Sunteți pe pagina 1din 65

Valuing Subscription-based Businesses Using

Publicly Disclosed Customer Data


Daniel M. McCarthy
Peter S. Fader
Bruce G. S. Hardie 1
May 19, 2016

1 Daniel

McCarthy is a Ph.D. Candidate in the Department of Statistics at the Wharton School of


the University of Pennsylvania (address: 400 Jon M. Huntsman Hal, 3730 Walnut Street, Philadelphia, PA 19104-6340; email: danielmc@wharton.upenn.edu, web: http://www-stat.
wharton.upenn.edu/danielmc/). Peter S. Fader is the Frances and Pei-Yuan Chia Professor of Marketing at The Wharton School of the University of Pennsylvania (address: 771 Jon M.
Huntsman Hall, 3730 Walnut Street, Philadelphia, PA 19104-6340; phone: (215) 898-1132; email:
faderp@wharton.upenn.edu). Bruce G. S. Hardie is Professor of Marketing, London Business
School (email: bhardie@london.edu; web: www.brucehardie.com). We acknowledge the
financial support of the Baker Retailing Center.

Abstract
Valuing Subscription-based Businesses Using Publicly Disclosed Customer Data
The growth of subscription-based commerce has changed the types of data firms report to external shareholders. More than ever before, companies are discussing/disclosing data such as
churn, customers acquired, customer lifetime value, and more. As such, there is increasing
interest in linking the value of a firms customers to the overall value of the firm. Although
a number of researchers in the fields of marketing and accounting have explored this idea,
previous work did not drive customer-based valuation estimates off of the widely accepted discounted cash flow (DCF) valuation model, nor did the underlying models of customer acquisition and retention fully reflect all of the empirical realities associated with these behaviors.
We develop a framework for valuing subscription-based firms that addresses both of these
issues, incorporating critical factors such as heterogeneity, duration dependence, seasonality,
and macroeconomic variables into a customer-based DCF model. Furthermore, our framework
explicitly acknowledges that publicly disclosed data is overly aggregated, with missingness.
We apply this methodology to data for Dish Network and Sirius XM Holdings, estimating
the value of the firms, analyzing customers unit economics, performing customer segmentation, and developing a new investor metric, the internal rate of return on a customer (IRRC).
Keywords: customer lifetime value; customer equity; valuation; shareholder value;
DCF

Introduction

The relevance and popularity of subscription-based businesses businesses whose customers


pay a periodically recurring fee for access to a product or service has grown considerably
in recent years. Previously dominated by newspapers, magazines, and telecommunications
companies, the subscription-based business model has made strong inroads into consumer software (Microsoft 365), food preparation (Blue Apron), health and beauty products (Dollar Shave
Club), and a large array of subscription-based software-as-a-service (SaaS) enterprises in the
B2B space, as businesses look to increase the predictability of their revenue streams. Many
experts have written in depth about this topic (Baxter (2015), Janzer (2015), Warrillow (2015)).
The future looks bright for example, Gartner predicts the global SaaS market may top $51
billion in 2018, more than double the $22.6B it had generated in 2013 (Columbus (2014)).
The increased popularity of subscription-based businesses has brought with it an increase in
the relevance, usage, and disclosure of customer data, including but not limited to churn/attrition,
customer/subscriber acquisition cost (CAC or SAC), average revenue per user (ARPU), and
customer lifetime value (CLV). An SEC text mining analysis performed by the authors shows
that 193 companies within the telecommunications industry alone (companies with SIC codes
4812 and 4813) have disclosed customer data in quarterly and annual filings. ConstantContact, a popular email marketing software firm, mentions CLV as a key growth driver in investor
presentations and even provided an estimate of CLV per user in its 2013 annual report: We currently estimate a customer lifetime-value of approximately $950 per unique customer. Similar
customer data references can be found in the filings of many other companies, including but not
limited to Avaya, Boingo Wireless, Vonage, Zipcar, and Shutterstock.
Customer data is important to investors and is being used by analysts as they make their recommendations. For example, analyst reports for video streaming/online rental business Netflix
from Thomas Weisel Partners, Vintage Research, First Albany Capital and Delafield Hambrecht
all strongly emphasize customer data in general (and the size of the total subscriber base over
time in particular) when justifying their investment recommendations (Stanford (2004)). The

2
price of a stock reflects and incorporates investors beliefs regarding the future cash flows a
company will generate, and for the subscription-based businesses we consider here, the primary source of future cash flows is customers. Customer data allows analysts to create more
grounded, accurate future cash flow forecasts.
The pioneering work of Gupta et al. (2004) (hereafter, GLS) was the first to explicitly link
firm value to CLV for public companies. However, treatment of the valuation problem by GLS
suffers from two major issues. First is that GLSs CLV calculations are performed assuming
a constant retention rate, which leads to pessimistic customer valuations, often dramatically
so (Fader and Hardie (2010)). The second issue is that GLSs valuation framework does not
incorporate key financial/accounting issues such as firm capital structure and non-operating
assets. Its practical applicability was further diminished by a reliance on external data sources
for key inputs for the few companies that did report customer data.
Many other authors have built upon GLSs seminal work, most notably Schulze et al. (2012,
SSW hereafter), whose work substantially improves upon the aforementioned finance/accounting
valuation aspects (see also Bauer and Hammerschmidt (2005)). However as we will discuss below, subsequent work still assumes a homogeneous retention rate, does not incorporate seasonality and macroeconomic conditions, does not address left censoring and missing data,
assumes quarterly cash flows while the flow of new customers is continuous and the contracts/subscriptions are almost always non-quarterly, and often does not distinguish between
contractual and non-contractual businesses (Fader and Hardie (2009)). Even though many authors have discussed DCF models in CBV papers, no extant models adopt the DCF model into
their financial valuation approach. Instead, they take a purely customer-based valuation (CBV)
approach best described/defined by Skiera and Schulze (2014) (discussed further below), and
they contrast CBV to DCF, treating CBV and DCF as mutually exclusive approaches to corporate valuation.
We show that CBV and DCF are complementary to one another. Our proposed model welds
the two approaches together into one customer-based DCF (CBDCF) model. DCF valuation

3
is the de facto standard valuation model within finance for good reason, and we believe that
allowing CBV models to benefit from the natural conveniences/advantages of DCF models is
an important contribution of our work.
We build a more realistic, accurate, and generally applicable valuation model using only
publicly disclosed customer data, making it suitable for passive external shareholders. If desired, company management could apply the same model with little modification using granular transaction-level data available within the organization. We build a stochastic model for
the acquisition and retention of customers over time, which parsimoniously incorporates factors that are essential to the dynamics of both processes heterogeneous retention propensities,
duration dependence, seasonality, macroeconomic conditions and household formation and
acknowledges the fact that our data is aggregated temporally and across customers (i.e., we
are only provided with quarterly company summaries) with missingness (i.e., all customer data
is not available for all periods). We summarize and compare our framework for analysis and
contributions to previous literature in Table 1 of the literature review in Section 2. We evaluate
the predictive performance of the proposed model by performing rolling 2-year predictions of
future customer metrics over all possible calibration periods. We perform this rolling validation
analysis for our CBDCF model, as well as GLS and SSWs models.
With this general-purpose model in hand, we estimate many managerially/practically important quantities, including the value of the firm, and the value/profitability of a newly acquired
customer (and how this compares to a long-time customer). We also introduce a new investment metric, the internal rate of return on a customer (IRRC), which is a scale-invariant measure
of the marginal profitability of customer acquisition spending. The public company real data
examples we study are Dish Network (Dish), a mature pay-TV company, and Sirius XM, a highgrowth satellite radio service provider. We develop our entire forecasting/valuation approach
for operating assets using only publicly disclosed data from both companies, without any inputs
or assumptions from outside sources.
We hope that this work drives increased adoption of customer-based valuation models, of-

4
fering marketing executives a great step forward to be able to coordinate with (and earn greater
respect from) their finance counterparts and other leaders across the organization.

Literature Review

A central concept underlying customer valuation is customer equity (CE, Blattberg and Deighton
(1996)). An excellent discussion of CE can be found in Kumar and Shah (2015). Customer equity is the net present value of the remaining (or residual) lifetime value (RLV) of all current
customers, also known as current customer equity or CCE (see Pfeifer (2011)), plus the CLV of
all future expected customers.
Wiesel et al. (2008) advocates reporting CE-related metrics within public company filings,
calling CE an integral part of financial reporting. GLS sparked widespread interest in the idea
of valuing companies from the bottom-up through CE. GLS added to its impactful legacy
through a series of subsequent publications (Gupta et al. (2006), Gupta and Zeithaml (2006),
Gupta and Lehmann (2005), Gupta (2009), Gupta and Lehmann (2006), Berger et al. (2006)).
However GLS was more of a proof of concept than an accurate, generally applicable valuation
framework. Its valuation estimates relied upon estimates from experts. Since then, there have
been a number of papers that have extended this valuation framework.
SSW provides thoughtful and accurate treatment of how CE relates to firm value using financial valuation theory in an empirical setting, considering/incorporating financial aspects such as
firm capital structure and non-operating assets. Kumar and Shah (2009) relate CE to public company shareholder value (SHV) through a regression requiring internal, individual-level
transaction data and covariates. However, all companies valued using only publicly disclosed
company data assume that all customers share one common time-invariant per-quarter propensity to churn, and when applicable, model future customer acquisition via a Bass diffusion
model or a simple variant.
A number of more theoretical papers have discussed other considerations without providing

5
explicit valuation estimates for real companies. Pfeifer (2011) shows how one must be careful with the timing of cash flows when estimating retention rates and current-customer equity
(CCE) using publicly disclosed company data, but stopped short of its own estimate of CCE,
let alone a value for the firm as a whole. Pfeifer echoes Bauer and Hammerschmidt (2005),
discussing the merits of a DCF approach like the one we propose but do not pursue the idea
further. Pfeifer and Farris (2004) study the theoretical effect of a homogeneous but duration
dependent retention rate upon customer valuation.
Skiera and Schulze (2014) compare and contrast DCF modeling with CBV modeling. They
offer an excellent discussion and reconciliation of the two valuation methods, but state that DCF
and CBV are fundamentally different, the biggest difference between them being whether the
cash flows are being modeled by time period or by customer. We believe that this conclusion is
more polarizing than it needs to be. Our proposed method marries DCF with CBV. We employ
a DCF model that is driven off of the same customer projections that are made in extant CBV
models a customer-based DCF model. GLS and SSW correctly note that customer-based
valuation methods are useful for valuing high-growth firms; however this does not necessarily
mean that DCF models are not useful in these cases. In our empirical analysis of a highgrowth firm (Sirius XM), we show that our customer-based valuation model performs very
well because, even though it is a DCF approach, it carefully accounts for customer acquisition,
retention, and profitability. We summarize the previous literature and our contributions to it in
Table 1.
It is encouraging to see that these ideas have been gaining attention and respect outside of
marketing. Within the accounting literature, for instance, Bonacchi et al. (2015) provides a
systematic analysis across multiple companies linking CCE as a metric to shareholder value
(further discussion is provided in Hand (2015)). Other work includes Andon et al. (2001),
Bonacchi et al. (2008), Bonacchi and Perego (2012), Boyce (2000), and Gourio and Rudanko
(2014).
With this large and growing base of literature on the importance of (and methods for) cus-

Gupta, Lehmann, and Stuart (2004)

Public
+ Experts
Wiesel, Skiera, and Villanueva (2008)
Public
Kumar and Shah (2009)
Private
Fader and Hardie (2010)
None
Pfeifer (2011)
None
Schulze, Skiera, and Wiesel (2012)
Public
Bonacchi, Kolev, and Lev (2015)
Public
Proposed
Public

Paper

No
No
Yes
No
No
No
Yes

No
No
No
No
No
No
Yes

Missing
Data

No

Heterog
eneity

No

ce
ependen
Duratio
nD

No
No
No
No
No
No
Yes

No

Covaria
tes

No
Yes
No
No
No
No
Yes

No

No
No
No
No
No
No
Yes

No

DCF M
ode

Valuation Elements
mated

No
No
No
No
Yes
Yes
Yes

No

WACC
Esti

Customer Dynamics

No
No
No
No
Yes
No
Yes

No

D
e
b
t
/
O
ther As
sets

Setting

urce
Data So

rs Value
d

Current + Future
Current
Current
Current
Current + Future
Current
Current + Future

Current + Future

Custom
e

Table 1: Literature Overview

7
tomer valuation, combined with ever-improving customer-level data disclosures by public companies, it is important to step back and develop a comprehensive customer-valuation model
which will attract the attention of the C-suite.
We structure the remainder of the paper as follows: in the next section, we describe a model
that enables us to forecast future customer acquisitions and losses which are then incorporated
into a standard DCF valuation framework. We then provide an empirical analysis that explores
how such a model can be fit to real public company data, validating the models goodness of fit
through a rolling validation analysis and comparison to alternative valuation specifications. We
spend the remainder of the paper discussing the sort of inferences that we can draw from the
model, including the profitability of new customers, the importance of customer tenure on the
residual value of the customer, and we develop a new investment performance metric called the
internal rate of return on a customer (IRRC). We conclude with a discussion of the results and
future work.

