\documentclass[a4paper]{article}
\usepackage[round]{natbib}
\usepackage{amssymb}
%% \usepackage{RJournal}
\usepackage{fancyvrb}
\usepackage{Sweave, url}
\usepackage{hyperref}


\DefineVerbatimEnvironment{Sinput}{Verbatim}{fontsize=\small,fontshape=sl}
\DefineVerbatimEnvironment{Soutput}{Verbatim}{fontsize=\small}
\DefineVerbatimEnvironment{Scode}{Verbatim}{fontsize=\small,fontshape=sl}

\SweaveOpts{echo = TRUE, results = verbatim, keep.source = TRUE}

\bibliographystyle{abbrvnat}

\begin{document}

\title{Performance Attribution for Equity Portfolios}
\author{Yang Lu\thanks{\href{mailto:yang.lu2014@gmail.com}{yang.lu2014@gmail.com}} and David Kane\thanks{\href{mailto:dave.kane@gmail.com}{dave.kane@gmail.com}}}

%%\VignetteIndexEntry{Using the pa package}
%%\VignetteDepends{pa}

\maketitle

\setkeys{Gin}{width=0.95\textwidth}

\section{Introduction}

Many portfolio managers measure performance with reference to a
benchmark. The difference in return between a portfolio and its
benchmark is the active return of the portfolio. Portfolio managers
and their clients want to know what caused this active
return. Performance attribution decomposes the active return. The two
most common approaches are the Brinson-Hood-Beebower (hereafter
referred to as the Brinson model) and a regression-based
analysis.\footnote{See \cite{brinson:gary} and \cite{jpmreg} for more
  information.}

Portfolio managers use different variations of the two models to
assess the performance of their portfolios. Managers of fixed income
portfolios include yield-curve movements in the model \cite{lord}
while equity managers who focus on the effect of currency movements
use variations of the Brinson model to incorporate ``local risk
premium'' \cite{localrisk}. In contrast, in this paper we focus on
attribution models for equity portfolios without considering any
currency effect.


The \textbf{pa} package provides tools for conducting both methods for
equity portfolios. The Brinson model takes an ANOVA-type approach and
decomposes the active return of any portfolio into asset allocation,
stock selection, and interaction effects. The regression-based
analysis utilizes estimated coefficients from a linear model to
attribute active return to different factors. After describing the
Brinson and regression approaches and demonstrating their use via the
\texttt{pa} package, we show that the Brinson model is just a special
case of the regression approach.


\section{Data}

We demonstrate the use of the \textbf{pa} package with a series of
examples based on real-world data sets from MSCI Barra's Global Equity
Model II(GEM2).\footnote{See www.msci.com and \cite{gem2} for more
  information.} MSCI Barra is a leading provider of investment
decision support tools to investment institutions worldwide. According
to the company:

% Note that the font size for this quote small and then go
% back to normal.

\small

\begin{quote}
  
  GEM2 is the latest Barra global multi-factor equity model. It
  provides a foundation for investment decision support tools via a
  broad range of insightful analytics for developed and emerging
  market portfolios. The latest model version provides:

\begin{itemize}
\item Improved accuracy of risk forecasts and increased explanatory
  power.

\item An intuitive structure that accommodates different investment
  processes in developed vs. emerging markets.

\item Greater responsiveness to market dynamics.

\item Comprehensive market coverage.

\end{itemize}

GEM2 leverages the decades of experience that MSCI Barra has in
developing and maintaining global equity multi-factor models and
indices, and offers important enhancements over GEM, which is utilized
by hundreds of institutional fund managers worldwide.
\end{quote}

\normalsize

The original data set contains selected attributes such as industry,
size, country, and various style factors for a universe of
approximately 48,000 securities on a monthly basis.

For illustrative purposes, this article uses three modified versions
of the original data set, containing 3000 securities, namely
\texttt{year}, \texttt{quarter}, and \texttt{jan}. The data frame,
\texttt{quarter}, is a subset of \texttt{year}, containing the data of
the first quarter. The data frame, \texttt{jan}, is a subset of
\texttt{quarter} with the data from January, 2010.

<<echo = FALSE, results = hide>>=
options(width = 50, digits = 2, scipen = 5)
library(pa)
@

<<>>=
data(year)
names(year)
@

\begin{itemize}

\item barrid: security identifier by Barra.

\item name: name of a security.

\item return: monthly total return in trading currency.

\item date: the starting date of the month to which the data belong.

\item sector: consolidated sector categories based on the
  GICS.\footnote{Global Industry Classification Standard}

\item momentum: capture sustained relative performance.

\item value: capture the extent to which a stock is priced
  inexpensively in the market.

\item size: differentiate between large and small cap companies.

\item growth: capture stock's growth prospects.

\item cap.usd: capitalization in model base currency USD.

\item yield: dividend of a security.

\item country: the country in which the company is traded.

\item currency: currency of exposure.

\item portfolio: top 200 securities based on \texttt{value} scores
  in January are selected as portfolio holdings and are held through
  December 2010. This is to avoid the complexity of trading in the
  analyses.

\item benchmark: top 1000 securities based on size each month. The
  benchmark is cap-weighted.

