Integrated Analytical Methods,
Systems, Process in IMC’s
Michael F. Gorman, Ph.D.
University of Dayton,
MFG Consulting
INFORMS General Meeting –Phoenix
October, 2012
Integration of Methods Systems, Processes at IMCs
1. Background/Introduction: The Industry and Situation
2. Strategic Operations Research Modeling: The Justification
3. Real time Decision Support: The Challenge
Integrated suite of five modules using a variety of methods:
1.
2.
3.
4.
5.
Supply and Demand Forecasts
Capacity Valuation Model
Load Accept Optimization
Fleet Inventory Target Setting
Load Routing Optimization
4. Extensions, Lessons learned, other applications
2
The five stages of analytical competition:
Where are the IMCs?
3
2) Analytics Methods Modeling
The growing need for OR at IMCs
• IMCs are becoming more asset based as RRs shed container
assets
– Shift in industrial organization, cost structures, decision rules
• Previous decision rules:
– Load acceptance: Take all profitable moves
– Load routing: Take lowest cost (available) route and equipment
• With a fixed fleet, these rules leads to missed profitable
loads, and undesirable situations with routing
• Change in philosophy:
– “Network view” rather than a “one way” view of each load
– Not all profitable moves are good moves:
• Repositioning costs – must balance fleet, unlike with rail-owned
• Opportunity costs - allocation of scarce, fixed fleet resources
– Lowest “one way” cost does not lead to lowest network cost
• Consider better opportunities for alternative orders at origin
• Equipment value from subsequent loads at destination
4
2) Strategic Fleet Allocation Model (FAM)
Conceptualization
Inputs
Model
Historical
Container
Traffic Flows
Recommended
Fleet
Assignments &
Utilization
Fleet
Contract
Provisions
Repositioning
Costs and
Times
Transit and
Street
Times
Outputs
Fleet Allocation
Model (FAM)
Network
Profitability
Ramp
Imbalances and
Repositioning
Cost
Strategic modeling shows the benefit, justifies the investment in real-time systems.
5
3) Real Time Decision Support
Project Scope Considerations and Constraints
• We wanted to take a conservative approach to an
aggressive project
– Relatively small company
– Relatively new to OR-based DSS
• To reduce project disruption, risk and cost, we applied
these constraints to our approach:
– Organizational:
• Business Process cannot change
• Marketing-centric Customer Service Representative (CSR) and
Operations-centric Dispatcher responsibilities cannot change
– Technological: optimization technology must fit within existing
production system and data structures
• In particular – “transactional”, one-at-a-time order processing
(both acceptance and dispatching)
6
3) Real Time Optimization – Organization and Process
Load Accept and Routing Decisions are Sequential
Load Accept
Optimization
(LAO)
Customer
Service Rep
Load Routing
Optimization
(LRO)
Dispatcher
First, the Load Acceptance decision (Marketing Decision)
– Should Customer Service Representative (CSR) accept
a tendered load?
To maximize contribution of scarce resources
Second, the Load Routing decision (Operating Decision)
– Which equipment/railroad should dispatcher choose?
To minimize cost, given a set of accepted loads
7
3) Real Time Optimization: Technological
Architecture Overview and Information Flow
Production
Transaction User Interface
Business Process
System
Network
Optimization
Program
(NOP)
Demand
And
Supply
Forecasts
Load
Accept
Optimization
Load
Routing
Optimization
Capacity
Valuation
Model
Fleet
Inventory
Targets
Fleet
Allocation
Model
8
3) Real Time Optimization
Modeling Component Overview
Forecasts
LAO
LRO
CVM
FIT
Network Optimization Program (NOP) suite of integrated tools
Real-time yield
management tool
determines load
acceptance
thresholds
Daily
forecasts
with real time
updates
Demand
And
Supply
Forecasts
CSR
Dispatcher
Load
Accept
Optimization
(LAO)
Load
Routing
Optimization
(LRO)
Capacity
Valuation
(CVM)
Fleet
Inventory
Target (FIT)
Daily estimates of the value
of capacity at a location on
a date; updated with
changing supply
Assigns accepted
loads to equipment
and routes in a
least-cost way
Seasonally, set daily
inventory targets for
container fleets
LAO and LRO interact directly with CSR and dispatcher.
Novel methods are detailed in the CVM and FIT support modules.
9
3.1) Daily Supply and Demand Forecasts
Forecasts
LAO
LRO
CVM
FIT
•
Container Supply forecast is in two parts:
– “Controlled” capacity – Timing of current and near-term customer
releases each day – quantity is known, timing is not
– “Predicted” capacity - “Street Sourced” – expert opinion on
available rail-owned supply off the “street”
•
Demand Forecast:
–
–
–
–
Two week horizon of daily orders
Time-series based demand forecast
Hundreds of estimated equations
Aggregate forecast of smaller lanes so that all demand is forecasted
•
Critical inputs for decision support models;
Valuable new source of information for managers
•
Challenge: What is the appropriate level of forecast detail to
support decision makers/models, but maintain accuracy?
10
3.2) Capacity Valuation Model – CVM
Overview
•
Forecasts
LAO
LRO
CVM
FIT
Capacity Valuation Model – CVM
– Establishes the value of an additional unit of capacity
– Output is Capacity Valuation Curve (CVC) –
relationship between expected margin and capacity
•
Capacity Valuation Curve (CVC) depends on
– Demand Forecast and forecast error distribution
– Margin distribution of market segments and ODs out of an origin
•
Capacity Value (CV) is a point on the CVC at a point in time
– Value is measured by the expected profit potential generated by the
unit of capacity
•
CV is an input to LAO and LRO
– for both establishing the opportunity cost of accepting a load at
origin, and
– for estimating the value of capacity at destination
11
3.2) CVC Illustration
Forecasts
LAO
LRO
CVM
FIT
Capacity Valuation Curve
$500
$450
Expected Profitability
$400
$350
$300
$250
$200
$150
$100
$50
$1
2
3
4
5
6
7
8
9
10
Container Supply
The CVC captures the diminishing marginal value of
container capacity as capacity on a date increases.
In this example, supply of 6 has CV=$72.30
12
3.3) Load Acceptance Optimization (LAO)
Overview of Methodology
Forecasts
LAO
LRO
CVM
FIT
•
Load Acceptance Optimization maximizes the anticipated value of a
tendered load against other future potential loads at that origin
•
Accept a tendered load if: Load Value >= Acceptance Threshold:
Tendered Load Profitability + Tendered Load Destination CV
>=
Origin Load CV + Average Destination CV
•
The simple decision rule captures the opportunity costs of accepting a
customer order as determined by the CVC
•
It also includes the future potential created by the accepted load at the
load’s destination, as compared to other loads’ future potential
13
3.4) Target Inventory Setting
Conceptualization
Forecasts
LAO
LRO
CVM
FIT
•
Target Fleet inventory balances the cost of a Fleet box against the
increased profit potential of a Fleet over other equipment types
•
Target Inventory is greater than zero because:
– Fleet cost advantages
– Mismatch in timing of inbound vs. outbound pattern of fleet flows
– Variation in IB (supply) and OB (demand) patterns
•
A Unique Two-step Algorithm was developed:
– 1) Target inventory setting - ideal inventory of fleet containers to
maximize the incremental profit potential over other capacity sources
– 2) Cost of Inventory Target Deviation - establish shortage and
surplus costs
•
Targets are set monthly as demand patterns change;
they are used daily in LRO for making routing decisions
14
3.4) Target Inventory Setting
Cost of Deviation from Target Inventory
Failed
Service to
Customer
Cost ($)
Higher-Cost
Rail-Owned
Container
Assignments
Forecasts
LAO
LRO
CVM
FIT
Fleet
Repositioned
Due to high
Surplus
Fleet Assigned to
Suboptimal Lanes
Due to excess Supply
Step 1: Establish
Fleet Target
Inventory (units)
We establish a fleet inventory target and estimate a cost of deviation from that target.
15
3.5) Load Routing Optimization Formulation:
Integration of all the pieces
One-way cost
of shipment
N M
Minimize cost
K
k1
M
CVM
FIT
Inventory and
Service failure costs
(1)
k 1 t 1
CVM Estimated
Capacity Value
at Destination
T