3
3.1

Model Development
Valuation 101

According to standard corporate valuation theory (Damodaran (2012), Greenwald et al. (2004),
Holthausen and Zmijewski (2013), Koller et al. (2010)), SHV is equal to the value of the operating assets (OA) plus the non-operating assets (NOA), minus the net debt (ND) of the firm (all
figures below assume a quarterly clock, consistent with firms external reporting period for
the financial data below):
SHVq = OAq + NOAq NDq ,

(1)

where q = 1 represents the first quarter of the companys commercial operations. Equation 1
reiterates SSW Equation 1. The primary elements of Equation 1 will be defined below, with
elements either of secondary importance to company valuation or directly available through

8
public sources elaborated upon in Web Appendix B, where we will also provide a detailed
summary of where these data elements may be obtained. The value of a firms operating assets
(OA) is equal to the sum of all future expected free cash flows (E(FCF)s) the firm will generate,
discounted at the weighted average cost of capital (WACC, assuming quarterly compounding):

X
E(FCF)q+q0
.
OAq =
q0
(1
+
WACC)
0
q =0

(2)

As noted in SSW, WACC is the appropriate firm-specifiic discount rate to apply to future free
cash flows.
We elaborate upon the expected free cash flows (E(FCF)) in the numerator of Equation 2
while WACC is discussed further in Web Appendix B. E(FCF) is equal to the net operating profit
after taxes (NOPAT) minus the difference between capital expenditures (CAPEX) and depreciation and amortization (D&A), minus the change in non-financial working capital (NFWC):

E(FCF)q = NOPATq (CAPEXq D&Aq ) NFWCq .

(3)

The most important ingredient of E(FCF) is NOPAT, which is a measure of the underlying
profitability of the operating assets of the firm. NOPAT is equal to revenues (REV) after variable
costs (VC%), minus fixed operating costs (FC), after taxes:

NOPATq = (REVq (1 VC%q ) FCq ) (1 TRq ),

(4)

where TR is the corporate tax rate for the firm. The other elements of Equation 3 make adjustments for balance sheet-related cash flow effects are generally of secondary importance to the
value of the firm.
The methodology summarized above is a discounted cash flow (DCF) model, which is the
de-facto industry standard way in which operating assets are valued within the financial community. While we are not the first to discuss DCF modeling (in particular, see Schulze and

9
Skiera in Skiera and Schulze (2014)), we are the first to build a CBV model on top of a DCF
model. We show how to do so in the next section.

3.1.1

Projecting Revenues with Customer Metrics

Customer data very naturally enters the DCF framework presented in Section 3.1, primarily
by providing further structure to REV within Equation 4. For subscription-based businesses,
revenue in a given time period must be equal to the average number of active users during the
period multiplied by the average revenue per user during the period. We operationalize our
revenue model by specifying sub-models for both processes the number of active users at the
end of each period (END), and average revenue per user (ARPU).
Subscription-based firms disclose the number of active users at the end of each quarter. Let
ENDq be the number of active users q quarters after the company begins commercial operations. Then ENDq is, by construction, equal to the cumulative historical number of customer
acquisitions (ADD), less the cumulative historical number of customer losses (LOSS), during
each prior quarter:
ENDq =

q
X

ADDi LOSSi ,

(5)

i=1

Therefore, our customer-based DCF entails constructing models for LOSS, ADD, and ARPU
using only publicly available data from which we can generate forecasts of REV. As we will
discuss at more length, valid modeling and projection of END require proper specification of
the acquisition of new customers, and the retention of customers once they have been acquired.
With more accurate revenue projections, the valuation estimates which result when we complete our DCF analysis must be improved as well. This is important, because coming up with
an appropriate model for future revenues is usually the most important and most challenging
valuation task analysts must perform. Before we can come up with these more accurate revenue
forecasts, however, we must first understand the data we need and how it compares to the data
which firms actually provide. We explore this in the next section.

3.2

Underlying Data Structure for Subscription-Based Businesses

10

In this section, we illustrate the data structure that firms with a subscription/contractual business
model should be reporting internally to account for their customer bases, and how it maps to
the data firms are increasingly providing publicly.
Let us assume the internal reporting occurs on a monthly cycle, and for ease of notation,
that the quarters always align with the months (i.e., that the first month of each quarter is
either January, April, July or October), and let C(m, m0 ) be the number of customers acquired in month m who are still active in month m0 , where m0 m. Then C(m, m) is the
number of customers acquired in month m. This gives us the upper-triangular matrix shown
in Figure 1. Each row is called a cohort, the group of customers born at the same time.
C(m, m0 )/C(m, m) gives us the empirical survival curve for cohort m as it represents the
proportion of the customers acquired in month m who are still active in month m0 (hence,
C(m, m0 ) C(m, m0 + m) m = 1, 2, . . .). The total number of customers the firm
has in month m, C(., m), is the number of customers acquired in month one who are still customers, plus the number of customers acquired in month two who are still customers, ..., plus
the number of customers acquired in month m:

C(., m) = C(1, m) + C(2, m) + + C(m, m).

(6)

While this matrix will be reported internally in more enlightened firms, firms do not release
this data to the public. When firms disclose customer data at all, they report quarterly summaries
the number of customers active at the end of each quarter (ENDq in Equation 5), and the
number of customers added and lost each quarter (ADDq and LOSSq in Equation 5). Publicly
disclosed customer data thus suffers from aggregation, because we observe quarterly summaries
when customers make acquisition and retention decisions at a higher temporal frequency.
We first explore how this upper triangular matrix maps to the reported numbers and then look
at how we can attempt to reconstruct and project it into the future from the publicly reported

11

Figure 1: Upper Triangular Cohort Framework for Customers

data. Assume that external reports are provided quarterly. It follows from Equation 5 that
quarter customer additions and total ending customers are equal to

ADDq = C(3q 2, 3q 2) + C(3q 1, 3q 1) + C(3q, 3q),


ENDq = C(., 3q).

and

(7)
(8)

We illustrate this mapping visually in Figure 1. Let L(m) denote the number of customers lost
in month m. Then

L(m) = C(., m 1) + C(m, m) C(., m),


LOSSq = L(3q 2) + L(3q 1) + L(3q).

and

(9)
(10)

Further complicating our modeling efforts is data missingness/left censoring quarterly


summary data is not available back to the start of commercial operations. For example, while
Sirius XM began commercial operations in Q4 2001, it started disclosing paying customer data

12
in Q3 2008. Similarly, while Netflix began commercial operations in calendar Q4 1999, they
started disclosing customer data in Q2 2001. Amazon, Ameritrade, Capital One, eBay and
E*Trade began commercial operations in 1994, 1971, 1993, 1995, and 1982, respectively, well
before GLS began modeling their respective customer data in 1996-1997. To properly handle
missing data, we must backcast the initial datapoints in addition to the in-filling we must
do to impute the monthly cohort-level numbers in the upper triangular matrix shown in Figure
1 from the quarterly summary figures which are disclosed publicly.
Our overarching modeling objective, then, is to take the aggregated, left-censored customer
addition and loss data that firms have made available, and use it to forecast into the indefinite
future the upper triangular matrix shown in Figure 1. This requires us to have well-specified
models of both customer acquisition and retention. Furtheremore, this matrix simply provides
us with the number of customers active at the end of every period, and not their spend. Therefore, we also need to develop a model for ARPU. We specify and estimate models for customer
retention, customer acqusition, and ARPU over the next four sections.

3.3

Retention Process

As noted above, C(m, m0 )/C(m, m) for m0 m gives us the empirical survival curve for
cohort m. Letting SR (m0 m|m) denote the probability that a customer acquired in month m
survives beyond month m0 ,

C(m, m0 ) = C(m, m) SR (m0 m|m).

(11)

Therefore, given knowledge of C(m, m), modeling and predicting C(m, m0 ) is a well-studied
problem. Our objective is to specify an accurate yet parsimonious survival model for the duration of customers tenure with the firm.
We want to capture the effects of heterogeneity and duration dependence, and incorporate
time-varying covariates for seasonality and macroeconomic conditions. As we will see later,

13
each of these factors has a significant effect upon the propensity to churn. Therefore we use a
Weibull baseline with covariates incorporated using proportional hazards, with cross-sectional
heterogeneity in the baseline propensity characterized by a gamma distribution. This is a wellaccepted model for duration data and has been proven to be quite effective and robust in numerous previous studies (Morrison and Schmittlein (1980), Moe and Fader (2002), Fader and
Hardie (2009), Schweidel et al. (2008b)).
Let SR (m0 m|m, , cR , X(m0 ), R ) denote the probability that a customer acquired in
month m will survive for at least m0 m more months (i.e., he/she survives beyond month
m0 ), given his/her individual-specific baseline propensity to churn (R ), homogeneous retention shape parameter (cR ), time-varying retention covariates up to and including month m0
(XR (m0 ) = [xR (1), xR (2), . . . , xR (m0 )]), and coefficients for retention covariates ( R ). Then
SR (m0 m|m, R , cR , X(m0 ), R ) = exp(R BR (m, m0 m)), where
(12)
0
m
X
T
BR (m, m0 m) =
[(i m)cR (i m 1)cR ]eR xR (i) . (13)
i=m+1

Following Schweidel et al. (2008b) and Jamal and Bucklin (2006), we expect cR > 1 if it is
significantly different from 1. When cR = 1, it reduces to an exponential baseline.
Assuming R is gamma(rR , R )-distributed across the population, the unconditional probability that a customer acquired in month m survives past month m0 is

S(m0 m|m, R , cR , X(m0 ), R )f (R |rR , R )dR


0
rR
R
=
(14)
R + BR (m, m0 m)

SR (m m|m, cR , X(m ), R , rR , R ) =

Plugging the survival curve in Equation 14 into Equation 11 allows us to predict the number
of active customers in future months for the month m cohort. Therefore, if we know C(m, m)
over all months, we can predict the remainder of the upper triangular matrix in Equation 1.
Prior to this work, the only retention model that has been used in empirical customer-based

14
corporate valuation work has been a constant geometric model. That is, all customers have
been assumed to share one common retention rate per quarter, r. Fader and Hardie (2010)
shows theoretically and by simulation that ignoring time dynamics in retention propensities
systematically undervalues the customer base.
Equation 14 also highlights how much the complexity of our modeling of losses increases
when we move away from a constant-retention world. In a constant-retention world, our estimate of the number of customer losses in a particular month m, L(m) from Equation 9, only
requires knowledge of the total number of customers from the end of the previous period. Assuming a constant monthly retention rate r,

b 0 ) = C(., m0 1) r.
L(m

(15)

In contrast, under our proposed model, the number of customers lost in month m0 requires
knowledge of the number of customers acquired in all months preceding month m0 :

b )=
L(m

0 1
m
X

C(m, m)[SR (m0 m 1|m, cR , XR (m0 ), R ) SR (m0 m|m, cR , XR (m0 ), R )] .

m=1

(16)
While this may be inconvenient, we will show that the cost of assuming a constant retention
rate can be much worse.
Estimates of LOSSq and ENDq follow from Equations 7 and 10 for all q = 1, 2, . . .. We
may then estimate the model parameters (cR , rR , R , R ) by minimizing the sum of squared
differences between our model-based estimates of LOSSq and the reported numbers. However
this assumes we know monthly customer additions C(m, m) for all m, which we do not know.
We only have quarterly customer additions ADDq and some of these observations may be missing. We therefore need to develop a model of the acquisition process whose parameters can be
estimated using the reported ADDq data. Using this model in conjunction with our retention
model, we can backcast and in-fill the C(m, m) numbers needed for estimating the retention
model. We may then forecast C(m, m) and C(m, m0 ) for m0 > m into the future. We fully

15
specify the acquisition process in the next section.

3.4

Acquisition Process

At first glance, developing a model for the acquisition of customers over time seems to be a
relatively simple exercise. Given a fixed population size, we can use a standard model for the
time to adoption such as the Bass model (Bass (1969)), or a simplified version of it, as in GLS.
However there are four problems with this:
1. It assumes that all churning customers disappear forever once an acquired customer has
churned, he/she cannot re-enter the pool of potential adopters once again. Libai et al.
(2009) extend the basic Bass model to allow for lost customers re-entering the pool of
potential adopters (but they assume a constant retention rate).
2. It assumes the population size is fixed, when we know that the number of potential customers is increasing over time due to household formation / population growth.
3. The Bass model and its simplified variants have a number of unfavorable properties, most
notably the fact that the resulting acquisition curve must be symmetric about the period
of peak acquisition. In real datasets, skewness about the peak is almost always present.
4. It ignores the effects of seasonality and macro-economic events.
We develop a new model from first principles which addresses these issues. Each month
m, a new prospect pool forms, the size of which is M (m). We set M (0) as the initial time 0
population size when the firm first commences operations (P OP (0)), who may then be acquired
as customers in month 1 and thereafter. M (1) is the growth in population by the end of the first
month of operations:
M (1) = P OP (1) P OP (0).
In general, the month m prospect pool, M (m), is equal to the growth in the population during

16
the month, plus the number of customers who churned in the previous month:

M (m) = P OP (m) P OP (m 1) + L(m 1),

m = 1, 2, . . .

