.56|  2.57|  2.56|  2.56|  2.56|  2.56|  2.56|  2.56|  2.56|  2.56|
|------------ |  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|
|Adapt²       |  2.13|  2.14|  2.14|  2.13|  2.13|  2.13|  2.13|  2.14|  2.13|  2.13|  2.14|
|Adapt²+AR(1) |  6.64|  6.64|  6.65|  6.64|  6.65|  6.64|  6.64|  6.65|  6.64|  6.65|  6.65|
|Adapt²+AR(2) |  8.06|  8.07|  8.06|  8.07|  8.07|  8.06|  8.07|  8.06|  8.06|  8.07|  8.07|
|Adapt²+AR(3) | 10.86| 10.87| 10.86| 10.86| 10.87| 10.86| 10.87| 10.86| 10.86| 10.87| 10.87|


Forecasts x[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Data         |  3.9|  3.9|    4|  4.1|  4.2|  4.2|  4.1|  4.1|  4.2|  4.1|     |
|-----------  | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt        | 3.89| 3.89| 3.89| 3.99| 4.09| 4.19| 4.19| 4.10| 4.10| 4.19| 4.10|
|Adapt+AR(1)  | 3.89| 3.89| 3.89| 3.99| 4.08| 4.19| 4.19| 4.09| 4.09| 4.19| 4.09|
|Adapt+AR(2)  | 3.91| 3.87| 3.89| 3.99| 4.07| 4.18| 4.18| 4.09| 4.10| 4.19| 4.07|
|Adapt+AR(3)  | 3.90| 3.87| 3.88| 3.99| 4.07| 4.18| 4.17| 4.08| 4.10| 4.19| 4.07|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 3.90| 3.90| 3.90| 4.00| 4.09| 4.30| 4.23| 4.04| 4.10| 4.23| 4.11|
|Adapt²+AR(1) | 3.80| 3.86| 3.86| 3.93| 4.02| 4.24| 4.23| 4.04| 4.05| 4.17| 4.11|
|Adapt²+AR(2) | 3.81| 3.79| 3.85| 3.92| 3.98| 4.20| 4.20| 4.06| 4.07| 4.15| 4.08|
|Adapt²+AR(3) | 3.81| 3.80| 3.84| 3.92| 3.98| 4.19| 4.19| 4.06| 4.08| 4.15| 4.08|


Tuning parameters rho[N-k+1|N-k]

|       |  k=10|   k=9|   k=8|   k=7|   k=6|   k=5|   k=4|   k=3|   k=2|   k=1|   k=0|
|:------|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|
|Adapt  | 0.001| 0.001| 0.001| 0.001| 0.086| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001|
|Adapt² | 0.999| 0.999| 0.999| 0.999| 0.999| 0.001| 0.349| 0.342| 0.699| 0.755| 0.840|
\end{CodeOutput}
\end{CodeChunk}

\pagebreak

\captionsetup{belowskip=7pt}
\captionsetup[subfloat]{farskip=10pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selection of the best forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_01.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_02.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_03.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_06.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_04.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_07.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_05.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_08.pdf}}
\caption{Comparison of the mean squared forecasts errors (MSFE) for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex5_msfe}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Forecasting results} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_10.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_14.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_11.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_15.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_12.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_16.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_13.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_17.pdf}}
\caption{Data and their forecasts for the forecasting methods Adapt, Adapt+AR(1), Adapt$^2$, Adapt$^2$+AR(1).}
\label{ex5_for}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selected tuning parameter} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_09.pdf}}
\caption{Selected tuning parameter for the forecasting methods Adapt, Adapt$^2$.}
\label{ex5_rho}
\end{figure}

\bigskip\bigskip

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Evaluation of forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_18.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_19.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex5_20.pdf}}
\caption{Forecasts errors for the forecasting methods Adapt+AR(1), Adapt$^2$+AR(1).}
\label{ex5_err}
\end{figure}

\clearpage

\captionsetup{belowskip=-2pt}
\captionsetup[subfloat]{farskip=18pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Testing for correlation in forecast errors} \vspace{6pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex5_21.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex5_25.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex5_22.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex5_26.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex5_23.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex5_27.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex5_24.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex5_28.pdf}}
\caption{Correlogram of forecasts errors for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex5_acf}
\end{figure}

