Imagine improving net recovery rates by 10-20% on your consumer charge-off portfolios. Using smart analytics and creative, out of the box thinking can deliver millions in improved bottom line.
Sheet1Financial Statement Analysis Paper Rubric - CatrinaMissingA.docx
Recovery Optimization
1. Op#mizing
Recoveries
for
Loan
Servicing
Por6olios
Through
Smart
Alloca#on
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
S.
Blair
Korschun
August
10,
2011
2. How
should
we
allocate
a
post
charge-‐off
consumer
loan
por6olio
to
op#mize
servicing
results?
First
we
need
to
ask
the
right
ques#ons.
The
typical
ques#ons
asked
include:
What
are
my
current
alloca#on
op#ons?
What
are
the
recovery
rates
for
each
servicing
op#on?
For
example
a
6
month
batch
liquida#on
rate
would
be
the
percentage
of
the
face
value
of
debt
owed
that
is
recovered
within
a
6
month
period
with
the
batch
being
the
por6olio
debt
placed
during
a
single
month
(the
batch).
Another
important
ques#on
is
–
What
is
the
cost
to
achieve
the
recovery
rate
for
each
servicing
op#on?
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
2
3. What
are
our
current
alloca#on
op#ons?
For
simplicity
let’s
assume
that
we
are
considering
only
fresh
charge-‐off
accounts
and
that
we’ve
been
using
one
internal
team
and
two
outside
recovery
agencies.
Assume
we’ve
given
40%
to
internal
team
and
30%
shares
each
to
the
two
agencies.
Are
these
the
only
op#ons
we
have
without
adding
people
or
a
new
outside
agency?
What
about
a
no
work
strategy
where
you
only
respond
to
inbound
calls?
Yes
there
would
be
low
results
but
there
would
also
be
very
low
cost.
What
about
using
an
automated
le_er
only
strategy
(no
outbound
calling)?
Again
this
would
be
rela#vely
low
cost
but
higher
than
the
no
work
strategy
op#on.
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
3
4. If
we
know
the
liquida#on
rate
then
how
should
we
allocate?
Let’s
ignore
the
no
work
and
le_er
only
op#ons
and
only
consider
the
internal
team
and
two
agencies.
Team
6Mth
Liq.
Rate
Internal
8.04%
Agency
A
9.35%
Agency
B
7.12%
How
should
we
allocate
the
accounts?
Should
we
give
more
to
Agency
A?
Should
we
fire
Agency
B?
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
4
5. What
other
ques#ons
should
we
ask
and
answer
before
changing
our
alloca#on
strategy?
We
need
to
understand
cost
as
the
net
recovery
rate
is
more
important
than
the
gross.
Let’s
assume
we
know
the
costs
as
a
percentage
of
recovery
dollars.
Team
6Mth
Liq.
Cost
Net
Liq.
Rate
Rate
Internal
8.04%
21%
6.35%
Agency
A
9.35%
30%
6.55%
Agency
B
7.12%
30%
4.99%
It
now
appears
that
Internal
is
doing
the
almost
as
well
as
Agency
A.
Should
we
give
more
to
Internal
and
Agency
A
and
fire
Agency
B?
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
5
6. We
s#ll
need
to
ask
more
ques#ons.
We
should
ask
if
the
sample
size
is
sufficient
to
give
us
confidence
that
the
differences
are
sta#s#cally
significant.
We
should
also
understand
if
the
results
fit
a
normal
distribu#on
curve.
In
our
example
of
recovery
performance
the
results
with
a
batch
are
not
normal
as
you
would
have
a
fat
tail
due
to
non-‐payers
(i.e.
lots
of
accounts
with
value
of
zero).
On
the
other
hand
if
we
have
3
years
of
monthly
results
then
the
results
per
batch
may
fit
a
normal
curve
–
but
we
don’t
want
to
wait
three
years
to
gather
data
to
verify
its
significance.
It
is
fairly
common
to
test
hypotheses
using
confidence
intervals.
Using
95%
confidence
is
typical
but
you
can
adjust
this.
Read
more
about
this
is
any
Sta#s#cs
textbook
or
see
the
CONFIDENCE
func#on
in
Excel.
Check
with
a
sta#s#cian
or
do
the
math
yourself
to
ensure
that
you
can
have
confidence
that
the
performance
differences
are
real
(significant).
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
6
7. Having
consistent
trends
in
our
data
is
cri#cal
to
being
able
to
act
with
confidence.
We
also
need
to
know
if
the
results
are
changing
over
#me.