Our model assumes that population growth is the only source of potential adopter growth over
time, aside from previously churned customers.1
Once a prospect pool has formed, some time will elapse until individuals within that pool
are acquired as customers. This is a timing decision for which we specify a hazard model. Let
FA (m0 m|m) denote the probability that a customer acquired in month m will be active by
the end of month m0 . Then the number of adopters in month m is equal to

C(m, m) =

m1
X

M (i) (FA (m i|i, xA (m), ) FA (m i 1|i, xA (m 1), )) . (17)

i=0

Our goal, then, is to specify an accurate yet parsimonious timing model for the duration of
time until a prospect becomes a customer. We model the time to adoption of individuals within
each prospect pool using a split-hazard model. Some individuals will never be acquired, but
for those who will be eventually, the duration of time until acquisition is characterized by a
Weibull baseline with covariate effects incorporated using the proportional hazard approach,
and cross-sectional heterogeneity in baseline propensity characterized by a gamma distribution.
Let FA (m0 m|m, , cR , X(m0 ), A , N A ) denote the probability that a customer from prospect
pool m will by the end of month m0 , m = 0, 1, ..., given his/her individual-specific baseline
propensity to be acquired (A ), homogeneous acquisition shape parameter (cA ), homogeneous
but possibly time varying acquisition covariates up to month m0 (XA (m0 )), coefficients for
acquisition covariates ( A ), and probability of never being acquired (N A ). Then
FA (m0 m|m, A , cA , xA (m0 ), A , N A ) = (1 N A )(1 exp(A BA (m, m0 m))), (18)
1
While this implies that product awareness is not changing over time, we do not have sufficient data to account
for such factors.

17

BA (m, m0 m) =

m
X

[(i m)cA (i m 1)cA ]e

T
A xA (i)

(19)

i=m+1

Assuming A is distributed across the population according to a gamma(rA , A ) distribution,


the unconditional probability that a customer from prospect pool m will be acquired by the end
of month m0 is


FA (m m|m, cA , XA (m ), A , rA , A , N A ) = (1N A ) 1
0

A
A + BA (m, m0 m)

rA 
(20)

This acquisition model is flexible yet quite parsimonious. Parsimony is especially important
in our limited data situation because, as was shown in Van den Bulte and Lilien (1997), illconditioning is a serious enough problem with small sample sizes that adding new predictors
to alleviate model misspecification concerns may make the resulting model fit (and forecast)
worse than it had been prior to the introduction of those covariates.

3.5

Parameter Estimation for Acquisition and Retention Processes

We estimate the above two models jointly using nonlinear least squares (Srinivasan and Mason
(1986)), minimizing the sum of squared differences between actual quarterly acquisitions and
losses during each period versus what we expected from our model via difference in CDFs.
Denoting the acquisition and retention parameters collectively by , so that

(rA , A , cA , A , N A , rR , R , cR , R ),

and letting Q be the number of quarters after the company commences commercial operations
to the end of the calibration period, if we observe the process from q = 1 onwards, our estimated

18
parameters are equal to
= argmin SSEF U LL (),

where

(21)

SSEF U LL S1 + S2F U LL ,
[ Q )2 ,
S1 (ENDQ END

and

(22)

S2F U LL

\ q )2 ,
[ q )2 + (LOSSq LOSS
(ADDq ADD

where

(23)

q=1

[q
ADD

3q
X

b
C(m,
m),

and

(24)

m=3q2

\q
LOSS

3q
X

b
L(m,
m).

(25)

m=3q2

b
b
Monthly customer additions and losses, C(m,
m) and L(m,
m) in Equations 24 and 25 respectively, are obtained using Equations 16 and 11. The size of the monthly prospect pools, M (m)
in Equation 11, are equal to

M (m) =

P OP (m),

m = 0,

\
P OP (m) P OP (m 1) + LOSS(m
1),

m = 1, 2, . . . , 3Q

(26)

We optimize over all parameters jointly because of the dependence of the acquisition process
upon the retention process (i.e., customers cannot churn until they have been acquired) and vice
versa (i.e., churning customers enter future prospect pools).
Equation 21 assumes that there is no missingness when in practice there often is missingness. Usually, not all data is available back to the beginning of commercial operations of the
company. Most companies begin disclosing total ending customer count some number of quarters after the company begins operations (call it qA ), then begin disclosing customer addition
and loss data in another later quarter number (call it qB , where qB qA ). In cases such as these,

19
we must replace Equation 21 with Equation 27:
= argmin SSEM ISS (),

SSEM ISS S1 + S2M ISS + S3 , where


Q
X
\ q )2 ,
[ q )2 + (LOSSq LOSS
(ADDq ADD
S2M ISS

(27)
(28)
(29)

q=
qB
qB 1

S3

X 

2
[
[
(ENDq ENDq1 ) (ENDq ENDq1 ) ,

(30)

q=
qA +1

and S1 is defined as before, in Equation 22. This accounts for the shortened contiguous customer addition and loss data through S2M ISS in Equation 29, and the missingness present at the
beginning of the time series through S3 in Equation 30.
The parameter estimates thus obtained fill in the upper triangular matrix shown in Figure
1 and forecast the evolution of the total active customer base into the indefinite future. The final
element needed to forecast future monthly revenues is average revenue per user (ARPU).

3.6

Average Revenue Per User

With estimates for the number of customers at the end of every month in hand, we need a model
for the profitability of those customers over time. Prior literature has assumed a constant dollar
profit margin per user, ignoring ARPU. This is problematic because customer profitability is
rarely constant over time and is often very noisy, whereas revenue per user is stable and reliable.
Therefore, instead of modeling and projecting customer margin directly, we model ARPU and
then consider margin separately.
If we had monthly ARPU in month m (ARPU(m)), we could easily predict revenue in

20
month m (R(m)) by multiplying ARPU(m) by the average number of users within month m2 :

R(m) =

C(., m 1) + C(., m)
2


ARPU(m).

(31)

Therefore, we need to model and project ARPU(m). While companies often publicly disclose
quarterly ARPU data, this data cannot be used in general because there are no well-accepted
standards for calculating ARPU. As Dish stated in its 2014 annual filing, We are not aware
of any uniform standards for calculating ARPU and believe presentations of ARPU may not be
calculated consistently by other companies in the same or similar businesses. Because there
is no standard definition of ARPU, different firms may have different definitions for it. This is
problematic because it implies revenue (the numerator of ARPU) often excludes certain sources
of revenue and thus may not be representative of all revenue derived from the customer base.
While revenue numbers are reliable, they are only provided quarterly, so we need to impute
monthly revenues. Assume without loss of generality that month m is within quarter q. Then
revenue in month m is equal to the customer-weighted total revenue in quarter q:

R(m) =

C(., m 1) + C(., m)
REVq ,
C(., 3q 3) + 2C(., 3q 2) + 2C(., 3q 1) + C(., 3q)

m {3q2, 3q1, 3q}


(32)

where C(., m) is defined as in Section 3.2 3 . After imputing R(m), we obtain ARPU(m) by
solving for ARPU(m) in Equation 31, given R(m) and C(., m).
Next, we apply time series methods to forecast ARPU(m). If we assume that pricing grows
at a fixed growth rate over time, time-trend regression would be appropriate, positing that

log(ARPU(m)) = 0 + 1 m + (m),
2

(m) N (0, 2 ).

(33)

Normally, one may consider multiplying ARPU(m) by C(., m) to obtain R(m). However this would not be
appropriate because ARPU is equal to total revenues divided by the average number of users within the period, not
total revenues divided by the number of users at the end of the period.
3
This calculation ignores changes in pricing within a particular quarter because monthly within-quarter changes
are effectively zero (rounding error amounts to less than 0.5%).

21
Had pricing growth been linear in nature, we would use the same regression above with respect
to ARPU(m) instead of log(ARPU(m)). However it is well known that in many economic and
financial time series, non-stationarity in the mean of the process is often present (Zivot and
Wang (2007); i.e., the level of macroeconomic aggregates, as well as product/service prices
like those considered here). In cases such as these, the fitted residuals of the regression shown
in Equation 33 may often fail tests for non-stationarity (Dickey and Fuller (1979)), the most
popular of which is the augmented Dickey-Fuller test (Elliott et al. (1996)). If this is the case,
estimates provided by the regression shown in Equation 33 are invalid, and we should use an
ARIMA(0,1,0) model instead:

log ARPU(m) = log ARPU(m 1) + 0 + (m),

(m) N (0, 2 ).

(34)

We estimate the coefficients (0 , 1 , 2 ) of our regression (either Equation 33 or Equation 34)


via maximum likelihood, which are then used to project ARPU(m) into the indefinite future.
Not included in our proposed model for ARPU is an estimate of heterogeneity in spend
propensity across the customer base. In addition to the fact that spend heterogeneity is not
identifiable from our sparse, aggregated data, ignoring spend heterogeneity does not affect our
resulting revenue point estimates, whereas incorporating retention-specific dynamics into the
retention model does affect the expected customer lifetime, often dramatically so (Fader and
Hardie (2010)). Furthermore, heterogeneity in spend across customers is generally tame in
contractual settings (and in the companies considered in our empirical analyses, in particular).
It is for these reasons that we believe it is important to incorporate heterogeneity in retention
but not ARPU.

3.7

Bringing It All Together

Taking a step back, our goal since Section 3.1.1 has been to improve the accuracy of revenue
forecasts (and thus valuation) by decomposing revenue into total users and ARPU, modeling

22
both components precisely, then bringing those components back together again to make revenue forecasts. We now have the ingredients necessary to make these forecasts:
1. Obtain monthly total users over time, C(., m) (Equation 6), by projecting out the upper
triangular matrix shown in Figure 1 using the parameters estimated from Section 3.5.
2. Obtain monthly ARPU over time, ARPU(m), via time series forecasting off of imputed
historical monthly ARPU via Equation 32 or 33, or a linear variant.
3. Given C(., m) and ARPU(m), obtain monthly revenues, R(m), using Equation 31.
With revenues estimated, the remainder of our valuation model is for all intents and purposes
the same as what a financial professional would do when building a DCF model. In the next
section, we bring this valuation model to life from start to finish with two public companies.

Empirical Analyses

To demonstrate the model and its use for valuation and financial forecasting, we first apply it
to Dish Network Corporation (Nasdaq: DISH), a large pay-TV service provider. We estimate
the parameters of the model, evaluate in-sample fit, and value Dishs shareholders equity. To
evaluate the predictive validity of the model, we perform rolling 2-year-ahead forecasts over
all possible calibration periods and compare our performance to those of alternative methods.
To further establish the robustness of the proposed model, we then show that the same model
provides a well-calibrated valuation for a second publicly-traded company, Sirius XM Holdings
(Nasdaq: SIRI), a satellite radio service provider.

4.1

Dish Network

23

Dish commenced operations in March 1996 (DISH (2014))4 . End-of-period customer counts
began being disclosed when Dish commenced commercial operations in Q1 1996. However
gross customer acquisition data is left censored the first time that gross customer additions
were disclosed was 7 quarters later, in Q1 1998. Therefore, qA = 0 while qB = 7 in our
estimation procedure, as per Equation 27.
All historical customer data (ADDq , LOSSq and ENDq , using the notation in Equation 21
and throughout) as well as quarterly subscriber revenues (REVq from Equation 4) is disclosed
in Dishs quarterly and annual reports, Forms 10-Q and 10-K, respectively. We model this
customer data up through and including Q1 2015.
The vast majority, but not all, of Dishs revenues come from its subscriber base. In Q1 2015,
0.9% of Dish sales were derived from equipment sales, which are not core to the business and
have not been growing over time. In general, customer-based corporate valuation is less useful
the larger the proportion of sales coming from non-subscriber sources happens to be.5
We incorporate 4 time-varying covariates into our acquisition and retention processes through
proportional hazards 3 quarterly dummy variables to capture seasonal fluctuations in the
propensity to acquire and retain services, and a Great Recession dummy variable to account
for the diminished propensity to acquire services and inreased propensity to drop services during that recession6 . New monthly prospect pools (P OP (m) in Equation 26) are driven off of
US household growth, provided by the Census Bureaus CPS/HVS data tables. Customer metrics and revenues are provided in the Web Appendix, and all other data is available from the
authors upon request.
4

While Dish Network was technically incorporated in 1980, the relevant starting date for our customer acquisition/retention models is when Dish actually commenced commercial operations and could thus begin acquiring
customers.
5
Dish has made investments in wireless spectrum over the past three years. Wireless spectrum is a nonoperating asset the company earns no revenue from it, and the core operations of the business do not depend
upon it. We will discuss our handling of the value of this investment (and of non-operating assets in general) in the
Web Appendix.
6
As per the Bureau of Labor Statistics, the recession began in Q4 2007 and ended in Q2 2009.

24
Many public companies summarize acquisition and retention spending in one category, usually denoted as marketing or sales and marketing expenses. This would require modelers
to estimate subscriber acquisition expenses. In contrast, because Dish publicly discloses subscriber acquisition expense, no such estimation is needed.

