The talk would be focusing on reasons and method for creating models which maximize sales price Gross Margin but still has high confidentiality that quote would be accepted by the client. Price changes are dynamic things that are impacted by many different elements like cost of input material, labor cost, transportation cost, scrap material due to different ordered quantities, etc. Besides input cost segments, output price is also impacted by different marketing campaigns (own and others), seasonality, past and future customer behavior as well as the behavior of the product we are selling.
6. Motivation
• For every product, we should be able to find
the optimal price that a customer is willing to
pay
• We would factor in highly specific insights
that would influence the price
• Customer has ~750k of custom products – manual
practices for setting prices make it virtually
impossible to see the pricing patterns
• Dynamic deal scoring integrated into the core
commercial process will help organization to
price smarter, streamline the approval process,
and win more deals
7. Fact finding mission
• Sales reps often argue that higher discounts
are necessary to win deals
• FALSE - research across companies consistently
shows the opposite to be true: successful deals
almost always have lower average discounts than
deals that were lost
• A pricing infrastructure can be difficult and
costly to create
• TRUE - It requires investing appropriately,
empowering the right people, articulating clear
targets and goals, and managing risk, but….
8. Dynamic Deal Scoring
• is certainly not new, having been pioneered by
airlines in the ‘70s and now widely used by
hotels, airlines, and car rental and tour
operators
• it’s not only being used in online
environments. Traditional bricks and mortar
companies are also revolutionizing how they
monetize, finding new ways to tap into
customers’ willingness to pay
• can be located within a broader move towards
customization and personalization with the
emergence of data
9. Dynamic Deal Scoring
• On average, a 1% price increase translates into
an 8.7% increase in operating profits (assuming
no loss of volume, of course)
• 30% of all pricing decisions companies make
every year fail to deliver the best price!!! –
a lot of lost revenue
• CAUTION! Dynamic Deal Scoring when used blindly
or in isolation , i.e. without human oversight
and a clear strategy, can at worst destroy
reputation and erode consumer trust
12. Hard work and long nights …
• Hypotheses
• Apply experiments
• Read measurement
• Deduce next steps
• We used CRISP methodology
13. Six steps to build the model
Data prepStep
▪ A piece of R code
takes the invoice
data and:
– Filters out
data that
should not be
considered in
the model
– Creates
synthetic
variables based
upon the
existing data
– Combines
invoices into
deals (there
are multiple
invoices
associated with
each deal)
▪ A piece of R code
calculates the
margin cutoffs
for A, B, C, D,
and F deals
within each node
▪ An tool shows the
distribution of
grades across
various
dimensions
including
Industry, and
Product Group.
Where there
exists a strong
skew towards A
grades or F
grades, the tool
will be used to
adjust margin
cutoffs to reduce
the uneven
distribution
▪ The DW team
provides invoice
data with margins
adjusted to match
the P&L and
additional
variables mapped
in from sources
outside the ERP
▪ The DDS model
embedded within
SalesForce
Detail
Input data DDS tree building Scoring Adjustments Output to tool
1 2 3 4 5 6
▪ Building the tree
involves two sub-
steps:
– Choosing
appropriate
variables based
on business
logic and
predictive
power
– Using the
regular method
to grow the
tree within the
bounds of best
practices. All
deals will be
sorted into
nodes of
similar
business
characteristics
and margin
3a
3b
14.
15.
16. Data prep
Data prep flow
This flow filters data, applies sales adjustments, and applies cost
adjustments
X
Sales adjustments
Cost adjustments
X
X
Filters
Combined invoice file
=
Sales + cost file by invoice
The output is a file with the sales and cost for each invoice
line. This file is mapped into the R data prep code
17. Data prep
Synthetic
variables
▪ Top Customer: This variables has 7 values; the 6 largest customers
and “other”
▪ Annual customer spend: The spend for the top customer for 2018
▪ Product Super Group: Assigned based on product group
▪ No of items in platform: sums up number of unique items within a
platform
▪ Deal size: Sum of invoice sales for that deal (as defined below)
▪ Spend per customer per item: % of a customer’s annual sales
represented by that item
▪ Platform: binary yes/no; is that item part of a platform
▪ Make vs Buy: binary make/buy
▪ No items in quote: “Quote” is same Sales Rep, same Customer, same
Order Date
Filters
▪ Intercompany sales
▪ Tooling and engineering charges
▪ Returns
▪ Invoices with sales > $10M or < -$10M
▪ Deals with margin > 100% or < 0%
▪ MOV/MOQ’s
Combining
invoices into
deals
▪ Deals are defined as the same customer, same item, same invoiced
price; all invoices fitting this criteria are combined
18. DDS tree building
Choosing
Variables
DDS Tree
3a
3b
▪ Before running the tree code, it is necessary to select final list of
variables to be used in tree
▪ The selection criteria should be a combination of the relative
predictive power and the business logic of the variables
▪ The tree should not branch on something that shouldn't be
prescriptive, even though it is descriptive of today’s behavior
(e.g., month of year)
▪ The variable should be capturable as an input during quotation
process
▪ With the list of variables finalized, manually select the branches of
the first two or three levels of the tree
▪ The remaining levels can be built automatically (allow the code to
select the most logical branching), but throughout the process, apply
best practices to guide and adjust the tree
▪ The “N” of terminal nodes should be no less than 50 and the P-value
of each branch should be below 0.05
▪ Best practice is for the R-square of the tree as a whole to be
greater than 30%.
19. Validate against Best Practice
Proposed modelBest practice
4220+
175-10+
31%130%+
Number of deal parameters analyzed
Number of deal parameters found predictive and
used for deal scoring
Predictive power of the scoring algorithm (R2)
Price realization uplift expected 2-3%2-7%1
23. Project outcome…
• Project engagement: 8 weeks over 3 month period
• Project cost: 90% of cost was pure R’n’R –
traveling and accommodation cost due client
world wide presence
• 3 for 3 : project was fully paid of by correct
and timely pricing within 3 months of use in
production and measured above product price
baseline
• now in Revenue share mode