Scenario
A
–
Inconsistent
results
over
#me
Team
Jan
Net
Liq
%
Feb
Net
Liq%
Mar
Net
Liq%
Internal
8.31%
5.65%
6.35%
Agency
A
5.65%
7.78%
6.55%
Agency
B
7.58%
3.69%
4.99%
Scenario
B
–
Consistent
results
over
#me
Team
Jan
Net
Liq
%
Feb
Net
Liq%
Mar
Net
Liq%
Internal
6.31%
6.65%
6.35%
Agency
A
6.65%
6.78%
6.55%
Agency
B
5.18%
4.89%
4.99%
We
want
to
see
data
that
is
consistent
over
#me
as
in
scenario
B.
Having
data
like
scenario
B
allows
us
to
build
strategies
around
the
results
to
improve
performance.
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
7
8. For
now
ignore
the
#me
series
results
and
assume
results
are
stable.
What
other
ques#ons
must
we
ask?
Are
there
segments
within
the
por6olio
for
which
the
liquida#on
results
vary
substan#ally
from
the
team
average
result?
Let’s
look
at
one
example
with
two
segments
–
good
phone
#
and
bad
phone
#
(oken
called
“skips”).
Team
Good
#
Net
Bad
#
Net
Liq
%
Liq
%
Internal
11.22%
2.15%
Agency
A
10.39%
2.55%
Agency
B
7.50%
4.13%
Now,
how
should
we
allocate
accounts?
Are
you
s#ll
thinking
that
firing
Agency
B
might
make
sense?
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
8
9. How
should
we
find
meaningful
segments
to
consider
in
our
alloca#on
strategy?
We
need
to
ask
what
data
do
we
have?
What
variables
can
we
measure
and
analyze?
Typical
variables
to
consider
would
include
balance
size,
credit
limit,
interest
rate,
days
since
account
opened,
days
since
last
payment,
credit
score,
cash
advances,
etc.
One
strategy
is
to
take
all
data
fields
available
and
run
all
of
them
through
one
or
more
modeling
techniques
(Regression,
Chiad,
Cluster
analysis,
etc)
to
find
meaningful
varia#on
in
results
by
score
band
or
by
segmenta#on/cluster.
If
we
test
all
known
data
fields
that
we
have,
then
are
we
done?
Is
there
more
that
we
can
or
should
do?
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
9
10. There
is
more
data
to
be
had
if
you
ask
the
right
people.
Who
should
we
ask?
One
error
that
sta#s#cians
and
analysts
make
is
not
talking
enough
to
the
people
on
the
front
lines.
We
should
ask
our
internal
collectors
what
factors
seem
to
ma_er
in
who
pays
and
who
doesn’t.
We
should
ask
our
agency
vendors
what
factors
they
consider
as
important.
We
should
ask
the
Opera#on
managers
and
supervisors
for
their
input.
They
may
tell
you
things
like:
-‐
several
states
have
non
garnishment
laws
–
i.e.
create
a
cluster
for
that
-‐
u#liza#on
ma_ers
(balance
divided
by
credit
limit)
–
i.e.
create
a
transforma#on
variable
from
two
others
-‐
first
payment
defaults
ma_er
(never
made
a
payment)
-‐
someone
who
made
many
small
payments
before
defaul#ng
is
likely
to
pay
You
should
get
lots
of
ideas
to
create
new
variables
or
clusters.
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
10
11. Aker
speaking
to
our
Ops
people
and
vendors
and
crea#ng
transforma#on
variables
is
there
s#ll
more
data
to
obtain?
Oken
financial
companies
have
different
systems
of
record
for
origina#ons
and
for
servicing.
So
you
might
obtain
more
data
if
you
can
study
the
origina#on
data
as
well.
At
a
cost
you
can
also
obtain
poten#ally
important
data
from
outside
sources.
Most
common
sources
include
the
major
credit
bureaus
which
can
supply
data
on:
-‐
Are
they
paying
other
bills
on
#me
or
at
all?
-‐
How
many
other
debts
are
delinquent
or
charged-‐off?
-‐
How
much
total
debt
and
total
credit
do
they
have?
-‐
Do
they
have
a
mortgage
or
auto
loans?
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
11
12. Assume
we’ve
created
the
best
net
recovery
scoring/
segmenta#on
model
possible.
Now,
how
should
we
allocate
the
por6olio?
Assume
we
se_led
on
four
segments
as
follows
and
that
sample
sizes/
confidence
intervals
are
good
and
#me
series
results
appear
stable.
Team
Segment
A
Segment
B
Segment
C
Segment
D
Internal
9.69% 8.44% 5.04% 2.09%
Agency
A
11.35% 6.97% 5.18% 1.79%
Agency
B
6.20% 4.54% 5.57% 3.34%
Firing
Agency
B
now
appears
to
be
a
mistake,
but
what
should
we
do?
We
could
give
100%
of
each
segment
to
the
best
performer
as
circled
above.
Would
this
be
smart?