4.1.1

Parameter Estimates and Evaluation of Fit

In
Training upon all customer data, we estimate the acquisition and retention parameters .
Figure 2, we plot model estimates for gross customer additions, losses, and end-of-period total
customer counts against what we actually observed (grey indicates the duration of the economic
downturn).
We must back-cast gross customer additions and losses because the only missingness Dish
suffers from is non-disclosure of additions and losses data until Q1 1998. Our resulting fits are
visually appealing we see a clear seasonal pattern within acquisitions and losses, and lower
acquisitions and higher losses during the recession of 2008. Dish appears to be past the point of
peak adoption, a sentiment echoed by Dish CEO Charlie Ergen in Dishs Q1 2015 conference
call: My general sense is that the linear pay television business probably peaked a couple of
years ago and that its in a very slight decline.
ARPU is modeled and projected as per Section 3.6. Linear growth is assumed, consistent
with comments made in Dishs annual financial reports and with the materially improved goodness of fit over a 10+ year period when we assume linear growth. Because the residuals of the
resulting model failed the Augmented Dickey Fuller unit root test, we model ARPU using an
autoregressive integrated moving average (ARIMA) (0,1,0) model with an intercept term (see
Section 3.6)7 . We will see in Figure 4 that this fit is very stable over time. Parameter estimates
are provided in Table 2 alongside their associated standard errors and the fitted models sum of
squared error (Equation 27) 8 .
7
8

Differencing the data simplifies the model, has an R2 of 99.2%, and rejects the unit root null hypothesis.
MLE and SSE estimates for 0 in the ARPU model are the same.

Figure 2: Dish Network: Gross Customer Additions, Losses and End-of-Period Customer Counts (Recession in Grey)

25

600
400

Act
Exp

200

Customers (000)

800

1000

Adds: Actual versus Expected

Q1 1996

Q1 1998

Q1 2000

Q1 2002

Q1 2004

Q1 2006

Q1 2008

Q1 2010

Q1 2012

Q1 2014

Calendar Quarter

600
400

Act
Exp

200

Customers (000)

800

1000

Losses: Actual versus Expected

Q1 1996

Q1 1998

Q1 2000

Q1 2002

Q1 2004

Q1 2006

Q1 2008

Q1 2010

Q1 2012

Q1 2014

Calendar Quarter

10000
5000

Act
Exp

Customers (000)

15000

Ending Customers: Actual versus Expected

Q1 1996

Q1 1998

Q1 2000

Q1 2002

Q1 2004

Q1 2006

Q1 2008

Q1 2010

Q1 2012

Q1 2014

Calendar Quarter

These parameters are also consistent with what Dish has disclosed in its public filings. Consider what Dish says about the seasonality of its operations in its 2015 annual report: Historically, the first half of the year generally produces fewer gross new subscriber activations than
the second half of the year, as is typical in the pay-TV industry. In addition, the first and fourth
quarters generally produce a lower churn rate than the second and third quarters. While this
statement is true, we provide a much richer depiction of seasonal fluctuation in Dishs acquisition and retention (i.e., the fact that Q3 churn is seasonally quite severe).

26

Table 2: Dish Network: Parameter Estimates and Standard Errors


Parameters

Acquisition
Retention
ARPU

Q1

Q2

Q3

Rec

N A

21.5

64407.1

2.03

.057

.063

.031

.096

.530

(8.4)

(23940.1)

(.007)

(.001)

(.001)

(.001)

(.001)

(.003)

1.62

84.2

1.44

.071

.044

.113

.122

(.75)

(27.4)

(.21)

(.002)

(.002)

(.002)

(.002)

0.226
(.085)

SSE
312,315

The negative effect of the 2008 recession on Dishs financials is unmistakeable its effect upon acquisition and retention propensities was greater than all of the respective seasonal
terms. The coefficient associated with acquisition is negative because customers have a lower
propensity to acquire services during recession, while the coefficient associated with retention is
positive because customers have a higher propensity to churn during recession, echoing Equation 14. Had we ignored the effects of recession upon acquisition and retention and instead
baked it into our baseline (as would be done by a model that assumes a time-invariant retention
rate), we would have materially underestimated future customer demand.

4.1.2

Predictive Validation and Comparison

While Section 4.1.1 shows our in-sample goodness of fit is strong for Dish, it does not tell us
(1) how much data is needed to make a proper forecast of firm value, (2) how well-validated
our models predictions are, and (3) how the predictive validity of our model compares to that
of GLS and SSW. To shed light on these questions, we perform rolling predictions of future
customer metrics. The quality of stock price estimates from our model is a direct function of
the quality of the customer metric projections coming from our proposed acquisition/retention
model. Forecasting customer metrics facilitates predictive comparisons across models because
the customer metric forecasts are automatic, requiring no human intervention. While it is tempting to consider rolling stock price predictions, doing so would require retrospective forecasts of
future gross margins, operating costs, and balance sheet adjustments, which would be subjec-

27
tive. It is hard to say what we would have predicted, if we only had data up to and including all
historical points in time.
We calibrate our model upon all data up to and including quarter Q, and then predict what
ADDQ+q , LOSSQ+q , and ENDQ+q will be over forecasting horizons q = 1, 2, . . . , 8 quarters
ahead. We let Q = 10, 11, . . . , 77, corresponding to all possible calibration periods after and
including Q2 1998. Because of missing data, only 3 quarters of ADD and LOSS data are
available when Q = 10, making it a reasonable starting point to the rolling analysis. As a result,
we evaluate the performance of GLS, SSW, and our proposed model with 508 predictions made
over 68 different calibration periods.
In Figure 3, we plot all resulting predictions over all calibration periods for ADD (first
column), LOSS (second column), and END (third column) using GLS (first row), SSW (second
row), and our proposed model (third row). The grey dots in Figure 3 correspond to estimated
model predictions at the end of the calibration period.
While the general patterns of over-estimation and under-estimation are similar between
SSW and GLS, we see that SSW has materially superior predictive validity to GLS in general.
GLS underestimates future ADD, LOSS, and END figures, often severely so. This is primarily
because of the inability of a Technological Substitution model for ADD and a constant retention rate for LOSS to model customer metrics over time. Because SSW models END with
the Technological Substitution model, SSWs resulting predictions for END are generally quite
well-behaved and well-calibrated. Both methods have the most difficulty forecasting ADD, as
evidenced by the large deviations between the predictions in grey and the actual data in black.
This is important because ADD is the primary input which is used for these models respective
valuation models.
The proposed model forecasts ADD, LOSS, and END very accurately, as evidenced by the
tight correspondence between the grey and black lines in the bottom row of Figure 3. Correspondence remains tight even when the calibration period being considered is short, which
is further proof of the robustness of the models predictions. We provide the RMSE of ADD,

Figure 3: Dish Network: Rolling 2-Year Predictive Validation and Comparison

20040328

20040328

1000

Proposed

20120328

Proposed

Proposed

600

200

19960328

20120328

Act
Exp

20120328

20040328
Date

20040328

19960328

Act
Exp

400

Customers (000)

20120328

Act
Exp

Date

800

Act
Exp

15000

19960328

20120328

200

1000

10000

15000

10000

Customers (000)

800

20040328

15000

10000

Customers (000)

5000

5000

1000

600

400

600

19960328

Act
Exp

800

Date

Customers (000)

20120328

Date

400

Customers (000)

0
20040328

SSW

Customers (000)

1000

19960328

Date

19960328

SSW

800

Date

19960328

Act
Exp

SSW

600

5000

800
20120328

Act
Exp

400

Customers (000)

20040328

200

Ending Customers: Actual versus Expected


GLS

Act
Exp

600
0

19960328

200

400
200

Customers (000)

800
600
400

Customers (000)

1000

Act
Exp

200

Losses: Actual versus Expected


GLS
1000

Adds: Actual versus Expected


GLS

28

20040328

20120328

19960328

20040328

20120328

LOSS, and END forecasts across all time horizons for each of the three models in Table 3. The
RMSE figures corresponding to SSW are generally 30-50% smaller than GLS, while the RMSE
figures of our proposed method are generally 40-70% smaller than SSW.
These conclusions are not affected by the inclusion of covariates in our model. We created
variants of GLS and SSW which allow retention, acquisition, and/or ending customers to be
affected by quarterly seasonality and the Great Recession (through a logit formulation for retention, and through proportional hazards for acquisition and ending customers), and the results
did not change.

Table 3: DISH: RMSE of Predictions by Forecasting Horizon for ADD, LOSS, and END
Metric
ADD

LOSS

END

4.1.3

Horizon
+1Q
+2Q
+3Q
+4Q
+5Q
+6Q
+7Q
+8Q
+1Q
+2Q
+3Q
+4Q
+5Q
+6Q
+7Q
+8Q
+1Q
+2Q
+3Q
+4Q
+5Q
+6Q
+7Q
+8Q

GLS
590.4
627.2
661.1
695.5
735.1
774.2
814.6
859.2
196.7
202.3
206.0
214.0
225.6
239.0
249.4
260.7
2370.3
2588.2
2831.7
3115.5
3413.9
3751.4
4116.5
4519.1

29

SSW Proposed
337.5
175.7
370.1
192.5
374.5
203.8
357.0
211.2
363.8
239.4
383.7
251.7
414.0
257.0
428.6
256.7
193.1
109.5
242.1
106.8
221.2
110.8
185.6
108.1
218.4
110.3
235.5
101.7
205.6
100.1
183.1
111.8
1203.4
229.1
1318.4
409.5
1446.8
580.0
1585.7
737.8
1734.4
920.3
1888.1 1094.2
2055.8 1251.8
2252.1 1381.4

Valuation Results and Comparison

We first project revenues (Section 3.7) far enough into the future (i.e., for 50 years) that all
subsequent profitability/loss has a negligible effect upon valuation. Long-term projections such
as these remove the need to estimate a terminal value, common in DCF models but often calculated in a relatively arbitrary manner (Courteau et al. (2001)). In Figure 4 we provide actual
and expected end-of-period total customer counts (top panel), monthly ARPU (middle panel),
and monthly sales figures (bottom panel).
Our revenue projections drive detailed financial projections used to estimate future free
cash flows, weighted average cost of capital, non-operating asset value, and net debt. For all
aspects of this procedure, see Appendix C which provides a step-by-step guide to the process

30
we followed. We then add the value of the operating assets to the non-operating assets and
subtract the net debt to arrive at our best estimate of shareholder value using Equation 1.
Figure 4: Dish Network: Summary of Projections

10000
4000

Act
Exp

Ending Customers (000)

Ending Customers: Actual versus Expected, Including Projections

2000

2010

2020

2030

2040

Date

150
100

Act
Exp

50

ARPU ($)

ARPU per Month: Actual vs Expected, Including Projections

2000

2010

2020

2030

2040

1000

1500

Total Subscriber Revenues: Actual versus Expected, Including Projections

500

Act
Exp

Total Subscriber Revenues ($MM)

Date

2000

2010

2020

2030

2040

Date

A summary of key valuation inputs and outputs from our analysis for Dish are provided in
Table 4, alongside the corresponding inputs and outputs for Sirius XM, which we will discuss
shortly. We estimate a stock price of 62.45 based on Q1 2015 results, which had been disclosed
on May 11th, 2015. The end-of-day stock price that day was 66.38, implying that we were
within 6% of the then-current stock price. Our price estimate is well-calibrated the stock had
been trading near our price estimate for over 10 months prior to Q1 2015 results, and fell back
down to the price we had estimated after results had been disclosed.

Table 4: Valuation Summary, as of Q1 2015 (Paying Customers)

4.2

Dish

Sirius XM

ARPU Growth Per Year


Subscriber Variable Cost Margin ((1 VCq ) in 4)
WACC
Total Paying Customers
Monthly ARPU ($)
Operating Margin
Average New Customer Lifetime (years)
SAC / Customer ($)

$2.71
23.3%
6.9%
13.8MM
$88.6
13.0%
5.60
$716.5

$0.49
49.3%
7.4%
22.9MM
$16.8
30.1%
4.20
$82.8

Value of Operating Assets


+ Non-Operating Assets - Net Debt
= Shareholder Value
/ Shares Outstanding
= E(Stock Price)
Actual Stock Price9
% Over(under)-estimation

$15.0B
$13.9B
$28.9B
462.1MM
$62.45
$66.38
(5.9%)

$25.6B
-$3.7B
$21.9B
5513.7MM
$3.97
$3.90
0.5%

31

Sirius XM

To test the robustness of the proposed valuation method, we repeat our valuation exercise for a
second company, Sirius XM, a large music services provider. Virtually all assumptions made
for Sirius XM are consistent with the approach outlined for Dish in Web Appendix C. As with
Dish, Sirius XM publicly discloses the magnitude of its subscriber acquisition expense, so no
estimation of this quantity is required.
Sirius XM is a complementary example for a number of reasons:
1. Sirius XM is a very high-growth business, while Dish is a mature business. ADD, LOSS,
and END are all past peak for Dish (Figure 2), while they are increasing for Sirius XM.
2. Sirius XM suffers from more severe missingness than Dish. While Sirius XM began commercial operations in Q4 200110 , no ADD, LOSS, and END data for paying customers
was publicly disclosed for almost 6 years, until Q3 2008. Data became available after its
predecessor, Sirius Satellite Radio, completed its acquisition of XM Satellite Radio. As
10

Both predecessor companies, Sirius Satellite and XM Satellite, began commercial operations at nearly the
same time February 2002 and November 2001, respectively.