\end{itemize}


Here is a sample of rows and columns from the data frame
\texttt{year}:

<<echo = FALSE>>=
sample.mat <- rbind(which(year$portfolio == 0 & year$benchmark == 0)[2],
                    which(year$portfolio == 0 & year$benchmark > 0)[2],
                    which(year$portfolio > 0 & year$benchmark == 0)[2],
                    which(year$return > 0 & year$portfolio > 0 & year$benchmark > 0)[500],
                    which(year$barrid == "CANAITC")[11])
year[sample.mat, c(-1, -6, -7, -9: -11, -13) ]
@

The portfolio has 200 equal-weighted holdings. The row for Canadian
Imperial Bank of Commerce indicates that it is one of the 200
portfolio holdings with a weight of 0.5\% in 2010. Its return was
2.61\% in August, and almost flat in November.

\section{The Brinson Model}

\subsection{Single-Period Brinson Model}

Consider an equity portfolio manager who uses the S\&P 500 as the
benchmark. In a given month, she outperformed the S\&P by 3\%. Part of
that performance was due to the fact that she allocated more weight of
the portfolio to certain sectors that performed well. Call this the
\emph{allocation effect}. Part of her outperformance was due to the
fact that some of the stocks she selected did better than their sector
as a whole. Call this the \emph{selection effect}. The residual can
then be attributed to an interaction between allocation and selection
-- the \emph{interaction effect}. The Brinson model provides
mathematical definitions for these terms and methods for calculating
them.

The example above uses sector as the classification scheme when
calculating the allocation effect. But the same approach can work with
any other variable which places each security into one, and only one,
discrete category: country, industry, and so on. In fact, a similar
approach can work with continuous variables that are split into
discrete ranges: the highest quintile of market cap, the second
highest quintile and so forth. For generality, we will use the term
``category'' to describe any classification scheme which places each
security in one, and only one, category.

Notations:

\begin{itemize}

\item $w^B_i$ is the weight of security $i$ in the benchmark.
\item $w^P_i$ is the weight of security $i$ in the portfolio.
\item $W^B_j$ is the weight of category $j$ in the benchmark. $W^B_j =
  \sum w^B_i$, $i$ $\in$ $j$.
\item $W^P_j$ is the weight of a category $j$ in the portfolio. $W^P_j
  = \sum w^P_i$, $i$ $\in$ $j$.
\item The sum of the weight $w^B_i$, $w^P_i$, $W^B_j$, and $W^P_j$ is
  1, respectively.
\item $r_i$ is the return of security $i$.
\item $R^B_j$ is the return of a category $j$ in the benchmark. $R^B_j
  = \sum w^B_ir_i$, $i$ $\in$ $j$.
\item $R^P_j$ is the return of a category $j$ in the portfolio. $R^P_j
  = \sum w^P_ir_i$, $i$ $\in$ $j$.
\end{itemize}

The return of a portfolio, $R_P$, can be calculated in two ways:

\begin{itemize}
  \item On an individual security level by summing over $n$ stocks: $R_P =
    \sum\limits_{i = 1}^n w^P_ir_i$.
  \item On a category level by summing over $N$ categories: $R_P = \sum\limits_{j =
    1}^N W^P_jR^P_j$.
\end{itemize}

Similar definitions apply to the return of the benchmark, $R_B$,

\begin{itemize}
  \item $R_B = \sum\limits_{i = 1}^n w^B_ir_i$.
  \item $R_B = \sum\limits_{j = 1}^N W^B_jR^B_j$.
\end{itemize}

Active return of a portfolio, $R_{active}$, is a performance measure
of a portfolio relative to its benchmark. The two conventional
measures of active return are arithmetic and geometric.  The \textbf{pa}
package implements the arithmetic measure of the active return for a
single-period Brinson model because an arithmetic difference is more
intuitive than a ratio over a single period.

The arithmetic active return of a portfolio, $R_{active}$, is the
portfolio return $R_P$ less the benchmark return $R_B$:

\begin{center}
  $R_{active} = R_P - R_B$.
\end{center}

Since the category allocation of the portfolio is generally different
from that of the benchmark, allocation plays a role in the active
return, $R_{active}$. The same applies to stock selection where
assuming that the portfolio has the exact same categorical exposures
as the benchmark does, equities within each category are
different. This contributes to $R_{active}$ as well. Allocation effect
$R_{allocation}$ and selection effect $R_{selection}$ over $N$
categories are defined as:
\begin{center}
  $R_{allocation} = \sum\limits_{j = 1}^N W^P_jR^B_j - \sum\limits_{j
    = 1}^N W^B_jR^B_j$,
\end{center}
\begin{center}
  $R_{selection} = \sum\limits_{j = 1}^N W^B_jR^P_j - \sum\limits_{j =
    1}^N W^B_jR^B_j$.
\end{center}

The intuition behind the allocation effect is that a portfolio would
produce different returns with different allocation schemes ($W^P_j$
vs. $W^B_j$) while having the same stock selection and thus the same
return ($R^B_j$) for each category. The difference between the two
returns, caused by the allocation scheme, is called the allocation
effect ($R_{allocation}$). Similarly, two different returns can be
produced when two portfolios have the same allocation ($W^B_j$) yet
dissimilar returns due to differences in stock selection within each
category ($R^p_j$ vs. $R^B_j$). This difference is the selection
effect ($R_{selection}$).