\clearpage
\pagebreak

\subsubsection*{Empirical example 3}

We use the time series of the US quarterly growth rate $\{G_t\}$ (in \%), seasonally adjusted, over the period 1950Q1 -- 2024Q4 with sample size $N=297$ (retrieved on 21/09/2025). We employ the package \pkg{quantmod} \citep{quantmod} to retrieve the data from FRED. This package depends on the \pkg{xts} package  \citep{xts} for handling the time series data. We exclude the quarters 2020Q1-Q3,  to mitigate the impact of the COVID-19 pandemic, reducing the sample size to $N=294$.

\begin{CodeChunk}
\begin{CodeInput}
R> library(quantmod)
R> getSymbols("GDP", src = "FRED", return.class = "xts")
R> G <- 100 * (GDP - lag(GDP, k = 1)) / lag(GDP, k = 1)
R> G <- window(G, start = "1950-01-01", end = "2024-10-01")
R> G <- G[!(index(G) %in% as.Date(c("2020-01-01", "2020-04-01", "2020-07-01")))]
\end{CodeInput}
\end{CodeChunk}
%
We use the function \code{forAD} to evaluate adaptive forecasting for $\{G_t\}$.
%
\begin{CodeChunk}
\begin{CodeInput}
R> library(forecastADAPT)
R> print(forAD(G, date_1 = "2025-01-01"))
\end{CodeInput}
\end{CodeChunk}

From the tables, we observe that the relative MSFE of the adapt$^2$, adapt$^2$+AR forecasts exceeds 1, while those of the adapt+AR are only slightly over 1. Together with the MSFE plots (Figure \ref{ex6_msfe}), this suggest that the simple adapt method is sufficient. Given the small differences in MSFE, the forecasts across the various methods are very similar, as  seen in the tables and forecast plots (Figure \ref{ex6_for}).

The tables show  that the estimated AR coefficients are not statistically significant (with one exception),
indicating that AR adjustments are not required. Moreover, the second-stage  adapt$^2$ procedure is not needed either: this  is evident from
the selected tuning parameters (Figure \ref{ex6_rho}) which are equal to 1, implying that the second-stage adapt just reduces to the sample mean.
This is further supported by the forecast error plot (see Figure 29), where applying adaptive forecasting to the adapt forecast errors yields values very close to zero.

Furthermore, the correlograms (Figure \ref{ex6_acf}) show no substantial autocorrelation in forecast errors across all methods, further supporting the use of the simple adapt approach.
 Based on the adapt method, the forecast of growth rate of GDP for 2025Q1 is evaluated at 1.24\%, compared to the observed value  0.8\%, reflecting a slowdown of the economy.

The user could  set \code{p_max = 0}, i.e. run \code{print(forAD(G, p_max = 0, date_1 = "2025-01-01"))}, to focus on the outputs of the adapt and adapt$^2$ methods only.

Overall, the adaptive forecasting performs rather well, particularly given that  only minimal pre-treatment of the data was applied around the COVID-19 period.

\vfill

\pagebreak

We have the following outputs:
%
\begin{CodeChunk}
\begin{CodeOutput}
-----------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output
-----------------------------------------------------------

Forecasts x[N+1|N], MSFE[N+1|N], RMSFE[N+1|N]

|             | Forecast| MSFE| Relative MSFE|
|:------------|--------:|----:|-------------:|
|Adapt        |     1.24| 0.42|             1|
|Adapt+AR(1)  |     1.20| 0.42|          0.99|
|Adapt+AR(2)  |     1.19| 0.41|          0.99|
|Adapt+AR(3)  |     1.19| 0.42|          1.01|
|------------ | --------| ----| -------------|
|Adapt²       |     1.23| 0.44|          1.06|
|Adapt²+AR(1) |     1.18| 0.45|          1.07|
|Adapt²+AR(2) |     1.17| 0.45|          1.07|
|Adapt²+AR(3) |     1.17| 0.46|          1.09|