32
a result, all back-casted figures represent the paying users of the combined entity. The
missigness is even more severe when we consider that Sirius XM is a younger company
than Dish. As a result, almost half of Sirius XMs customer data is missing.
3. Sirius XM is a high fixed-cost business because its satellite radios are pre-installed in
most new vehicles, while Dish Networks is a high variable-cost business. A regression of
operating expenses (net of SAC) upon revenues implies that nearly 40% of Sirius XMs
Q1 2015 operating costs are fixed in nature, while the corresponding regression for Dish
implies nearly virtually all of Q1 2015 costs are fixed. This substantially increases the
marginal profitability of Sirius XMs new users, all else being equal.
4. Sirius XM has a very different customer base and customer profile than Dish. Sirius
XM has a larger number of customers, each of whom generates less revenue but is much
cheaper to acquire (Table 4). We will discuss how the unit economics of customers compare in Section 4.3.
5. Sirius XM sells almost entirely into cars, whereas Dish sells almost entirely into homes.
All else being equal, this makes Sirius XM a more cyclical business than Dish.
Despite these differences, we proceed with virtually the same model. The only modifications we make to the proposed model for Sirius XM are to account for the fact that Sirius XM
sells primarily into the car and not into the home

11

. The parameter estimates from our fitted

model as well as their associated standard errors and the models sum of squared error (Equation 27) are provided in Table 5. Unlike with Dish, non-stationarity in the mean of the ARPU
process was neglible with Sirius XM, so the ARPU coefficient estimates (0 , 1 ) provided in
Table 5 come from a simple linear regression model (Equation 33) and not an ARIMA(0,1,0)
model (Equation 34).
11

The market size for Sirius XM is equal to the number of vehicles on the road, as provided by the Bureau
of Transportation Statistics. Corresponding, we use Total Vehicle Sales, as defined/provided by St. Louis Fed,
as our macroeconomic covariate. We denote the coefficient associated with the Total Vehicle Sales covariate by
T V S . Forecasts for these statistics are provided by two national automotive organizations, IHS and the National
Automobile Dealers Association. For more detail, please see Web Appendix D

33

Table 5: Sirius XM: Parameter Estimates


Parameters

Acquisition
Retention
ARPU

9.64

.041

.223

.002

78.9

249389.4

1.55

(17.0)

(82080.1)

2.05
(.769)

Q1

Q2

Q3

TVS

N A

.116 .057

.051

.015

.032

(.047)

(.003)

(.005)

(.004)

(.001)

(.156)

397.3

2.00

.031

.037

.002

.003

(207.3)

(.287)

(.004)

(.005)

(.005)

(.001)

SSE
180,604

In Figure 5, we plot model estimates for ADD, LOSS, and END against what we actually
observed (grey indicates the duration of the Great Recession). We overlay 2-year rolling predictions as we had done for Dish in Section 4.1.2. As was the case with Dish, the in-sample and
out-of-sample fits for Sirius XM in terms of all three customer metrics are satisfactory.
As with Dish, we project revenues 50 years into the future. In Figure 6 we provide actual
and expected end-of-period total paying customer counts (top panel), monthly ARPU (middle
panel), and monthly subscriber sales (bottom panel). ARPU is modeled assuming linear growth,
because as with Dish, goodness of fit is materially better under this assumption. ARPU is
modeled as a simple linear function of time because the residuals of this fit did not fail the
Augmented Dickey Fuller unit root test. As with Dish, the goodness of fit of this simple ARPU
model is high, with an R2 of 87%.
From the top panel of Figure 6, we see that the proposed model predicts continued growth
in the total number of paying customers for approximately another decade before declining. Estimated monthly ARPU growth for Sirius XM is nearly the same as what we had estimated for
Dish after accounting for the absolute level of monthly ARPU for the two companies (approximately 3% of Q1 2015 ARPU; see Table 4). As a result, we predict that subscriber revenues
will grow for the next two decades before plateauing.
We perform a detailed margin and cash flow analysis to turn the revenue projections in
the bottom panel of Figure 6 into monthly free cash flow projections. From Table 4, we infer
that paying customers are far cheaper to acquire for Sirius XM, generate smaller (but more

Figure 5: Sirius XM: Gross Customer Additions, Losses and End-of-Period Customer
Counts (Recession in Grey)

34

1500

Act
Exp
Rolling Predictions

500

Customers (000)

Adds: Actual versus Expected

Q4 2001

Q4 2003

Q4 2005

Q4 2007

Q4 2009

Q4 2011

Q4 2013

Date

500 1000

Act
Exp
Rolling Predictions

Customers (000)

Losses: Actual versus Expected

20011228

20031228

20051228

20071228

20091228

20111228

20131228

Date

25000

Act
Exp
Rolling Predictions

10000

Customers (000)

Ending Customers: Actual versus Expected

20011228

20031228

20051228

20071228

20091228

20111228

20131228

Date

profitable) ongoing revenue streams thereafter, and churn more quickly, with average customer
lifetimes that are approximately 25% shorter. We predict that Sirius XM will experience very
favorable margin expansion over the next decade, as its fixed costs become a smaller proportion
of its rapidly growing revenue base. All assumptions made are documented in Web Appendix
D.
We estimate Sirius XMs operating assets to be worth $25.6B. After adding non-operating
assets (Sirius XM has approximately $1.1B in net operating loss carry-forwards) and subtracting
net debt, we arrive at our best estimate of shareholder value of $21.9B using Equation 1. This
implies a fair value per share of 3.97, which is 0.5% higher than Sirius XMs stock price the

35

Figure 6: Sirius XM: Summary of Projections

25000
10000

Act
Exp

Ending Customers (000)

Ending Customers: Actual Versus Expected, Including Projections

2000

2010

2020

2030

2040

2050

2060

Date

30
20

Act
Exp

10

ARPU ($)

40

ARPU per Month: Actual Versus Expected, Including Projections

2000

2010

2020

2030

2040

2050

2060

Date

1500
500

Act
Exp

Total Subscriber Revenues ($k)

Total Subscriber Revenues: Actual Versus Expected, Including Projections

2000

2010

2020

2030

2040

2050

2060

Date

day that Q1 2015 earnings we released on April 28 2015.


In summary, our proposed model slightly over-estimates Sirius XMs stock price and slightly
under-estimates Dishs stock price, further evidence that our models price estimates are wellcalibrated.

4.3

Additional Insights

Confident that our model provides sensible valuation estimates, we return to Dish to study other
insights that we are able to draw from the model beyond stock price estimates. We look at the
remaining lifetime, CLV, and residual lifetime value (RLV) of Dish customers as a function
of their tenure with the firm. We first compare and contrast expected remaining lifetime and

36
RLV for customers with differing tenures with the firm, before looking at the entire probability
distribution. Next, we decompose Dishs current customer equity (CCE) by tenure. We finish
by introducing a customer profitability metric derived purely from firm financial and customer
data which investors may use to identify companies that earn high rates of return on their SAC.

4.3.1

Comparison of Customer Value by Tenure

Let us consider a Dish customer acquired in Q1 2015 who we call Recent Robin, and another
Dish customer acquired 10 years earlier in Q1 2005 who we call Longtime Larry. One quantity
of managerial interest is the expected future lifetime of Recent Robin and how it compares to
the expected future lifetime of Longtime Larry. Under GLS, all customers are assumed equal
and thus Recent Robin and Longtime Larry are expected to share the same expected future
lifetime. We intuitively know, however, that Longtime Larry is likely to remain a customer for
a longer period of time because his long history with Dish thus far implies he has a lower churn
propensity. Under our model, the expected residual lifetime of a customer acquired in month
m, active in and relative to month M , is by definition equal to
Z

E(m,M )
i=0

S(M + i m|m, cR , X(M + i), R , rR , R )


di.
S(M m|m, cR , X(M ), R , rR , R )

(35)

Because there is no closed form expression for Equation 35, we estimate it via Monte Carlo
simulation. For details regarding this procedure, please see Appendix A.
The expected total lifetime of Recent Robin and Longtime Larry are 5.6 years and 9.5
years, respectively, in line with our intuition. Investors want to know the expected lifetime
of customers. Longer expected customer lifetimes imply more stable future cash flows, all else
being equal, because future cash flows are less reliant on the acquisition of new customers.
At Dish, we see not only that older customers have longer residual lifetimes, but also that all
customers live for a relatively long time, which should be heartening to investors. Reducing
investors perceived risk of future cash flows improves cost of capital, raising firm valuation.

37
Another quantity of managerial interest is (pre-tax) residual lifetime value (RLV), i.e., the
net present value of all variable profit associated with a customer. Calculating this using nothing
but information provided in firm financial statements requires careful consideration of what expenses are fixed versus variable, and proper handling of subscriber acquisition cost (for details,
see Appendix A). We estimate that the pre-tax RLV associated with Recent Robin is $1,407,
excluding average initial acquisition costs of $854 (implying a pre-tax CLV for newly acquired
users of $554), while Longtime Larry is worth $1,931.
This information is useful to many stakeholders:
1. Investors may track CLV relative to SAC over time, viewing these metrics as financial
barometers of customer health. Unfavorable trends in these figures (as has been evident
at Dish, for example) are indicative of decreasing customer (and thus firm) profitability.
2. Competitors, comparable companies, and investors will be interested in the absolute level
of CLV for benchmarking purposes. If a competitor estimates its own CLV to be less than
Dishs, there may be opportunities to close the gap, identifying what it could be that is
causing the gap in average customer profitability. Investors may ask the same question
and demand that changes be made to improve CLV. In this regard, it is interesting to note
that the pre-tax CLV associated with a newly acquired customer at Sirius XM is $354,
excluding average initial acquisition costs of $83 while the expected post-acquisition
profit is smaller at Sirius XM than at Dish, the cost of acquisition is smaller still. We will
revisit this concept when we study the internal rate of return on newly acquired customers
below.
While all of the preceding analysis pertains to point estimates of future lifetime and value,
our probability model provides us with the entire distribution for both across all possible Recent
Robins and Longtime Larries. In Figure 7 we plot the distribution of pre-tax CLV for Recent
Robin and Longtime Larry. The fact that we estimate expected CLV via Monte Carlo simulation
naturally gives us the distribution of CLV as well.

38
Figure 7 provides us with additional color regarding the riskiness of future cash flows associated with new and existing customers at Dish, for example, we estimate that there is a 41%
chance that the company will earn incur a loss on a Recent Robin (i.e., the present value of the
pre-tax contribution margin for Recent Robin falls below her initial acquisition cost) 12 . Unsurprisingly, we infer a long right tail to Longtime Larrys pre-tax RLV this drives up Longtime
Larrys expected pre-tax RLV, but also implies a much higher variance about that expectation.
Longtime Larry is more valuable but is also more risky (McCarthy et al. (2016)).

3e04

Longtime Larry
Recent Robin

0e+00

1e04

2e04

Density

4e04

5e04

Figure 7: Probability Distribution for PV of Pre-Tax RLV


Recent Robin versus Longtime Larry

1000

2000

3000

4000

5000

PV of Pretax Contribution Margin ($)

4.3.2

Customer-Base Decomposition

The raw data available from virtually any public source reveals nothing about the tenure of
existing customers or how these lifetimes vary across the customer base. As hinted in the ex12

We tested for the existence of an on contract segment which is unable to churn for a two-year period, before
being able to churn. The model inferred only 1 segment of customers.

39
amples of Recent Robin and Longtime Larry, this can be vital information to outside investors.
Fortunately, our proposed model allows managers to easily infer these lifetimes and to perform
customer base segmentation on this basis. The proportion of currently active customers (i.e.,
active in quarter Q) who were born in month m is equal to
b
C(m,
3Q)
,
pm,3Q P3Q
b 3Q)
C(i,
i=1

(36)

using Equations 11 and 17.


While knowing the count of customers within each segment is helpful, the value of those
customers is what matters most to investors and managers. The proportion of total current
customer equity as of quarter Q coming from customers who were born in month m is the
CLV-weighted analog of Equation 36:
b
C(m,
3Q)E(Pre-tax CLVm,3Q )
,
pm,3Q P3Q
b 3Q)E(Pre-tax CLVi,3Q )
C(i,

(37)

i=1

where CLV is defined as in Section 4.3.1. The resulting barplot of customer base and firm CLV
proportions segmented by tenure is provided in Figure 8. We estimate, for example, that approximately one-eighth of Dishs customer base is comprised of highly loyal/inertial customers who
have been Dish subscribers for 10+ years. We also infer that longer-lived segments comprise
more of the total value of the customer base because they are inferred to have higher residual
lifetime values, as is evident from our comparison of Recent Robin and Longtime Larry above.