Interaction effect, $R_{interaction}$, is the result of subtracting
return due to allocation $R_{allocation}$ and return due to selection
$R_{selection}$ from the active return $R_{active}$:
\begin{center}
  $R_{interaction} = R_{active} - R_{allocation} - R_{selection}$.
\end{center}


\subsection{Weakness of the Brinson Model}

The Brinson model allows portfolio managers to analyze the relative
return of a portfolio using any attribute of a security, such as
country or sector. One weakness of the model is to expand the analysis
beyond two categories.\footnote{\citet{garytwo} proposed a framework
  to include two variables in the Brinson analysis.} As the number of
categories increases, this procedure is subject to the \textit{curse
  of dimensionality}.

Suppose an equity portfolio manager wants to find out the
contributions of any two categories (for instance, country and sector)
to her portfolio based on the Brinson model. She can decompose the
active return into three broad terms -- $R_{allocation}$,
$R_{selection}$, and $R_{interaction}$. The allocation effect can be
further split into country allocation effect, sector allocation effect
and the product of country and sector allocation effects:
\begin{center}
  $R_{allocation} = R_{country \, allocation} + R_{sector \,
    allocation} + R_{country \, allocation}R_{sector \, allocation}$.
\end{center}

Specifically, the country allocation effect is the return caused by
the difference between the actual country allocation and the benchmark
country allocation while assuming the same benchmark return within
each level of the category country, that is,
\begin{center}
  $R_{country \, allocation} = \sum\limits_{j = 1}^N
  {}_CW^P_j{}_CR^B_j - \sum\limits_{j = 1}^N {}_CW^B_j{}_CR^B_j$,
\end{center}
where 
\begin{itemize}
  \item ${}_CW^P_j$ and ${}_CW^B_j$ refer to the weight of each
    country $j$ ($N_C$ countries in total) in the portfolio and that in
    the benchmark, respectively.
  \item ${}_CR^B_j$ refers to the benchmark return of any country $j$.
\end{itemize}

Similarly, the sector allocation effect is the difference in return
between a portfolio's sector allocation and the benchmark's sector
allocation while having the same benchmark returns:
\begin{center}
  $R_{sector \, allocation} = \sum\limits_{j = 1}^N {}_SW^P_j{}_SR^B_j
  - \sum\limits_{j = 1}^N {}_SW^B_j{}_SR^B_j$,
\end{center}
${}_SW^P_j$ and ${}_SW^B_j$ refer to the weight of the sector $j$ in
the portfolio and the weight of the sector $j$ in the benchmark,
respectively. ${}_SR^B_j$ is the benchmark return of any given sector
$j$ of all $N_S$ sectors.

In the same vein, the return as a result of the selection effect
$R_{selection}$ is the sum of country selection effect, sector
selection effect, and the product of country and sector selection
effects:
\begin{eqnarray*}
  R_{selection} & = & R_{country \, selection} + R_{sector \, selection} \\
  & & + R_{country \, selection} * R_{sector \, selection}\\ 
  & = & \sum\limits_{j = 1}^N {}_CW^B_j{}_CR^P_j - \sum\limits_{j = 1}^N {}_CW^B_j{}_CR^B_j \\ 
  & + & \sum\limits_{j = 1}^N {}_SW^B_j{}_SR^P_j - \sum\limits_{j = 1}^N {}_SW^B_j{}_SR^B_j \\
  & + & (\sum\limits_{j = 1}^N {}_CW^B_j{}_CR^P_j - \sum\limits_{j = 1}^N {}_CW^B_j{}_CR^B_j) \\
  & * & (\sum\limits_{j = 1}^N {}_SW^B_j{}_SR^P_j - \sum\limits_{j = 1}^N {}_SW^B_j{}_SR^B_j).
\end{eqnarray*}

The interaction effect, $R_{interaction}$, includes the interaction
between country allocation and sector selection and that between
country selection and sector allocation.

Therefore, in the case of $Q$ categories where $Q > 1$, the Brinson
model becomes very complex (assume $Q \ge 3$):
\begin{eqnarray*}
  R_{allocation} & = & \sum\limits_{j = 1}^Q R_{allocation_j} +
  \sum\limits_{j = 1}^Q\sum\limits_{k = 1}^Q
  R_{allocation_j}R_{allocation_k} \\ 
  & + & \sum\limits_{j = 1}^Q\sum\limits_{k = 1}^Q\sum\limits_{p = 1}^Q
  R_{allocation_j}R_{allocation_k}R_{allocation_p} \\
  & = & \dots,  
\end{eqnarray*}

\begin{eqnarray*}
  R_{selection} & = & \sum\limits_{j = 1}^Q R_{selection_j} +
  \sum\limits_{j = 1}^Q\sum\limits_{k = 1}^Q
  R_{selection_j}R_{selection_k} \\ & + & \sum\limits_{j = 1}^Q
  \sum\limits_{k = 1}^Q\sum\limits_{p = 1}^Q
  R_{selection_j}R_{selection_k}R_{selection_p} \\
  & = & \dots, 
\end{eqnarray*}
where $R_{allocation_j}$ is the allocation effect of any given
category $j$, $j$ $\in$ $Q$ and $R_{selection_j}$ is the selection
effect of any given category $j$, $j$ $\in$ $Q$. $i$, $j$, $k$ have
different values.

As the number of categories grows, the numbers of terms for the
allocation and the selection effects grow exponentially. $Q$
categories will result in $2^Q - 1$ terms for each of the allocation
and selection effect.