AR coefficients and standard errors (adapt+AR)

|      |     |  const|    ar1|    ar2|     ar3|
|:-----|:----|------:|------:|------:|-------:|
|AR(1) |     |  -0.03|   0.08|       |        |
|      |s.e. | (0.05)| (0.05)|   (NA)|    (NA)|
|AR(2) |     |  -0.03|   0.08|   0.05|        |
|      |s.e. | (0.05)| (0.05)| (0.05)|    (NA)|
|AR(3) |     |  -0.02|   0.09|   0.06| -0.12**|
|      |s.e. | (0.05)| (0.05)| (0.05)|  (0.05)|

AR coefficients and standard errors (adapt²+AR)

|      |     |   const|     ar1|     ar2|     ar3|
|:-----|:----|-------:|-------:|-------:|-------:|
|AR(1) |     |  -0.005|   0.013|        |        |
|      |s.e. | (0.054)| (0.060)|    (NA)|    (NA)|
|AR(2) |     |  -0.003|   0.011|   0.047|        |
|      |s.e. | (0.054)| (0.060)| (0.060)|    (NA)|
|AR(3) |     |  -0.001|   0.015|   0.047|  -0.096|
|      |s.e. | (0.054)| (0.060)| (0.060)| (0.060)|


------------------------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output x[N-k+1|N-k]
------------------------------------------------------------------------

Relative mean squared forecast errors RMSFE[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt        |    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|
|Adapt+AR(1)  | 1.00| 1.00| 0.98| 0.99| 1.00| 0.99| 1.00| 1.01| 1.00| 0.99| 0.99|
|Adapt+AR(2)  | 0.99| 0.98| 0.98| 1.00| 1.00| 0.99| 1.00| 1.01| 0.99| 0.98| 0.99|
|Adapt+AR(3)  | 1.00| 1.00| 1.00| 1.01| 1.01| 1.00| 1.03| 1.03| 1.01| 1.01| 1.01|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 1.07| 1.05| 1.06| 1.06| 1.06| 1.08| 1.07| 1.07| 1.08| 1.06| 1.06|
|Adapt²+AR(1) | 1.07| 1.06| 1.07| 1.07| 1.07| 1.09| 1.08| 1.08| 1.09| 1.06| 1.07|
|Adapt²+AR(2) | 1.07| 1.06| 1.09| 1.06| 1.07| 1.08| 1.07| 1.10| 1.07| 1.07| 1.07|
|Adapt²+AR(3) | 1.08| 1.07| 1.09| 1.08| 1.09| 1.10| 1.09| 1.11| 1.09| 1.09| 1.09|


Forecasts x[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Data         | 1.80| 1.75| 1.60| 1.06| 1.87| 1.17| 1.15| 1.37| 1.23| 1.18|     |
|-----------  | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt        | 2.38| 2.11| 1.91| 1.79| 1.45| 1.66| 1.47| 1.34| 1.34| 1.29| 1.24|
|Adapt+AR(1)  | 2.35| 2.04| 1.85| 1.74| 1.35| 1.67| 1.40| 1.28| 1.31| 1.25| 1.20|
|Adapt+AR(2)  | 2.27| 2.03| 1.82| 1.71| 1.34| 1.62| 1.42| 1.25| 1.29| 1.25| 1.19|
|Adapt+AR(3)  | 2.13| 2.21| 1.82| 1.78| 1.37| 1.67| 1.51| 1.20| 1.36| 1.29| 1.19|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 2.37| 2.11| 1.90| 1.78| 1.44| 1.65| 1.46| 1.33| 1.33| 1.28| 1.23|
|Adapt²+AR(1) | 2.35| 2.03| 1.84| 1.72| 1.33| 1.67| 1.38| 1.26| 1.30| 1.23| 1.18|
|Adapt²+AR(2) | 2.21| 2.02| 1.79| 1.69| 1.30| 1.59| 1.43| 1.21| 1.27| 1.24| 1.17|
|Adapt²+AR(3) | 1.97| 2.35| 1.80| 1.82| 1.37| 1.67| 1.59| 1.12| 1.38| 1.31| 1.17|


Tuning parameters rho[N-k+1|N-k]

|       | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt  | 0.66| 0.53| 0.48| 0.53| 0.53| 0.59| 0.59| 0.59| 0.57| 0.55| 0.55|
|Adapt² | 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99| 0.99|
\end{CodeOutput}
\end{CodeChunk}