4.3.3

Internal Rate of Return on a Customer

Investors are interested in simple profitability metrics that serve as proxies for corporate health,
and often filter/screen on stocks using these metrics to identify companies they may consider for
investment. One of the most popular metrics amongst financial analysts is return on invested
capital (ROIC, Damodaran (2007), Greenblatt (2006)), a measure of a companys ability to
create operating profit per dollar of investment into the business. Businesses that are able to earn

Figure 8: Inferred Customer Base and Firm CLV Decomposition by Tenure, Q1 2015

40

0.15
0.00

0.05

0.10

% of Customer Base

0.20

0.25

0.30

Customer Count
CLV

02

25

510

10+

Tenure (Years)

higher rates of return on invested dollars will generate faster-growing and thus more valuable
future cash flow streams.
While ROIC is a useful measure of a companys overall ability to generate high rates of
return, it is also valuable to know the marginal profitability of investments in newly acquired
customers. Investors may consider investment in companies that are able to earn high rates
of return on their subscriber acquisition dollars because these companies will also be able to
generate future cash flow at a fast rate by continuing to invest in customers, even if previous
investments in infrastructure/distribution encumber traditional metrics like ROIC.
For this reason, we propose a financial metric called the internal rate of return on a customer (IRRC). IRRC represents the internal rate of return on a marginal customer the rate of
return which makes the customer investment a net present value zero project (see Section 4.3).
A mathematical expression is provided in Appendix A.
IRRC is an important quantity for investors to take heed of if IRRC falls below WACC,
the firm would be earning an inadequate return on its acquisition spending and investors should

41
demand that the firm cease such spending. Conversely, a very high IRRC could indicate that
the firm may want to increase its acquisition budget, because the expected return on customer
investments is well in excess of what the capital markets are demanding of the firm as a whole.
Dishs IRRC is currently 25%, a respectably high rate of return on acquisition dollars. This
rate is high enough that Dish should certainly not curtail subscriber acquisition spending yet,
even though IRRC has been in decline. In contrast, Sirius XMs IRRC is a staggering 108%, far
higher than Dishs, indicative of far superior unit economics and a harbinger of growing future
profitability. This high rate of return on new customer acquisition spending, coupled with the
healthy number of new customers we expect Sirius XM to acquire as per Figure 6, drive Sirius
XMs valuation premium vis a vis Dish.
This profitability metric and other granular inferences that can be drawn from our model
can provide useful context to investors. In some sense, the overall corporate valuation shown
earlier isnt necessarily very insightful by itself; it merely captures the voice of the financial
market and suggests that the company is a bit overvalued in this case. But the real value of our
proposed approach is the ability to go beyond the macro valuation to offer useful, operational
diagnostics to better understand where that value is coming from, and what it might mean to the
firm, its competitors, investors, and possibly even public policy makers. In the case of Dish, the
considerable amount of value coming from long-lived customers is indicative of a very mature
business, and implies that the valuation of the business as a whole will be much more dependent
on and sensitive to changes in the companys ability to retain existing companies, rather than
acquire new ones. In contrast, Sirius XMs valuation is far more dependent on the the firms
ability to acquire new users and to earn a high rate of return on the firms investment in those
new users. This is important information for investors and managers alike.

General Discussion and Future Work

42

As noted at the outset of the paper, our objective has been to create the right model for customer
acquisition and retention, and embed it within the right framework for corporate valuation. But
beyond the methods developed here, we hope this paper will serve as a call to action for firms
to perform these kinds of analyses on a more regular and rigorous basis. We have provided
a number of use cases for the analytics that can be derived from the model, including but not
limited to benchmarking the internal rate of return on customers, tracking CLV and IRRC over
time as key performance indicators for the business, comparing CLV and IRRC across comparable/competing firms, performing customer/firm value segmentation, and providing investors
with improved forward-looking sales visibility. All of this is possible because we have a flexible, general-purpose model.
While our model is particularly suited to publicly traded companies using their public disclosures, we contend that this same exercise can and should be done internally as well. Firms
can easily implement an equivalent version of this model using internal company data, enhancing its overall versatility. While estimation may differ a bit, the model for customer acquisition/retention and the valuation framework proposed here would remain essentially the same.
Measuring, tracking, and acting upon CLV can improve the ROI of company acquisition and
retention spending, and our valuation framework gives company executives the ability to estimate how much value this ROI improvement has created for the overall value of the firm. This
provides executives with an important key performance indicator to hold themselves and their
marketing managers accountable to.
While our model is considerably more flexible than previously published customer-based
corporate valuation models in terms of the dynamics that it can accommodate, it must nevertheless remain parametrically parsimonious because available data is so limited and will likely
stay that way for the foreseeable future. For example, it is highly unlikely that firms will begin
to disclose the kinds of data required to properly account for other sources of customer value,
such as the referral value of a customer (Kumar et al. (2007), Kumar et al. (2010) Kemper

43
(2009)), the impact of social media (Yu et al. (2013), Luo et al. (2013)), or customer satisfaction (Anderson et al. (2004), Homburg et al. (2005), Luo and Bhattacharya (2006)). At the same
time, indirect proxies for these factors may be obtainable in some cases through external data
sources for a small subset of companies.
Furthermore, it may seem tempting to add in other bells and whistles to further enrich the
model specifiction used here. We are open to such possibilities but are cautious about our ability
to do so. For instance, it may be the case that individual-level acquisition and retention propensities are correlated (i.e., customers who take longer to acquire may have a lower propensity to
churn once they have been acquired or vice versa see Schweidel et al. (2008a)), but our ability
to empirically identify such a correlation is very limited, increasing the risk that we over-burden
the limited data we have available. Many other theoretically reasonable extensions (e.g., allowing for cross-cohort effects (Gopalakrishnan et al. (2015)), or specifying a more complicated
market potential model) will likely suffer from similar issues. Bodapati and Gupta (2004) note
that when data is highly aggregated, even the ability to identify heterogeneity (in their setting,
latent class structure) can be challenging. For this reason, even though fits and forecasts dramatically improve when we allow for heterogeneity, it may be that some of the heterogeneity
we have inferred is due to other causes, including but not limited to state dependence. Their
paper reinforces the need for model parsimony.
Another limitation we readily acknowledge is that our framework is appropriate only for
subscription-based businesses, for whom attrition is observed (Fader and Hardie (2009)). It
may be possible, in theory, for an acquisition/retention model to be specified and fit for a
non-contractual business (i.e., with latent attrition), and/or for an upper triangular matrix to
be modeled and projected upon purchases instead of active customers as in Figure 1, in practice
non-contractual firms disclose less customer data than do contractual businesses, while also being more complex to model due to attrition being unobservable. It is not even clear what kinds
of metrics a non-contractual company would have to disclose in order for an outside analyst to
be able to perform a valuation task in a valid manner.

44
There are several other areas of future research that build upon the valuation framework
proposed in this paper. We have strictly focused our attention upon one company, but our
predictive accuracy may be improved if we were to develop a hierarchical Bayesian model,
estimating the parameters for many companies at the same time. This may alleviate some of
our data inadequacy issues by borrowing strength across firms, but will require considerable
methodological advancements to properly share information across firms and handle aggregated
missing data. Our analysis also motivates acquisition/retention spending optimization. This
would only be possible however for firms that disclose acquisition and retention spending, under
the imposition of (possibly strong) assumptions about the return on spending in both categories.
Beyond the methodological improvements, our valuation framework provides perspective
to the ongoing discussion amongst marketing scholars regarding the accounting of customer
equity and advertising spending. Consistent with Srinivasan (2015), the vast majority of Dishs
SAC is expensed and not capitalized (82% in Q1 2015) the primary component of SAC which
is capitalized is spending to purchase satellite receivers, which are then owned by Dish and
depreciated over a useful life of approximately 4 years. In contrast, acquired customers have
a much longer useful life of 5.6 years, on average (Section 4.3), and yet are not considered
assets (Wiesel et al. (2008)). As a result, SAC creates costs that are inflicted immediately while
benefits are received in the future, making the income statement not reflective of the underlying
economic condition of the business. It is no surprise, then, that Dish was generally unprofitable
earlier in its history, and profitable in recent years.
As companies increasingly recognize the importance/merit of customer-centric business
strategies (Fader (2012)), and in turn disclose customer data on a more regular and thorough
basis, there will be a growing opportunity for marketing scholars to study the behavior of large,
publicly traded companies through their customer data in conjunction with their financial statements. We hope that this paper lays a sound foundation for how future analyses will incorporate,
and shed further light upon, company valuation.

45

Customer Valuation Derivations

In this Section, we elaborate upon the procedure used to compute the results provided in Section
4.3.
The expected remaining or residual lifetime of a customer acquired in month m, active in
and relative to month M , is equal to the average of many samples drawn from its true distribution:
K
1 X (k)
E(m,M ) =

,
K k=1 m,M

(38)

(k)

where K is a large number (in practice, we set K = 100, 000). m,M , the kth sampled residual
lifetime of a customer born in month m, given he/she is active in month M , relative to month
M , is obtained using a simple conditional probability based upon Equation 16:
(k)

P (m,M = ) =

SR ( 1|m) SR ( |m)
,
1 SR (M m|m)

{1, 2, }.

(39)

In general, the pre-tax CLV of a customer born in month m, active in and relative to month
M , is equal to the average of pre-tax CLV of this customer over many sampled lifetimes:

E(Pre-tax CLVm,M ) =
(k)

Pre-tax CLVm,M =

K
1 X
(k)
Pre-tax CLVm,M ,
K k=1

(40)

1(m m,M )
EBITDASAC(m + m )
,
b m + m 1) + C(.,
b m + m ))/2 (1 + W ACC) m12
(
C(.,

m =1

(k)

b m) is the total number of active customers in month m, using Equation 11;


where C(.,
EBITDASAC(m+m ) is equal to earnings before interest, taxes, depreciation, amortization and
subscriber acquisition costs in month m + m (monthly EBITDA and SAC are estimated using
(k)

the procedure outlined in Web Appendix B and operationalized in Web Appendix C); m,M is
the kth sampled residual lifetime of a customer born in month m, given he/she is still active in
month M , relative to month M (Equation 39); and W ACC is the firms weighted average cost

46
of capital. In English, we are computing the discounted sum of all future cash flows associated
with a customer, averaging over sampled realizations of his/her residual lifetime.
The IRRC of the company in quarter q is equal to

IRRCq : SACq +

(k)
K

1(m 3q,3q+m
)
1 XX
EBIDASAC(3q + m )
=0

b 3q + m 1) + C(.,
b 3q + m ))/2 (1 + IRRC)m /12
K k=1 m =1 (C(.,
(41)

where SACq represents the subscriber acquisition expense per user acquired in quarter q;
EBIDASAC(3q + m ) represents earnings before interest, depreciation, amortization, and subscriber acquisition cost in month 3q + m (monthly EBITDA, SAC, and taxes are estimated
using the procedure outlined in Web Appendix B and operationalized in Web Appendix C);
(k)

b m) is the total number of active customers in month


m,M is defined as in Equation 39; and C(.,
m, using Equation 11.

Web Appendix W1

In this Appendix, we provide additional detail regarding our DCF model so that scholars may
reproduce our analysis and perform similar analyses upon other companies. We begin by elaborating further upon the weighted average cost of capital (WACC) in Equation 2. WACC is equal
to the proportion of debt within the capital structure multiplied by the after-tax cost of debt (rD ),
plus the proportion of equity within the capital structure multiplied by the cost of equity (rE ):

WACC =

E
D
rD (1 TR) +
rE ,
D+E
D+E

(42)

where D is the market value of the companys debt, and E is the market value of the companys
equity, rD is equal to the value-weighted average current yield on the companys debt instruments, and rE is equal to the risk-free rate (rf ) plus the levered stock beta () multiplied by the

47
equity risk premium (ERP, or the required rate of return on the market (rm ) over and above rf ):

rE = rf + ERP.

(43)

The final key elements of Equation 1 are NOA, which is equal to the market value of all assets
that are not required for the ongoing operations of the business (i.e., deferred tax assets, ownership interests in unrelated businesses, etc.), net of associated liabilities, and ND, which is the
firms net debt.
Next, we further define the key financial items presented in Section 3.1:
NFWC is equal to current assets (CA) minus current liabilities (CL), excluding excess
cash and cash equivalents (XCASH) and short term debt (STD):

NFWCm = CAm CLm (XCASHm STDm ).

(44)

The variable operating margin (1 VC%) is equal to the gross margin, net of variable
operating expenses in a given period, expressed as a percentage of revenues:

1 VC%m =

GPm VOPEXm
.
REVm

(45)

VOPEXm and fixed operating expenses (FOPEXm ), are obtained by regressing total operating expenses (OPEXm ) against total revenues:

E(VOPEXm ) = 1 REVm ,

where

OPEXm = FOPEXm + 1 REVm + m .

(46)
(47)

All public, domestic companies report income statement and balance sheet data via SEC
EDGAR quarterly; for reference, relevant financial metrics for our real data example are
available upon request by the authors). ND (Equation 1); CAPEX and D&A (Equation

48
3); REV, and TR (Equation 4); CA, CL, XCASH, and STD (Equation 44); and OPEX are
all directly obtainable through EDGAR in this way.
The companys equity, E (Equation 42), is conventionally set equal to the then-current
market capitalization of the companys stock (available through data sources such as Yahoo!Finance.
The market value of the firms debt and cost of debt (D and rD , respectively, from Equation
42) can be obtained from the market price of corporate bonds traded in secondary markets,
available through FINRAs Trace and other vendors.
The risk-free rate, rf , within Equation 43 is commonly taken to be the then-current yield
on 10-year Treasury bonds, and the same rate is used in the estimate of ERP (available
from the US Treasurys website).
within Equation 43 is available through data vendors (i.e., Bloomberg), and is commonly benchmarked off of comparable firms.
ERP estimates are made available on the website of Aswath Damodaran (Damodaran
(2015)).
WACC (Equation 42) may theoretically vary over time due to fluctuations in each of these
components, WACC is often assumed fixed, particularly for firms who appear to target a
particular ratio of debt to equity within their capital structures (i.e., Dish).
The market value of non-operating assets (Equation 1) is sometimes difficult to estimate,
because reported balance sheet figures represent their book value and not their market
value. Care must be taken when estimating the value of these assets. Historical nonoperating asset value estimates can be obtained from analyst reports.
For more detail regarding valuation theory, see Damodaran (2012). For reference, all relevant historical financial information for Dish are available upon request from the authors.