Due to the interaction between allocation and selection of each of the
$Q$ categories (it could be interaction between 2, 3 or even all $Q$
categories), the number of terms included in the interaction effect
grows exponentially to take into all the interaction effects among all
categories.
\begin{eqnarray*}
    R_{interaction} & = & \sum\limits_{j = 1}^Q \sum\limits_{k = 1}^Q R_{allocation_j}R_{selection_k} \\
    & + & \sum\limits_{j = 1}^Q \sum\limits_{k = 1}^Q \sum\limits_{p = 1}^Q R_{allocation_j}R_{selection_k}R_{allocation_p} \\
    & + & \dots.
\end{eqnarray*}
%% In the case of $Q$ categories where $Q > 1$,
%% there are $3^Q - 2Q - 1$ terms for the interaction effect between any
%% of the two categories. The interaction effect can be written as,
%% Another way to think about a multivariate Brinson model is to
%% visualize the effects in cells. Three different characteristics can be
%% assigned to each of the $Q$ categories. They are allocation effect,
%% selection effect and \textit{no effect}. The allocation and selection
%% effects are the same as those in the univariate Brinson model. No
%% effect means a category has no effect on itself or other categories
%% when calculating overall allocation, selection, and interaction
%% effects. Thus, $Q$ categories can break the portfolio into $3^Q$
%% cells. Each cell has a combination of allocation, selection or no
%% effect for each of the $Q$ categories. In the case of 3 categories, if
%% a cell consists of the allocation effect of category $i$, no effect of
%% category $j$, and no effect of category $k$, it represents the
%% allocation effect of category $i$. Should the cell include the
%% allocation effect of category $i$, the selection effect of category
%% $j$, and no effect of category $k$, the cell represents the
%% interaction between the allocation effect of category $i$ and the
%% selection effect of category $j$. To calculate the allocation effect,
%% one can sum up all the cells which contain one of the allocation
%% effects from the $Q$ categories and no effect from other
%% categories. The selection effect is the sum of the returns in all the
%% cells which have one of the selection effects from the $Q$ categories
%% and no effect from other categories. The interaction effect is the sum
%% of the returns in all cells which have more than one allocation or
%% selection effect.
$Q$ categories has $2^{2n} - 2^{n + 1} + 1$ terms of interaction
effects.

For instance, when there are 3 categories, the allocation effect and
the selection effect each have $2^3 - 1 = 7$ terms. The interaction
effect has $2^6 - 2^4 + 1 = 49$ terms. When there are 4 categories,
$2^4 - 1 = 15$ terms have to be estimated for the allocation effect as
well as the selection effect, respectively. $2^8 - 2^5 + 1 = 225$
terms have to be calculated for the interaction effect of 4
categories. This poses a significant computational challenge when a
portfolio manager performs a multivariate Brinson analysis.

To some extent, the regression-based model detailed later solves the
problem of multivariate attribution.


\subsection{Single-Period Brinson Tools}

Brinson analysis is run by calling the function \texttt{brinson} to
produce an object of class \texttt{brinson}. Below we show the tools
provided in the \textbf{pa} package to analyze a single period portfolio
based on the Brinson model.

<<>>=
data(jan)
br.single <- brinson(x = jan, date.var = "date", 
                 cat.var = "sector",
                 bench.weight = "benchmark", 
                 portfolio.weight = "portfolio", 
                 ret.var = "return")
@

The data frame, \texttt{jan}, contains all the information necessary
to conduct a single-period Brinson analysis. \texttt{date.var},
\texttt{cat.var}, and \texttt{return} identify the columns containing
the date, the factor to be analyzed, and the return variable,
respectively. \texttt{bench.weight} and \texttt{portfolio.weight}
specify the name of the benchmark weight column and that of the
portfolio weight column in the data frame.

Calling \texttt{summary} on the resulting object \texttt{br.single} of
class \texttt{brinson} reports essential information about the input
portfolio (including the number of securities in the portfolio and the
benchmark as well as sector exposures) and the results of the Brinson
analysis.

<<>>=
summary(br.single)
@

The \texttt{br.single} summary shows that the active return of the
portfolio, in January, 2010 was 1.47\%. This return can be decomposed
into allocation effect (-0.14\%), selection effect (1.42\%), and
interaction effect (0.19\%).

\begin{figure}
\centering
\vspace*{.1in}
<<fig = TRUE>>=
plot(br.single, var = "sector", type = "return")
@ 
\caption{\label{figure:return}
  Sector Return.}
\end{figure}

Figure \ref{figure:return} is a visual representation of the return of
both the portfolio and the benchmark sector by sector in January,
2010. This plot shows that in absolute terms, Utilities performed the
best with a gain of more than 5\% and Consumer Discretionary, the
worst performing sector, lost more than 10\%. Utilities was also the
sector with the highest active return in the portfolio.

\subsection{Multi-Period Brinson Model}

To obtain Brinson attribution on a multi-period data set, one
calculates allocation, selection and interaction within each period
and aggregates them across time. There are five methods for this --
arithmetic, geometric, optimized linking by \citet{multi}, linking by
\citet{davies}, and linking by \citet{frongello}. We focus on the
first three methods in this paper. Arithmetic measure calculates
relative performance of a portfolio and its benchmark by a difference;
geometric measure does so by a ratio. Arithmetic measure is more
intuitive but a well-known challenge in arithmetic attribution is that
active returns do not add up over multiple periods due to geometric
compounding.\footnote{See \cite{practical} for a complete discussion
  of the complexity involved.} Geometric is able to circumvent the
adding-up problem. \cite{multi} discussed various linking algorithms
to connect arithmetic return with geometric return and argued that the
\emph{optimized linking algorithm} is the best way to link attribution
over time.