\pagebreak

\captionsetup{belowskip=7pt}
\captionsetup[subfloat]{farskip=10pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selection of the best forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_01.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_02.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_03.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_06.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_04.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_07.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_05.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_08.pdf}}
\caption{Comparison of the mean squared forecasts errors (MSFE) for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex6_msfe}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Forecasting results} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_10.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_14.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_11.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_15.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_12.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_16.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_13.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_17.pdf}}
\caption{Data and their forecasts for the forecasting methods Adapt, Adapt+AR(1), Adapt$^2$, Adapt$^2$+AR(1).}
\label{ex6_for}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selected tuning parameter} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_09.pdf}}
\caption{Selected tuning parameter for the forecasting methods Adapt, Adapt$^2$.}
\label{ex6_rho}
\end{figure}

\bigskip\bigskip

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Evaluation of forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_18.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_19.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex6_20.pdf}}
\caption{Forecasts errors for the forecasting methods Adapt+AR(1), Adapt$^2$+AR(1).}
\label{ex6_err}
\end{figure}

\clearpage

\captionsetup{belowskip=-2pt}
\captionsetup[subfloat]{farskip=18pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Testing for correlation in forecast errors} \vspace{6pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex6_21.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex6_25.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex6_22.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex6_26.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex6_23.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex6_27.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex6_24.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex6_28.pdf}}
\caption{Correlogram of forecasts errors for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex6_acf}
\end{figure}

\clearpage
\pagebreak

\subsubsection*{Empirical example 4}

We consider the US GDP time series $\{GDP_t\}$  (in billions of USD), seasonally adjusted at an annual rate, over the period 1950Q1 -- 2024Q4, with sample size $N=297$ (retrieved on 21/09/2025). The data are retrieved from FRED using the \pkg{quantmod} package \citep{quantmod}, which depends on the \pkg{xts} package \citep{xts} for handling time series data.
To mitigate the effect of extreme values (outliers)
associated with the COVID-19 period, we exclude the quarters 2020Q1–2020Q3. This reduces the sample size to $N=294$.

\begin{CodeChunk}
\begin{CodeInput}
R> getSymbols("GDP", src = "FRED", return.class = "xts")
R> GDP <- window(GDP, start = "1950-10-01", end = "2024-10-01")
R> GDP <- GDP[!(index(GDP) %in% as.Date(c("2020-01-01", "2020-04-01", "2020-07-01"
+         )))]
\end{CodeInput}
\end{CodeChunk}
%
We use the function \code{forAD} to evaluate adaptive forecasting for $\{GDP_t\}$.
%
\begin{CodeChunk}
\begin{CodeInput}
R> library(forecastADAPT)
R> print(forAD(GDP, date_1 = "2025-01-01"))
\end{CodeInput}
\end{CodeChunk}

From the tables, we see that the relative MSFE of the adapt+AR, adapt$^2$, adapt$^2$+AR forecasts are well below 1. Together with the MSFE plots (Figure \ref{ex7_msfe}),
this suggest that the simple adapt method is not sufficient. Among the alternative methods, the adapt$^2$ achieves a slightly lower MSFE. The rather weak performance of the adapt forecasts is also evident from the tables and forecast plots (Figure \ref{ex7_for}). Implementing  AR(1) dynamics and the second-stage adapt$^2$ forecast procedure, visibly improves  the  forecasting performance of  the  adapt method.

The tables show that  the estimated AR coefficients  are  not significant for the adapt$^2$+AR method, while they are  significant for the adapt+AR method.
The second level adapt$^2$ adjustment is required, as the selected tuning parameters (Figure \ref{ex7_rho}) for the adapt$^2$ method lie away from the upper bound of 1. This is also seen from the forecast error plots (Figure \ref{ex7_err}), where applying   adaptive forecasting  to the adapt forecast errors, tracks a non-zero time-varying mean.
 In addition, the correlograms (Figure \ref{ex7_acf}) show slow-decaying autocorrelation in the adapt forecast errors, reinforcing the need for a second round of adaptive adjustment.
  Using the adapt$^2$
 method, the forecast of GDP for 2025Q1 is 30,083, compared to the observed value of 29,962.