49
Using the results presented above, we may easily obtain historical free cash flow estimates.
Estimating the stock price requires the following additional steps:
Projecting monthly future revenues: This represents the core of our valuation model and
its implementation is presented in Section 3.6, which concludes with future monthly subscriber revenue estimates in Equation 31.
Projecting non-subscriber revenues: Even though subscriber revenues should represent
the majority of total revenues (in the case of Dish, for example, subscriber revenues represented 99% of total sales in Q1 2015), value derived from non-subscriber cash flows
should not be ignored.
Projecting NOPAT (see Equation 4): We must make a number of margin projections:
We project SAC expenses, driving these expenses off of the historical evolution of
SAC costs relative to customer additions.
We project non-SAC G&A expenses through a decomposition into fixed and variable cost, the variable component of which is driven off of our revenue projections.
We project capital expenditures, D&A and other expense, which are often fixed in
nature for the existing operations of mature businesses.
We project the corporate income tax rate.
Projecting future changes in NFWC: Changes in NFWC are typically driven off of historical sales over the previous twelve months.
We discuss operationalization of these additional steps for Dish, as well as the specific assumptions driving our projections, in Section C.

Web Appendix W2

50

In this Web Appendix, we provide additional information needed to replicate our analysis of
Dish Network. In Table 6, we provide a table containing Dish Networks most important customer metrics, as well as total subscriber revenues each quarter. All other data corresponding
to Dish is available upon request from the authors. For most of Dishs corporate history, all
relevant figures are contained in Other Data, within the first table of the Results of Operations section of Managements Discussion and Analysis of Financial Condition and Results
of Operations.
Below we provide all assumptions made to obtain FCF projections for Dish which are not
described above.
Estimating WACC: The total market value of Dish equity as of the filing date of Q1 2015
results is $30.7B. The total estimated value of Dishs debt is $14.4B. This implies that
68% of Dishs capital structure is comprised of equity. The proportion of equity within the
capital structure has been relatively stable over time, implying a constant future WACC
assumption is reasonable. Dishs levered beta is 1.05, implying an unlevered beta of
0.82. For ERP, we use the smoothed ERP estimate from Aswath Damodarans website
(Damodaran (2015)) as of May 2015 of 6.2%. The risk free rate is taken to be the thencurrent long-term Treasury Bond rate of 2.13%. We estimate a cost of debt of 5.5%,
estimated via then-current market yields on public debt.
Projecting non-subscriber revenues: Non-subscriber revenues for Dish represent only 1%
of total sales in Q1 2015, and thus are not material to the valuation of the business. This
business primarily relates to the sale of equipment. We estimate future equipment sales by
estimating trend-line growth over the prior 7 year period from Q1 2008 to Q1 2015, after
accounting for fluctuations in equipment sales in 2011-2012 due to end of year accounting
adjustments. As a result, we expect future growth for this segment of approximately 0.5%
per year, in line with the slow growth we have seen historically.

51
Projecting subscriber gross margin: We predict future subscriber gross margin to be equal
to the Q1 2015 subscriber gross margin, 36.4%. Because Dishs gross margin has experienced some gross margin erosion in recent quarters, taking a 3-year historical average is
likely to over-state future gross profitability. Gross margin is purely variable in nature, in
line with convention.
Projecting non-subscriber gross margin: Equipment sales gross margin had been historically relatively stable prior to the Q4 2014-Q1 2015 period. Therefore we assume future
equipment sales gross margin equal to the historical average over the period Q1 2014-Q3
2014, 26.6%. As with subscriber gross margin, this implies that equipment gross margin
is assumed to be variable in nature.
Projecting SAC expense: Our acquisition model provides us with monthly future expected
gross customer acquisitions. We multiple this by the expected ratio of SAC expense to
b
historical gross customer acquisitions, C(m,
m) (see Equation 17), estimated using linear
extrapolation off of the recent trend-line growth observed over the past 8 years. This implies future growth in acquisition expense per gross customer addition of approximately
3.4% per year.
Projecting non-SAC G&A expense: While G&A is often argued to be purely fixed in
nature, non-SAC G&A at Dish has historically been remarkably stable as a percentage
of total sales. This suggests that non-SAC G&A is effectively a pure variable cost. We
project future non-SAC G&A to be equal to the trailing 1 year average, or approximately
5.6% of quarterly sales.
Projecting depreciation and amortization expense: We project D&A expense each future
quarter to be equal to the year-earlier period, because we are valuing the existing operations of the business and not incorporating any future growth options the firm may
explore in the future into our revenue projections. This implies D&A fluctuates between
$246MM and $287MM per quarter into the future.

52
Projecting capital expenditures: We project capital expenditures to be equal to D&A,
which should be a reasonable maintenance level.
Projecting corporate tax rate: We assume a standard corporate tax rate of 40%.
Projecting changes in NFWC: We assume changes in NFWC as a percentage of trailing
twelve months will be equal to the prior two year average to smooth out quarterly noise,
or -0.5%.
Infilling costs: All expense items discussed thus far are reported quarterly, while our
modeling is performed using estimated monthly revenue figures. To infill expenses, we
assume all variable costs are a fixed percentage of monthly sales within each quarter, and
all fixed costs represent one third of quarterly total fixed costs.
Estimating non-operating asset value and net debt: Dish has made a number of very
large wireless spectrum asset purchases over the past 3 years, most recently in Q1 2015.
These assets are not related to the ongoing operations of the core business. Because
our customer-based corporate valuation model does not offer us any special insights into
the value of non-operating assets such as these, we estimate the following valuations via
the midpoint of Wall Street analyst reports (see, for example, Swinburne et al. (2014),
Manning (2015), and Hodulik et al. (2015)):
Wireless asset value post-tax: $25B
Hughes Retail Group Tracking Stock: $0.5B
Net present value of Tax Assets: $2.2B

Web Appendix W3

In this Web Appendix, we provide additional information needed to replicate our analysis of
Sirius XM. In Table 7, we provide a table containing Sirius XMs most important customer

53
metrics, as well as total subscriber revenues each quarter13 . All other data corresponding to
Sirius XM is available upon request from the authors.
Data regarding the number of vehicles on the road is provided by the Bureau of Transportation Statistics. We make this time series isotonic (monotone increasing) to ensure non-negativity
over time.
Below we provide all assumptions made to obtain FCF projections for Sirius XM which are
not described above.
Revenues: We add Subscriber Revenue, Advertising Revenue and Other Revenue together
into the Revenues column in Table 7 because revenue generated through all three channels
is strongly driven off of the size of the customer base.
Estimating WACC: The total market value of Sirius XM equity as of the filing date of
Q1 2015 results is $21.6B. The total estimated value of Sirius XMs debt is $5.1B. This
implies that 81% of Dishs capital structure is comprised of equity. The proportion of
equity within the capital structure has been relatively stable in recent years, implying
a constant future WACC assumption is reasonable. Sirius XMs levered beta is 1.01,
implying an unlevered beta of 0.91. For ERP, we use the smoothed ERP estimate from
Aswath Damodarans website (Damodaran (2015)) as of April 2015 of 6.2%. The risk
free rate is taken to be the then-current long-term Treasury Bond rate of 2.00%. We
estimate a cost of debt of 6.4%.
Projecting non-subscriber revenues: As was the case with Dish, non-subscriber sales are
a very small proportion of total revenues, at approximately 2.2%. Unlike Dish, Sirius
XMs equipment sales are growing modestly at a linear rate, at approximately $0.6MM
per quarter. We project this modest growth will continue into the future.
Projecting subscriber gross margin: Unlike Dish, Sirius XMs subscriber gross margin
13

As noted above, Sirius XM disclosed its own revenue figures prior to Siriuss merger with XM Satellite Radio,
and neither Sirius nor XM disclosed paying customer metrics until after the merger took place. This makes premerger data not meaningful.

54
has been relatively stable. We predict future subscriber gross margin to be equal to its
3-year historical average, or 62.9%. Results are not sensitive to the choice of look-back
period. Gross margin is purely variable in nature, in line with convention.
Projecting non-subscriber gross margin: Equipment sales gross margin had been historically relatively stable. We assume the non-subscriber gross margin will be equal to its
prior 5-year average of 57.1%. Results are not sensitive to the look-back period (the
full-year 2014 non-subscriber gross margin is 57.6%).
Projecting SAC expense: Unlike with Dish, Sirius XMs ratio of SAC to ADD has been
declining instead of rising, but at a declining rate. Linear extrapolation of this ratio into
the future would eventually result in negative numbers, and would not be consistent with
the deceleration evident in this ratio over time. We assume that this decline in the cost
to acquire new paying customers will continue for another 5 years (estimated using a
quadratic fit), driving the acquisition cost of a new paying user to $73 from its most
recent level of $80. Thereafter, we assume a flat SAC/ADD ratio.
Projecting non-SAC operating expenses: We run a regression of real sales against real
operating expenses (inclusive of depreciation and amortization) net of subscriber acquisition costs using data from Q4 2008 through Q1 2015. The resulting regression estimates
imply total real fixed costs of approximately $99MM per quarter, and variable costs equal
to 13% of real revenues. We predict that future fixed costs will grow at the rate of inflation
while future variable costs are equal to 13% of predicted future revenues.
Projecting capital expenditures: Unlike Dish, Sirius XM is still growing and investing in
growth capital expenditures, over and above the capital expenditures required to maintain
the existing operations of the business. We assume capital expenditures are equal to their
historical 2 year average of $85.1MM per quarter (depreciation and amortization averages
to $65.4MM over the same period).

Projecting corporate tax rate: We assume a standard corporate tax rate of 40%.

55

Projecting changes in NFWC: We assume changes in NFWC as a percentage of trailing


twelve months will be equal to the prior two year average to smooth out quarterly noise,
or -19.7%.
Infilling costs: All expense items discussed thus far are reported quarterly, while our
modeling is performed using estimated monthly revenue figures. To infill expenses, we
assume all variable costs are a fixed percentage of monthly sales within each quarter, and
all fixed costs represent one third of quarterly total fixed costs.
Estimating non-operating asset value and net debt: Sirius XM has a substantial amount
of net operating loss carry-forwards (NOLs), which will reduce Sirius XMs future tax
payments, due to previously incurred losses. The net present value of these NOLs is
$1.1B.

References
Anderson, Eugene W, Claes Fornell, Sanal K Mazvancheryl. 2004. Customer satisfaction and
shareholder value. Journal of Marketing 68(4) 172185.
Andon, Paul, Jane Baxter, Graham Bradley. 2001. Calculating the economic value of customers
to an organisation. Australian Accounting Review 11(23) 6272.
Bass, Frank. 1969. A new product growth for model consumer durables. Management Science
15(5) 215227.
Bauer, Hans H, Maik Hammerschmidt. 2005. Customer-based corporate valuation: Integrating
the concepts of customer equity and shareholder value. Management Decision 43(3) 331
348.

56
Baxter, Robbie Kellman. 2015. The Membership Economy: Find Your Super Users, Master the
Forever Transaction, and Build Recurring Revenue. New York, NY: McGraw Hill Professional.
Berger, Paul D, Naras Eechambadi, Morris George, Donald R Lehmann, Ross Rizley, Rajkumar
Venkatesan. 2006. From customer lifetime value to shareholder value theory; Empirical
evidence, and issues for future research. Journal of Service Research 9(2) 156167.
Blattberg, Robert C, John Deighton. 1996. Manage marketing by the customer equity test.
Harvard Business Review 74(4) 136.
Bodapati, Anand V, Sachin Gupta. 2004. The recoverability of segmentation structure from
store-level aggregate data. Journal of Marketing Research 41(3) 351364.
Bonacchi, Massimiliano, Mascia Ferrari, Massimiliano Pellegrini. 2008. The lifetime value
scorecard: From e-metrics to internet customer value. Managerial and Financial Accounting
18 193226.
Bonacchi, Massimiliano, Kalin Kolev, Baruch Lev. 2015. Customer franchise: A hidden, yet
crucial asset. Contemporary Accounting Research 32 10241049.
Bonacchi, Massimiliano, Paolo Perego. 2012. Measuring and managing customer lifetime
value: A CLV scorecard and cohort analysis in a subscription-based enterprise. Management Accounting Quarterly 14(1) 2739.
Boyce, Gordon. 2000. Valuing customers and loyalty: The rhetoric of customer focus versus
the reality of alienation and exclusion of (de-valued) customers. Critical Perspectives on
Accounting 11(6) 649689.
Columbus, Louis. 2014. IDC predicts saas enterprise applications will be a $50.8b market by
2018. Forbes URL http://www.forbes.com/sites/louiscolumbus/2014/

57
12/20/idc-predicts-saas-enterprise-applications-will-be-a50-8b-market-by-2018/.
Courteau, Lucie, Jennifer L Kao, Gordon D Richardson. 2001. Equity valuation employing the
ideal versus ad hoc terminal value expressions. Contemporary Accounting Research 18(4)
625661.
Damodaran, Aswath. 2007. Return on capital (ROC), return on invested capital (ROIC) and
return on equity (ROE): Measurement and implications (July 2007) Available at SSRN:
http://ssrn.com/abstract=1105499.
Damodaran, Aswath. 2012. Investment valuation: Tools and Techniques for Determining the
Value of Any Asset. Hoboken, N.J.: John Wiley & Sons.
Damodaran, Aswath. 2015. Website of Aswath Damodaran. URL http://pages.stern.
nyu.edu/adamodar/.
Dickey, David A, Wayne A Fuller. 1979. Distribution of the estimators for autoregressive time
series with a unit root. Journal of the American Statistical Association 74(366a) 427431.
DISH. 2014. 10-K Report. URL http://www.sec.gov/Archives/edgar/data/
1001082/000110465915012777/a14-26838_110k.htm.
Elliott, Graham, Thomas J Rothenberg, James H Stock. 1996. Efficient tests for an autoregressive unit root. Econometrica 64(4) 813836.
Fader, Peter. 2012. Customer Centricity: Focus on the Right Customers for Strategic Advantage.
Philadelphia, PA: Wharton Digital Press.
Fader, Peter S, Bruce GS Hardie. 2009. Probability models for customer-base analysis. Journal
of Interactive Marketing 23(1) 6169.
Fader, Peter S, Bruce GS Hardie. 2010. Customer-base valuation in a contractual setting: The
perils of ignoring heterogeneity. Marketing Science 29(1) 8593.