\texttt{Arithmetic Attribution.} The arithmetic attribution model
calculates active return and contributions due to allocation,
selection, and interaction in each period and sums them over multiple
periods. The arithmetic active return over $T$ periods
$R_{arithmetic}$ is expressed as:
\begin{center}
  $R_{arithmetic} = \sum\limits_{t = 1}^T R_t^{active}$,
\end{center}
and $R_t^{active}$ is the active return in a single period $t$.

\texttt{Geometric Attribution.} The geometric attribution is to
compound various returns over $T$ periods where,
\begin{center}
  $1 + R_P = \prod\limits_{t = 1}^T (1 + R_t^P)$,
\end{center}
\begin{center}
  $1 + R_B = \prod\limits_{t = 1}^T (1 + R_t^B)$,
\end{center}
and $R_t^P$ and $R_t^B$ are portfolio and benchmark returns in a
single period $t$, respectively. Geometric return $R_{geometric}$ is
thus the difference between $R_p$ and $R_B$:
\begin{center}
  $R_{geometric} = R_p - R_B$.
\end{center}

\texttt{Optimized Linking Algorithm.} The well-known challenge faced
in arithmetic attribution is that the actual active return over time
is not equal to the arithmetic summation of single-period active
returns,
\begin{center}
  $ R_{geometric} \not= R_{arithmetic}$,
\end{center}
i.e.,
\begin{center}
  $R_P - R_B \not= \sum\limits_{t = 1}^T R_t^{active}$.
\end{center}

\cite{multi} proposed an optimized linking coefficient $b^{opt}$ to
link arithmetic returns of individual periods with geometric returns
over time,
\begin{center}
  $R_p - R_B = \sum\limits_{t = 1}^T b_t^{opt}R_t^{active}$,
\end{center}
where $b_t^{opt}$ is the optimized linking coefficient in a single
period $t$.

The optimized linking coefficient $b_t^{opt}$ is the summation of a
\emph{natural scaling} $A$ and an \emph{adjustment} $a_t$ specific to
a time period $t$,
\begin{center}
  $b_t^{opt} = A + a_t$,
\end{center}
where $A$ is an coefficient for linking from the single-period to the
multi-period return and $a_t$ is an adjustment to eliminate
residuals\footnote{See \cite{linking} for more information on the
  optimized linking coefficients.}.

Since active return over time $R_P - R_B$ is a summation of active
return in each period adjusted to the optimized linking algorithm, the
following is true:
\begin{center}
  $R_P - R_B = \sum\limits_{t = 1}^T b_t^{opt}(R_t^{allocation} +
  R_t^{selection} + R_t^{interaction})$,
\end{center}
where $R_t^{allocation}$, $R_t^{selection}$, and $R_t^{interaction}$
represent allocation, selection and interaction in each period $t$,
respectively.

Within each period $t$, the adjusted attribution is thus expressed as
\begin{center}
  $\hat{R}_t^{allocation} = b_t^{opt}R_t^{allocation}$,
\end{center}
\begin{center}
  $\hat{R}_t^{selection} = b_t^{opt}R_t^{selection}$,
\end{center}
and
\begin{center}
  $\hat{R}_t^{interaction} = b_t^{opt}R_t^{interaction}$. 
\end{center}

Therefore, across $T$ periods, active return $R_{active}$, the
difference between portfolio return $R_P$ and benchmark return $R_B$,
can be written as
\begin{center}
  $R_{active} = \sum\limits_{t= 1}^T (\hat{R}_t^{allocation} +
  \hat{R}_t^{selection} + \hat{R}_t^{interaction})$,
\end{center}
where $R_{active} = R_P - R_B$.

\subsection{Multi-Period Brinson Tools}

In practice, analyzing a single-period portfolio is meaningless as
portfolio managers and their clients are more interested in the
performance of a portfolio over multiple periods. To apply the Brinson
model over time, we can use the function \texttt{brinson} and input a
multi-period data set (for instance, \texttt{quarter}) as shown
below. 

<<>>=
data(quarter)
br.multi <- brinson(quarter, date.var = "date",
                    cat.var = "sector",
                    bench.weight = "benchmark",
                    portfolio.weight = "portfolio",
                    ret.var = "return")
@


The object \texttt{br.multi} of class \texttt{brinsonMulti} is an
example of a multi-period Brinson analysis.

<<>>=
exposure(br.multi, var = "size")
@

The \texttt{exposure} method on the class \texttt{br.multi} object
shows the exposure of the portfolio and the benchmark based on a
user-defined category. Here, it shows the exposure on
\texttt{size}. We can see that the portfolio overweights the benchmark
in the lowest quintile in \texttt{size} and underweights in the
highest quintile.