The user could  set \code{p_max = 0}, i.e. run
\code{print(forAD(GDP, p_max = 0, date_1 = "2025-01-01"))}, to focus on the outputs of the adapt and adapt$^2$ methods only.


Overall, adaptive forecasting performs very well, particularly given that only minimal pre-treatment of the data was applied around the COVID-19 period.



\vfill

\pagebreak

We have the following outputs:
%
\begin{CodeChunk}
\begin{CodeOutput}
-----------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output
-----------------------------------------------------------

Forecasts x[N+1|N], MSFE[N+1|N], RMSFE[N+1|N]

|             | Forecast|  MSFE| Relative MSFE|
|:------------|--------:|-----:|-------------:|
|Adapt        |    29723| 67576|             1|
|Adapt+AR(1)  |    30029| 14430|          0.21|
|Adapt+AR(2)  |    30049| 13235|          0.19|
|Adapt+AR(3)  |    30062| 13320|          0.19|
|------------ | --------|  ----| -------------|
|Adapt²       |    30083| 12971|          0.19|
|Adapt²+AR(1) |    30391| 13543|          0.20|
|Adapt²+AR(2) |    30411| 13944|          0.20|
|Adapt²+AR(3) |    30423| 14803|          0.21|


AR coefficients and standard errors (adapt+AR)

|      |     |    const|     ar1|     ar2|     ar3|
|:-----|:----|--------:|-------:|-------:|-------:|
|AR(1) |     | 18.98***| 0.82***|        |        |
|      |s.e. |   (5.46)|  (0.03)|    (NA)|    (NA)|
|AR(2) |     |  12.90**| 0.54***| 0.34***|        |
|      |s.e. |   (5.25)|  (0.05)|  (0.05)|    (NA)|
|AR(3) |     |  11.03**| 0.48***| 0.25***| 0.16***|
|      |s.e. |   (5.26)|  (0.05)|  (0.06)|  (0.05)|

AR coefficients and standard errors (adapt²+AR)

|      |     |   const|     ar1|     ar2|     ar3|
|:-----|:----|-------:|-------:|-------:|-------:|
|AR(1) |     |   2.724|   0.013|        |        |
|      |s.e. | (4.394)| (0.060)|    (NA)|    (NA)|
|AR(2) |     |   2.689|   0.013|   0.016|        |
|      |s.e. | (4.422)| (0.061)| (0.061)|    (NA)|
|AR(3) |     |   2.664|   0.013|   0.016|   0.007|
|      |s.e. | (4.450)| (0.061)| (0.061)| (0.061)|


------------------------------------------------------------------------
Adaptive+AR and adaptive²+AR one-step ahead forecast output x[N-k+1|N-k]
------------------------------------------------------------------------

Relative mean squared forecast errors RMSFE[N-k+1|N-k]

|             | k=10|  k=9|  k=8|  k=7|  k=6|  k=5|  k=4|  k=3|  k=2|  k=1|  k=0|
|:------------|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|----:|
|Adapt        |    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|    1|
|Adapt+AR(1)  | 0.25| 0.22| 0.21| 0.20| 0.23| 0.22| 0.21| 0.25| 0.24| 0.22| 0.21|
|Adapt+AR(2)  | 0.22| 0.21| 0.19| 0.23| 0.21| 0.20| 0.19| 0.23| 0.21| 0.20| 0.19|
|Adapt+AR(3)  | 0.22| 0.20| 0.19| 0.23| 0.21| 0.20| 0.24| 0.22| 0.21| 0.20| 0.19|
|------------ | ----| ----| ----| ----| ----| ----| ----| ----| ----| ----| ----|
|Adapt²       | 0.22| 0.20| 0.18| 0.21| 0.20| 0.24| 0.21| 0.19| 0.22| 0.20| 0.19|
|Adapt²+AR(1) | 0.22| 0.20| 0.19| 0.21| 0.20| 0.24| 0.21| 0.20| 0.22| 0.21| 0.20|
|Adapt²+AR(2) | 0.22| 0.20| 0.24| 0.22| 0.20| 0.23| 0.22| 0.26| 0.23| 0.21| 0.20|
|Adapt²+AR(3) | 0.23| 0.21| 0.25| 0.22| 0.21| 0.24| 0.23| 0.26| 0.23| 0.22| 0.21|