58
Gopalakrishnan, Arun, Eric Bradlow, Peter Fader. 2015. A cross-cohort changepoint model for
customer-base analysis Available at SSRN: http://ssrn.com/abstract=2119337.
Gourio, Francois, Leena Rudanko. 2014. Customer capital. The Review of Economic Studies
81(3) 11021136.
Greenblatt, Joel. 2006. The Little Book That Beats the Market. Hoboken, N.J.: John Wiley &
Sons.
Greenwald, Bruce, Judd Kahn, Paul D Sonkin, Michael Van Biema. 2004. Value Investing:
From Graham to Buffett and Beyond. Hoboken, N.J.: John Wiley & Sons.
Gupta, Sunil. 2009. Customer-based valuation. Journal of Interactive Marketing 23(2) 169
178.
Gupta, Sunil, Dominique Hanssens, Bruce Hardie, Wiliam Kahn, V Kumar, Nathaniel Lin,
Nalini Ravishanker, S Sriram. 2006. Modeling customer lifetime value. Journal of Service
Research 9(2) 139155.
Gupta, Sunil, Donald R Lehmann. 2005. Managing Customers as Investments: The Strategic
Value of Customers in the Long Run. Upper Saddle River, NJ: Wharton School Publishing.
Gupta, Sunil, Donald R Lehmann. 2006. Customer lifetime value and firm valuation. Journal
of Relationship Marketing 5(2-3) 87110.
Gupta, Sunil, Donald R Lehmann, Jennifer Ames Stuart. 2004. Valuing customers. Journal of
Marketing Research 41(1) 718.
Gupta, Sunil, Valarie Zeithaml. 2006. Customer metrics and their impact on financial performance. Marketing Science 25(6) 718739.
Hand, John RM. 2015. Discussion of customer franchise a hidden, yet crucial, asset. Contemporary Accounting Research 32 10501052.

59
Hodulik, John, Batya Levi, Lisa Friedman. 2015. Dish Network: Still dont know how the story
ends. UBS Analyst Report .
Holthausen, Robert W, Mark E Zmijewski. 2013. Corporate Valuation: Theory, Evidence and
Practice.
Homburg, Christian, Nicole Koschate, Wayne D Hoyer. 2005. Do satisfied customers really
pay more? A study of the relationship between customer satisfaction and willingness to pay.
Journal of Marketing 69(2) 8496.
Jamal, Zainab, Randolph E Bucklin. 2006. Improving the diagnosis and prediction of customer
churn: A heterogeneous hazard modeling approach. Journal of Interactive Marketing 20(3-4)
1629.
Janzer, Anne H. 2015. Subscription Marketing. Mountain View, CA: Cuesta Park Consulting.
Kemper, Andreas. 2009. Valuation of network effects in software markets: a complex networks
approach. Springer Science & Business Media.
Koller, Tim, Marc Goedhart, David Wessels. 2010. Valuation: Measuring and Managing the
Value of Companies. Hoboken, N.J.: John Wiley & Sons.
Kumar, V, Lerzan Aksoy, Bas Donkers, Rajkumar Venkatesan, Thorsten Wiesel, Sebastian Tillmanns. 2010. Undervalued or overvalued customers: Capturing total customer engagement
value. Journal of Service Research 13(3) 297310.
Kumar, V, Denish Shah. 2009. Expanding the role of marketing: From customer equity to
market capitalization. Journal of Marketing 73(6) 119136.
Kumar, V., Denish Shah. 2015. Handbook of Research on Customer Equity in Marketing.
Edward Elgar Pub.
Kumar, Viswanathan, J Andrew Petersen, Robert P Leone. 2007. How valuable is word of
mouth? Harvard Business Review 85(10) 139.

60
Libai, Barak, Eitan Muller, Renana Peres. 2009. The diffusion of services. Journal of Marketing
Research 46(2) 163175.
Luo, Xueming, CB Bhattacharya. 2006. Corporate social responsibility, customer satisfaction,
and market value. Journal of Marketing 70(4) 118.
Luo, Xueming, Jie Zhang, Wenjing Duan. 2013. Social media and firm equity value. Information Systems Research 24(1) 146163.
Manning, Kevin. 2015. Dish Network: Takeaways from our launch. BMO Capital Markets
Analyst Report .
McCarthy, Daniel, Peter Fader, Bruce Hardie. 2016. V(CLV): Examining variance in models of
customer lifetime value. Available at SSRN: http://ssrn.com/abstract=2739475.
Moe, Wendy W, Peter S Fader. 2002. Using advance purchase orders to forecast new product
sales. Marketing Science 21(3) 347364.
Morrison, Donald G, David C Schmittlein. 1980. Jobs, strikes, and wars: Probability models
for duration. Organizational Behavior and Human Performance 25(2) 224251.
Pfeifer, Phillip E. 2011. On estimating current-customer equity using company summary data.
Journal of Interactive Marketing 25(1) 114.
Pfeifer, Phillip E, Paul W Farris. 2004. The elasticity of customer value to retention: The
duration of a customer relationship. Journal of Interactive Marketing 18(2) 2031.
Schulze, Christian, Bernd Skiera, Thorsten Wiesel. 2012. Linking customer and financial metrics to shareholder value: The leverage effect in customer-based valuation. Journal of Marketing 76(2) 1732.
Schweidel, David A, Peter S Fader, Eric T Bradlow. 2008a. A bivariate timing model of customer acquisition and retention. Marketing Science 27(5) 829843.

61
Schweidel, David A, Peter S Fader, Eric T Bradlow. 2008b. Understanding service retention
within and across cohorts using limited information. Journal of Marketing 72(1) 8294.
Skiera, Bernd, Christian Schulze. 2014.

Handbook of service marketing research, chap.

Customer-based vluation: similarities and differences to traditinoal discounted cash flow


models. Edward Elgar Publishing, Cheltenham, UK, 123134.
Srinivasan, Shuba. 2015. Capitalizing advertising spending. Presentation to the Marketing
Accountability Standards Board; Chicago, IL.
Srinivasan, V, Charlotte H Mason. 1986. Nonlinear least squares estimation of new product
diffusion models. Marketing science 5(2) 169178.
Stanford, Law School. 2004. Case Summary, Netflix, Inc. Securities Litigation. URL http:
//securities.stanford.edu/filings-case.html?id=103168.
Swinburne, Benjamin, Ryan Fiftal, Bronson Kussin, Thomas Yeh. 2014. Dish Network: The
builder. Morgan Stanley Analyst Report .
Van den Bulte, Christophe, Gary L Lilien. 1997. Bias and systematic change in the parameter
estimates of macro-level diffusion models. Marketing Science 16(4) 338353.
Warrillow, John. 2015. The Automatic Customer: Creating a Subscription Business in Any
Industry. Westminster, U.K.: Penguin.
Wiesel, Thorsten, Bernd Skiera, Julian Villanueva. 2008. Customer equity: An integral part of
financial reporting. Journal of Marketing 72(2) 114.
Yu, Yang, Wenjing Duan, Qing Cao. 2013. The impact of social and conventional media on
firm equity value: A sentiment analysis approach. Decision Support Systems 55(4) 919926.
R
Zivot, Eric, Jiahui Wang. 2007. Modeling Financial Time Series with S-Plus .
Springer

Science & Business Media.

62

Table 6: Dish Networks Publicly Disclosed Historical Customer Data


Customer Data (000) and Subscriber Revenues ($MM) [Web Appendix]
Period
Q1 1996
Q2 1996
Q3 1996
Q4 1996
Q1 1997
Q2 1997
Q3 1997
Q4 1997
Q1 1998
Q2 1998
Q3 1998
Q4 1998
Q1 1999
Q2 1999
Q3 1999
Q4 1999
Q1 2000
Q2 2000
Q3 2000
Q4 2000
Q1 2001
Q2 2001
Q3 2001
Q4 2001
Q1 2002
Q2 2002
Q3 2002
Q4 2002
Q1 2003
Q2 2003
Q3 2003
Q4 2003
Q1 2004
Q2 2004
Q3 2004
Q4 2004
Q1 2005
Q2 2005
Q3 2005

Adds Losses Ending Revenues


0
0
0
0
100
6
190
15
350
40
479
58
590
76
820
99
1040
122
204
44
1200
140
236
36
1400
161
267
67
1600
187
395
55
1940
218
400
40
2300
271
414
114
2600
318
500
100
3000
372
580
170
3410
392
585
95
3900
492
618
218
4300
577
648
148
4800
634
705
245
5260
668
688
248
5700
797
656
286
6070
886
684
324
6430
925
692
292
6830
998
619
289
7160
1019
642
342
7460
1080
722
402
7780
1124
804
404
8180
1207
687
337
8530
1293
701
431
8800
1343
746
461
9085
1366
759
419
9425
1438
785
425
9785
1494
851
511
10125
1661
897
547
10475
1733
907
477
10905
1837
801
476
11230
1894
799
574
11455
1990
900
645
11710
2008

Period
Adds
Q4 2005 897
Q1 2006 794
Q2 2006 824
Q3 2006 958
Q4 2006 940
Q1 2007 890
Q2 2007 850
Q3 2007 904
Q4 2007 790
Q1 2008 730
Q2 2008 752
Q3 2008 825
Q4 2008 659
Q1 2009 653
Q2 2009 731
Q3 2009 887
Q4 2009 847
Q1 2010 833
Q2 2010 747
Q3 2010 819
Q4 2010 653
Q1 2011 681
Q2 2011 572
Q3 2011 656
Q4 2011 667
Q1 2012 673
Q2 2012 665
Q3 2012 739
Q4 2012 662
Q1 2013 654
Q2 2013 624
Q3 2013 734
Q4 2013 654
Q1 2014 639
Q2 2014 656
Q3 2014 691
Q4 2014 615
Q1 2015 554

Losses
567
569
629
663
590
580
680
794
705
695
777
835
761
747
705
646
598
596
766
848
809
623
707
767
645
569
675
758
648
618
702
699
646
599
700
703
678
688

Ending
12040
12265
12460
12755
13105
13415
13585
13695
13780
13815
13790
13780
13678
13584
13610
13851
14100
14337
14318
14289
14133
14191
14056
13945
13967
14071
14061
14042
14056
14092
14014
14049
14057
14097
14053
14041
13978
13844

Revenues
2137
2183
2325
2374
2541
2552
2676
2699
2764
2810
2876
2886
2883
2865
2878
2863
2933
3036
3141
3186
3181
3199
3311
3229
3232
3225
3296
3267
3277
3348
3453
3464
3500
3556
3645
3648
3656
3689

63

Table 7: Sirius XMs Publicly Disclosed Historical Paying Customer Data


Customer Data (000) and Subscriber Revenues ($MM) [Web Appendix]
Period
Q4 2001
Q1 2002
Q2 2002
Q3 2002
Q4 2002
Q1 2003
Q2 2003
Q3 2003
Q4 2003
Q1 2004
Q2 2004
Q3 2004
Q4 2004
Q1 2005
Q2 2005
Q3 2005
Q4 2005
Q1 2006
Q2 2006
Q3 2006
Q4 2006
Q1 2007
Q2 2007
Q3 2007
Q4 2007
Q1 2008
Q2 2008

Adds Losses Ending Revenues


0
0
0
0
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>
<NM>

Period
Adds
Q3 2008
Q4 2008 1189
Q1 2009 909
Q2 2009 911
Q3 2009 962
Q4 2009 1182
Q1 2010 1014
Q2 2010 1164
Q3 2010 1182
Q4 2010 1292
Q1 2011 1126
Q2 2011 1331
Q3 2011 1353
Q4 2011 1385
Q1 2012 1329
Q2 2012 1514
Q3 2012 1502
Q4 2012 1629
Q1 2013 1488
Q2 2013 1447
Q3 2013 1479
Q4 2013 1539
Q1 2014 1380
Q2 2014 1538
Q3 2014 1624
Q4 2014 1778
Q1 2015 1621

Losses
830
1023
926
926
935
944
860
924
941
1005
968
989
1010
1029
1051
1131
1100
1183
1024
1106
1127
1207
1158
1244
1269
1227

Ending
15191
15550
15436
15421
15457
15704
15774
16078
16336
16687
16808
17170
17534
17909
18208
18671
19042
19570
19875
20298
20670
21082
21255
21635
22015
22523
22917

Revenues
477
632
577
580
608
657
650
681
700
715
708
727
747
761
788
821
850
870
879
922
944
974
974
1008
1034
1061
1144

S-ar putea să vă placă și