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
12
13. We
should
ask
are
there
logical
constraints
to
consider
in
op#mizing
our
alloca#on
strategy.
Some
logical
constraints
could
include:
-‐
Corporate,
Risk
or
Legal
considera#ons
including
requirements
to
always
have
two
or
more
vendors
or
possibly
limi#ng
share
to
no
more
than
70%
to
any
single
vendor.
-‐
We
may
need
to
keep
X
number
of
internal
employees
which
would
require
a
minimum
account
volume.
Likewise
there
may
be
a
hiring
limit
or
freeze
which
could
limit
new
volume
placements
-‐
We
should
keep
a
minimum
alloca#on
of
each
segment
to
each
vendor
to
watch
for
result
trend
changes
over
#me
which
do
occur
-‐
Some
vendors
might
have
capacity
limits
and
their
results
may
fall
if
given
too
many
addi#onal
accounts
too
quickly
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
13
14. Maximizing
net
recoveries
across
segments
with
many
alloca#on
requirements
/
constraints
may
be
best
solved
with
linear
programming.
Our
goal
objec#ve
would
be
to
maximize
net
recovery
dollars.
Assume
we
have
10,000
accounts
per
month
to
allocate.
Constraints
might
include
items
like:
-‐
Internal
min
=
2,000
and
max
=
5,000
with
a
change
of
no
more
than
X%
per
month
-‐
Each
Agency’s
share
must
be
>=10%
and
<=70%;
Agency
share
can’t
change
more
than
+/-‐
1,000
per
month
-‐
Agency
A
has
an
upper
capacity
limit
of
4,000
-‐
Each
team
must
get
at
least
100
accounts
per
segment
per
month
You
could
write
a
simple
Linear
Programming
Model
to
solve
/
op#mize
this
problem
using
“SOLVER”
in
Excel
or
choose
from
many
other
programs
.
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
14
15. What
else
should
we
consider?
Make
sure
that
before
your
models
/
segmenta#ons
are
finalized
that
other
departments
have
signed
off.
For
example
Legal/Risk
would
likely
not
let
you
use
full
Zip
code
as
a
variable
as
it
could
be
considered
red
lining.
Also
confirm
with
Opera#ons,
Vendor
Management
and
HR
what
you
are
planning.
Sudden
volume
shiks
are
likely
to
hurt
results
and
hiring/training
may
take
#me.
Opera#ons
likes
to
have
predictable
volumes.
Also
consider
the
difficulty
and
cost
of
gexng
certain
data.
Maybe
you
can
get
90%
of
the
model’s
power
from
using
only
three
variables.
If
true,
then
do
you
really
need
12
variables
in
your
model?
Also
it
is
very
important
to
publish
and
share
data
results
and
to
step
the
ground
rules
for
Internal
and
the
agencies.
Performance
has
a
way
of
improving
quickly
when
measured
and
reported
publicly.
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
15
16. Let’s
review
our
current
strategy’s
results.
Our
original
alloca#on
was
40%
to
Internal
and
30%
each
to
our
two
agencies.
We
will
assume
alloca#on
was
consistent
in
share
across
our
four
defined
segments.
We
will
assume
we
have
10,000
accounts
per
monthly
batch.
Segment A Segment B Segment C Segment D Total
Internal 1,256 740 944 1,060 4,000
Agency A 942 555 708 795 3,000
Agency B 942 555 708 795 3,000
Total 3,140 1,850 2,360 2,650 10,000
Avg. Bal $ $3,250 $5,105 $2,841 $3,088 $3,454
This
original
distribu#on
with
our
liquida#on
results
from
page
12
predicts
a
monthly
batch
net
recovery
of
$2,123,206.
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
16
17. Let’s
review
our
LP
Model
constraints.
• Sum
of
all
segments
=
10,000
• Agency
B
must
have
<=70%
share
• Sum
of
alloca#on
to
Internal
+
• Agency
A
has
a
capacity
limit
of
4000
Agency
A
+
Agency
B
=
10,000
• All
solved
values
must
be
integers
• Sum
of
segment
A
distribu#on
=
total
• Internal
Segment
A
>=
100
of
segment
A
• Sum
of
segment
B
distribu#on
=
total
• Internal
Segment
B
>=
100
of
segment
B
• Internal
Segment
C
>=
100
• Sum
of
segment
C
distribu#on
=
total
• Internal
Segment
D
>=
100
of
segment
C
• Agency
A
Segment
A
>=
100
• Sum
of
segment
D
distribu#on
=
total
of
segment
D
• Agency
A
Segment
B
>=
100
• Internal
must
have
at
least
2000
• Agency
A
Segment
C
>=
100
accounts
• Agency
A
Segment
D
>=
100
• Internal
can't
have
more
than
5000
accts
• Agency
B
Segment
A
>=
100
• Agency
A
must
have
>=10%
share
• Agency
B
Segment
B
>=
100
• Agency
A
must
have
<=70%
share
• Agency
B
Segment
C
>=
100
• Agency
B
must
have
>=10%
share
• Agency
B
Segment
D
>=
100
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
17
18. Now
let’s
see
our
op#mized
LP
Model
results.