<<>>=
returns(br.multi, type = "linking")
@ 

The \texttt{returns} method shows the results of the Brinson analysis
applied to the data from January, 2010 through March, 2010. The
optimized linking algorithm is applied here by setting the type to
\emph{linking}. The first portion of the \texttt{returns} output shows
the Brinson attribution in individual periods. The second portion
shows the aggregate attribution results. The portfolio formed by top
200 value securities in January had an active return of 12.7\% over
the first quarter of 2010. The allocation and the selection effects
contributed 0.95\% and 1.73\% respectively; the interaction effect
made a loss of 1.42\%.

\begin{figure}
\centering
\vspace*{.1in}
<<fig = TRUE, height = 5, width = 9>>=
plot(br.multi, type = "return")
@
\caption{\label{figure:multireturn} Sector Return Across Time.}
\end{figure}

Figure \ref{figure:multireturn} depicts the returns of both the
portfolio and the benchmark of the allocation effect from January,
2010 through March. 2010. This plot shows that for the portfolio,
\texttt{Utilities} performed the best with a gain of more than 5\% in
January and February, 2010 but tanked in March, 2010.

\section{Regression}

\subsection{Single-Period Regression Model}

One advantage of a regression-based approach is that such analysis
allows one to define their own attribution model by easily
incorporating multiple variables in the regression formula. These
variables can be either discrete or continuous.

Suppose a portfolio manager wants to find out how much each of the
value, growth, and momentum scores of her holdings contributes to the
overall performance of the portfolio. Consider the following linear
regression without the intercept term based on a single-period
portfolio of $n$ securities with $k$ different variables:
\begin{center}
  $\mathbf{r}_n = \mathbf{X}_{n,k}\mathbf{f}_k + \mathbf{u}_n$
\end{center}
where 
\begin{itemize}
\item $\mathbf{r}_n$ is a column vector of length $n$. Each element in
  $\mathbf{r}_n$ represents the return of a security in the portfolio.
\item $\mathbf{X}_{n,k}$ is an $n$ by $k$ matrix. Each row represents
  $k$ attributes of a security. There are $n$ securities in the
  portfolio.
\item $\mathbf{f}_k$ is a column vector of length $k$. The elements
  are the estimated coefficients from the regression. Each element
  represents the \emph{factor return} of an attribute.
\item $\mathbf{u}_n$ is a column vector of length $n$ with residuals
  from the regression.
\end{itemize}

In the case of this portfolio manager, suppose that she only has three
holdings in her portfolio. $r_3$ is thus a 3 by 1 matrix with returns
of all her three holdings. The matrix $\mathbf{X}_{3,3}$ records the
score for each of the three factors (value, growth, and momentum) in
each row. $\mathbf{f}_3$ contains the estimated coefficients of a
regression $\mathbf{r}_3$ on $\mathbf{X}_{3, 3}$.

The active exposure of each of the $k$ variables, $X_{i}$, $i$
$\in$ $k$, is expressed as
\begin{center}
  $X_i = \mathbf{w}_{active}\prime \mathbf{x}_{n,i}$,
\end{center}
where $X_i$ is the value representing the active exposure of the
attribute $i$ in the portfolio, $\mathbf{w}_{active}$ is a column
vector of length $n$ containing the active weight of every security in
the portfolio, and $\mathbf{x}_{n, i}$ is a column vector of length
$n$ with attribute $i$ for all securities in the portfolio. Active
weight of a security is defined as the difference between the
portfolio weight of the security and its benchmark weight.

Using the example mentioned above, the active exposure of the
attribute \texttt{value}, $X_{value}$ is the product of
$\mathbf{w}_{active}\prime$ (containing active weight of each of the
three holdings) and $\mathbf{x}_{3}$ (containing value scores of the
three holdings).

The contribution of a variable $i$, $R_i$, is thus the product of the
factor returns for the variable $i$, $f_i$ and the active exposure of
the variable $i$, $X_i$. That is,
\begin{center}
  $R_i = f_iX_i$.
\end{center}
Continuing the example, the contribution of value is the product of
$f_{value}$ (the estimated coefficient for value from the linear
regression) and $X_{value}$ (the active exposure of value as shown
above).

Therefore, the active return of the portfolio $R_{active}$ is the sum
of contributions of all $k$ variables and the residual $u$
(a.k.a. the interaction effect),
\begin{center}
  $R_{active} = \sum\limits_{i = 1}^kR_i + u$.
\end{center}

For instance, a hypothetical portfolio has three holdings (A, B, and
C), each of which has two attributes -- size and value.

<<echo = FALSE>>=
test.df <- data.frame(Return = c(0.3, 0.4, 0.5),
                      Name = c('A', 'B', 'C'),
                      Size = c(1.2, 2, 0.8), 
                      Value = c(3, 2, 1.5),
                      Active_Weight= c(0.5, 0.1, -0.6))
test.df
## model <- lm(Ret ~ Size + Value, data = test.df)
## test.df[,1] %*% test.df[,5]
## model$coefficients[2] * t(test.df[,3]) %*% test.df[,5] ## active exposure of size 
## model$coefficients[3] * t(test.df[,4]) %*% test.df[,5] ## active exposure of value 
@ 

Following the procedure as mentioned, the factor returns for size and
value are -0.0313 and -0.1250. The active exposure of size is 0.32
and that of value is 0.80. The active return of the portfolio is -11\%
which can be decomposed into the contribution of size and that of
value based on the regression model. Size contributes 1\% of the
negative active return of the portfolio and value causes the portfolio
to lose the other 10.0\%.

\subsection{Single-Period Regression Tools}

Another conventional attribution methodolody is the regression-based
analysis. As mentioned, the \textbf{pa} package provides tools to analyze
both single-period and multi-period data frames.