Forecasts x[N-k+1|N-k]

|             |  k=10|   k=9|   k=8|   k=7|   k=6|   k=5|   k=4|   k=3|   k=2|   k=1|   k=0|
|:------------|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|
|Data         | 26272| 26734| 27164| 27453| 27967| 28296| 28624| 29016| 29374| 29723|      |
|-----------  |  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|
|Adapt        | 25805| 26271| 26733| 27163| 27453| 27967| 28296| 28623| 29016| 29374| 29723|
|Adapt+AR(1)  | 26303| 26667| 27128| 27534| 27707| 28408| 28583| 28909| 29357| 29687| 30029|
|Adapt+AR(2)  | 26293| 26737| 27157| 27568| 27765| 28360| 28665| 28926| 29354| 29717| 30049|
|Adapt+AR(3)  | 26386| 26734| 27187| 27580| 27783| 28376| 28649| 28961| 29356| 29715| 30062|
|------------ |  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|  ----|
|Adapt²       | 26385| 26798| 27227| 27629| 27831| 28412| 28691| 28988| 29393| 29742| 30083|
|Adapt²+AR(1) | 26888| 27195| 27624| 28000| 28081| 28860| 28980| 29275| 29737| 30058| 30391|
|Adapt²+AR(2) | 26888| 27266| 27652| 28034| 28138| 28809| 29065| 29290| 29733| 30088| 30411|
|Adapt²+AR(3) | 26977| 27276| 27683| 28049| 28157| 28827| 29050| 29325| 29735| 30086| 30423|


Tuning parameters rho[N-k+1|N-k]

|       |  k=10|   k=9|   k=8|   k=7|   k=6|   k=5|   k=4|   k=3|   k=2|   k=1|   k=0|
|:------|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|
|Adapt  | 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001| 0.001|
|Adapt² |  0.55|  0.52|  0.50|  0.52|  0.50|  0.56|  0.56|  0.55|  0.55|  0.54|  0.54|
\end{CodeOutput}
\end{CodeChunk}

\pagebreak

\captionsetup{belowskip=7pt}
\captionsetup[subfloat]{farskip=10pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selection of the best forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_01.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_02.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_03.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_06.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_04.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_07.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_05.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_08.pdf}}
\caption{Comparison of the mean squared forecasts errors (MSFE) for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex7_msfe}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Forecasting results} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_10.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_14.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_11.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_15.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_12.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_16.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_13.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_17.pdf}}
\caption{Data and their forecasts for the forecasting methods Adapt, Adapt+AR(1), Adapt$^2$, Adapt$^2$+AR(1).}
\label{ex7_for}
\end{figure}

\clearpage

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Selected tuning parameter} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_09.pdf}}
\caption{Selected tuning parameter for the forecasting methods Adapt, Adapt$^2$.}
\label{ex7_rho}
\end{figure}

\bigskip\bigskip

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Evaluation of forecasting method} \vspace{-2pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_18.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_19.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={33 47 29 2},clip]{plots_ex7_20.pdf}}
\caption{Forecasts errors for the forecasting methods Adapt+AR(1), Adapt$^2$+AR(1).}
\label{ex7_err}
\end{figure}

\clearpage

\captionsetup{belowskip=-2pt}
\captionsetup[subfloat]{farskip=18pt}

\begin{figure}[h!]
\centering
\caption*{\small \textbf{{\color{gray}Testing for correlation in forecast errors} \vspace{6pt}}}
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex7_21.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex7_25.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex7_22.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex7_26.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex7_23.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex7_27.pdf}}
\\
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex7_24.pdf}} \qquad
\subfloat{\includegraphics[width=0.47\textwidth,trim={5 5 5 5},clip]{plots_ex7_28.pdf}}
\caption{Correlogram of forecasts errors for the forecasting methods Adapt, Adapt+AR, Adapt$^2$, Adapt$^2$+AR.}
\label{ex7_acf}
\end{figure}

\clearpage
\pagebreak


\bibliography{forecastADAPT}

\end{document}