We
assume
the
same
popula#on
and
distribu#on
of
segments
solved
to
maximize
net
recovery
subject
to
the
constraints
on
the
prior
page.
Segment A Segment B Segment C Segment D Total
Internal 100 1650 150 100 2,000
Agency A 2,940 100 100 100 3,240
Agency B 100 100 2,110 2,450 4,760
Total 3,140 1,850 2,360 2,650 10,000
Avg. Bal $ $3,250 $5,105 $2,841 $3,088 $3,454
This
op#mized
distribu#on
with
our
liquida#on
results
from
page
12
predicts
a
monthly
batch
net
recovery
of
$2,540,549.
This
predicts
a
lik
of
$417K
per
monthly
batch
or
19.66%
or
a
net
annual
improvement
of
$5M
per
year
on
a
batch
basis.
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
18
19. Are
there
other
constraints
to
consider?
Yes,
this
model
is
only
a
simple
example.
There
are
many
other
issues
to
consider
including
the
profitability
of
the
servicing
work
for
both
internal
and
external
vendors.
Collec#on/Recovery
agencies
usually
follow
the
unit
yield
on
their
client
assigned
paper.
Unit
Yield
=
Liquida#on
Rate
x
Average
$Balance
x
Commission
%
If
the
expected/actual
unit
yield
drops
significantly
the
vendor
will
either
be
forced
to
pull
resources
off
of
the
por6olio
or
they
could
actually
resign
from
being
a
servicer.
On
the
flip
side,
client
por6olios
with
a
high
unit
yield
can
demand
be_er
staffing
ra#os
and
more
experienced
staff.
Such
considera#ons
are
important
when
working
with
agencies
and
should
be
reflected
as
part
of
any
LP
Model’s
constraints.
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
19
20. Before
changing
to
the
new
op#mized
alloca#on
are
there
other
issues
to
consider?
Yes!
There
are
many
issues
to
think
through
before
making
the
changes.
• Internal’s
share
will
be
cut
in
half.
This
would
mean
cuxng
or
realloca#ng
half
the
current
internal
staff.
Are
we
willing
to
do
this?
Should
we
give
our
Internal
group
#me
to
improve
its
results?
• Will
Internal’s
cost
structure
change
with
a
large
reduc#on
in
volume?
• Agency
B
would
receive
58.7%
more
volume.
Can
they
handle
this
increase
and
if
so
then
how
quickly?
• Should
we
iden#fy
to
Internal
and/or
to
the
Agencies
which
accounts
are
which
segments
so
they
can
work
harder
on
the
higher
liquida#on
accounts?
How
will
we
share
results?
• Should
we
change
the
Agency
commission
rate
based
on
segments?
• How
oken
should
we
verify
the
results
and
alter
the
alloca#ons?
• How
oken
should
we
rebuild
the
segmenta#on
model?
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
20
21. Conclusions
on
Op#mizing
Recoveries
through
Smart
Alloca#on
Using
these
smart
analy#c
techniques
could
easily
improve
net
recoveries
by
10-‐20%
or
more
verses
a
tradi#onal
straight
share
alloca#on
method.
– Remember
to
ask
lots
of
ques#ons
– Measure
your
goal
objec#ve
(i.e.
net
recoveries
over
some
batch
period)
– Consider
cost
– Look
for
all
relevant,
usable
data
to
create
segmenta#ons
– Talk
to
Opera#ons
and
your
vendors;
talk
to
Legal,
Risk,
HR,
etc
for
their
input
– Check
sample
size
and
significance
(Hypothesis
tes#ng
and
Confidence
Intervals)
– Make
sure
the
trend
is
tracked
and
is
meaningful
(i.e.
don’t
want
to
see
wild
swings
in
performance)
– Consider
the
80/20
rule
when
building
a
model
/
segmenta#on
–
is
it
worth
the
complexity?
– If
there
are
many
constraints,
then
consider
using
LP
modeling
– Measure
and
publish
/share
the
results
by
segment
(Shine
a
light
on
things)
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
21
22. For
More
Informa#on:
For
more
informa#on
about
this
presenta#on
you
may
contact
the
author
at:
Blairkorschun@aol.com
LinkedIn:
www.linkedin.com/in/blairkorschun
Using
Data
Analy#cs
&
Cri#cal
Thinking
to
Beat
Your
Compe#tors
22