<<>>=
rb.single <- regress(jan, date.var = "date",
                ret.var = "return",
                reg.var = c("sector", "growth",
                  "size"),
                benchmark.weight = "benchmark",
                portfolio.weight = "portfolio")
exposure(rb.single, var = "growth")
@

\texttt{reg.var} specifies the columns containing variables whose
contributions are to be analyzed. Calling \texttt{exposure} with a
specified \texttt{var} yields information on the exposure of both the
portfolio and the benchmark by that variable. If \texttt{var} is a
continuous variable, for instance, \texttt{growth}, the exposure will
be shown in 5 quantiles. Majority of the high \texttt{value}
securities in the portfolio in January have relatively low
\texttt{growth} scores.

<<>>=
summary(rb.single)
@

The \texttt{summary} method shows the number of securities in the
portfolio and the benchmark, and the contribution of each input
variable according to the regression-based analysis. In this case, the
portfolio made a loss of 2.91\% and the benchmark lost
4.38\%. Therefore, the portfolio outperformed the benchmark by
1.47\%. \texttt{Sector}, \texttt{growth}, and \texttt{size}
contributed 0.32\%, 0.05\%, and 0.29\%, respectively.


\subsection{Multi-Period Regression Model}

The same challenge of linking arithmetic and geometric returns is
present in multi-period regression model. We apply the optimized
linking algorithm proposed by \cite{linking} in the regression
attribution as well.

Within each period $t$, 
\begin{center}
  $R_t^{active} = \sum\limits_{i = 1}^kR_{i,t} + u_t$,
\end{center}
where $R_{i,t}$ represents the contribution of a variable $i$ of the
time period $t$ and $u_t$ is the residual in that period.

Across $T$ periods, the active return can be expressed by a product of
the optimized linking coefficient $b_t^{opt}$ and the individual
contribution of each of the $k$ attributes. The adjusted contribution
of each of the $k$ variables $i$, $\hat{R}_{i,t}$, is expressed by
\begin{center}
  $\hat{R}_{i,t} = b_t^{opt}R_{i,t}$.
\end{center}

Thus, the overall active return $R_{active}$ can be decomposed into
\begin{center}
  $R_{active} = \sum\limits_{t = 1}^T\sum\limits_{i =
    1}^k\hat{R}_{i,t} + U$,
\end{center}
where $U$ is the residual across $T$ periods.

\subsection{Multi-Period Regression Tools}
<<>>=
rb.multi <- regress(quarter, date.var = "date",
                ret.var = "return",
                reg.var = c("sector", "growth", 
                  "size"),
                benchmark.weight = "benchmark",
                portfolio.weight = "portfolio")
rb.multi
@

Regression-based analysis can be applied to a multi-period data frame
by calling the same method \texttt{regress}. By typing the name of the
class object \texttt{rb.multi} directly, a short summary of the
analysis is provided, showing the starting and ending period of the
analysis, the methodology, and the average number of securities in
both the portfolio and the benchmark.

<<>>=
summary(rb.multi)
@

The regression-based summary shows that the contribution of each input
variable in addition to the basic information on the portfolio. The
summary suggests that the active return of the portfolio in year 2010
is 1.27\%. The \texttt{Residual} number indicates the contribution
of the interaction among various variables including \texttt{sector},
\texttt{growth}, and \texttt{growth}.

Visual representation of relative performance of a portfolio against
its benchmark is best viewed across a longer time span. Here, we use
the data frame \texttt{year} for illustrative purposes. 

<<>>=
rb.multi2 <- regress(year, date.var = "date",
               ret.var = "return",
               reg.var = c("sector", "growth", 
                 "size"),
               benchmark.weight = "benchmark",
               portfolio.weight = "portfolio")
returns(rb.multi2, type = "linking")
@ 

We obtained an object \emph{rb.multi2} of class \emph{regressMulti}
based on the data set from January, 2010 through December, 2010. The
portfolio beat the benchmark by 10.1\% over this period. Based on the
regression model, \texttt{size} contributed to the lion share of the
active return.

\begin{figure}
\centering
\vspace*{.1in}
<<fig = TRUE>>=
plot(rb.multi2, var = "sector", type = "return")
@
\caption{\label{figure:regmultiattrib}
  Performance Attribution.}
\end{figure}

Figure \ref{figure:regmultiattrib} displays both the cumulative
portfolio and benchmark returns from January, 2010 through December,
2010. It suggests that the portfolio, consisted of high \texttt{value}
securities in January, consistently outperformed the benchmark in
2010. Outperformance in May and June helped the overall positive
active return in 2010 to a large extent.

\section{Brinson as Regression}

Another way to think about the analysis as \cite{brinson:gary} have
done is to consider it in the context of a regression
model. Conducting a Brinson attribution is similar to running a linear
regression without the intercept term. Estimated coefficients will
then be the mean return of each category of the attributed specified
in the universe, a.k.a. the factor return of each category. The mean
return of each category also appears in the Brinson analysis. The
equivalent to the allocation effect for the universe in the Brinson
model is the sum of the product of the estimated coefficient and the
active weight of each category. 

Using the same regression model as before,
\begin{eqnarray*}
  R_{allocation} & = & \sum\limits_{j = 1}^N W^P_jR^B_j -
  \sum\limits_{j = 1}^N W^B_jR^B_j \\ 
  & = & (\mathbf{W}^P - \mathbf{W}^B)\prime \mathbf{f},
\end{eqnarray*}
where $\mathbf{W}^P$ is a column vector indicating the portfolio
weight of each category within the attributed specified by the
manager; $\mathbf{W}^B$, a column vector indicating the benchmark
weight of each category, and $\mathbf{f}$ is the column vector which
has benchmark return of all the categories. Assuming that in this
case, the benchmark is the universe and the portfolio holdings are all
from the benchmark, $\mathbf{R^B}$ can be estimated by regressing
returns on the attribute specified by the portfolio manager:
\begin{center}
  $\mathbf{r}_n = \mathbf{X}_{n,p}\mathbf{f} + \mathbf{U}$,
\end{center}
where 
\begin{itemize}
\item $\mathbf{r}_n$ is a column vector of length $n$. Each element in
  $\mathbf{r}_n$ represents the return of a security in the portfolio.
\item $\mathbf{X}_{n, p}$ is an $n$ by $p$ matrix where $n$ refers to
  the number of securities in the portfolio and $p$ refers to the
  number of levels within the attribute specified.
\item $\mathbf{f}$ is the estimated coefficients on the regression
  without the intercept term. The estimated coefficient of each
  attribute is the mean return for each of the attribute.
\item $\mathbf{U}$ is the column vector with all the residual terms. 
\end{itemize}
  
Since $\mathbf{R}^B$ is the same as $\mathbf{f}$, the allocation
effect in the Brinson model is a special case of the regression
approach.

In order to estimate the selection effect in the Brinson model, 
one can calculate the mean return of each category within the attribute 
in both the portfolio and the benchmark under a regression framework and 
use the benchmark weights to calculate the selection effect. 

\begin{eqnarray*}
    R_{selection} & = & \sum\limits_{j = 1}^N W^B_jR^P_j -
    \sum\limits_{j = 1}^N W^B_jR^B_j \\ 
    & = & \mathbf{W^B}\prime
    (\mathbf{f^P} - \mathbf{f^B}),
\end{eqnarray*}
where $\mathbf{W^B}$ is the column vector with the benchmark weight of
each category within the attribute specified; $\mathbf{f^P}$ and
$\mathbf{f^B}$ are the column vectors indicating the mean return of
the portfolio and that of the benchmark, respectively. As mentioned
above, $\mathbf{f^P}$ and $\mathbf{f^B}$ can be estimated by running a
linear regression without the intercept term with respect to stocks in
the portfolio and benchmark separately. Hence, the selection effect in
the Brinson model can be calculated by using linear regression.

Interaction effect is the difference between a portfolio's actual
return and the sum of the allocation and selection effects.

An numerical example of showing that the Brinson model is a special
case of the regression approach is as follows.

Suppose that an equity portfolio manager has a portfolio named
\texttt{test} with the universe as the benchmark.

%% allocation effect
<<>>=
data(test)
test.br <- brinson(x = test, date.var = "date", 
                   cat.var = "sector",
                   bench.weight = "benchmark", 
                   portfolio.weight = "portfolio", 
                   ret.var = "return")
returns(test.br)
@ 

When we apply the standard single-period Brinson anaysis, we obtain an
active return of -35.9 bps which can be further decomposed into
allocation (-3.4 bps), selection (-42.5 bps), and interaction (10.1
bps).

We can also show the allocation effect by running a regression model
based on sector only.

<<>>=
test.reg <- regress(x =test, 
                    date.var = "date", 
                    ret.var = "return", 
                    reg.var = "sector",
                    benchmark.weight = "benchmark", 
                    portfolio.weight = "portfolio") 
returns(test.reg)
@ 

The contribution from sector based on the regression approach (-3.4
bps) matches the allocation effect from the Brinson model as shown above.

%% <<>>=
%% lm.1 <- lm(return ~ sector - 1, data = test)
%% exposure(br.single, var = "sector")

%% (exposure(br.single, var = "sector")[ , 1] -
%%   exposure(br.single, var = "sector")[ , 2]) %*% 
%%   lm.1$coefficients
%% @ 

However, in order to calculate the selection effect from the
regression approach, we need to apply another regression model to a
universe limited to the securities held in the portfolio. Using the
factor returns from the regress class object, \texttt{test.reg}, and
those from the linear regression, we can obtain the selection effect
(-42.5 bps) via the regression approach.

%% selection effect
<<>>=
lm.test <- lm(return ~ sector - 1, 
              data = test[test$portfolio != 0, ])
lm.test$coefficients 
exposure(br.single, var = "sector")[ ,2] %*% 
  (lm.test$coefficients - test.reg@coefficients)
@ 

\section{Conclusion}


In this paper, we describe two widely-used methods for performance
attribution -- the Brinson model and the regression-based approach,
and provide a simple collection of tools to implement these two
methods in \texttt{R} with the \texttt{pa} package. We also show that
the Brinson model is a special case of the regression method. A
comprehensive package, \texttt{portfolio} \cite{kane:david}, provides
facilities to calculate exposures and returns for equity
portfolios. It is possible to use the \texttt{pa} package based on the
output from the \texttt{portfolio} package. Further, the flexibility
of \texttt{R} itself allows users to extend and modify these packages
to suit their own needs and/or execute their preferred attribution
methodology. Before reaching that level of complexity, however,
\texttt{pa} provides a good starting point for basic performance
attribution.


\bibliography{pa}


\end{document}