LRO
(Cijk + TIikt, + TIjkt - CVjkt + Invjkt + SVjkt ) xijkt
Subject to:
K
LAO
T
 
i 1 j 1
FIT Orig/Dest
Target
inventory
Forecasts
xijkt = Dijt
for all (i,j)
(2)
xijkt <= Sikt
for all k, for all i
(3)
t 1
T

j 1 t 1
16
4) Successes and Lessons Learned
• What we did well:
– Delivered wins early; earned quick paybacks
• Strategic model “paid for” more aggressive production models;
Forecasts were used for planning months before optimization was
in place
– Mapped the business process, information flow, decision
points and systems integration before modeling
• Stayed on scope, under budget – increased returns
– Stayed small for minimal disruption - reduced risk
• No change to the organization, the load tendering process, or the
system’s user interface
• Lessons we learned
– We could have started with simpler costing, then added
complexity
– Simplicity gains confidence– compliance is hard.
(and some sophisticated costing has little impact on decisions)
– Incurring near-term costs to obviate future costs or gain future
benefits is challenging in an economic downturn!
17
4) Extensions
•
•
•
Assignment of equipment to load also affects origin and
destination dray
Dray carriers/location/box holding affects network economics
Requires
–
–
–
–
–
•
Very detailed data
Very detailed forecasts
Longer look-ahead
Managing shipper loading behavior
Managing dray company behavior
Bid response/Pricing
– Very challenging to price given dramatic change in cost structure due
to network effects when a change in product mix
– Estimation of customer response to pricing initiatives a challenge
– Especially true in an all-or-nothing bid style pricing problem
18
Thank you!
Michael F. Gorman, Ph.D.
University of Dayton
Phoenix INFORMS
October, 2012
Flows including dray moves increase problem size
• The number of OD pairs will be much larger, and will depend on
the level of detail for customer/consignee ‘location’
–
–
–
–
Ramp area
Dray zone
City/state/zip location
Customer lat/long
• How we estimate dray efficiency changes with ‘location’
specification (more specific gives better estimate of dray
efficiency)
– Aggregated location:
• Imbalance Based – net flow balance contributes to efficiency
– Specific location:
• Triangularization – pairing IB/OB loads; empty move distance
20
Customer Specific Locations
•
Centroid of Dray zone or zip code might misstate specific cross-region
dray distances/costs
– Both overstate for proximate locations, and understate for remote ones
•
Specific customer locations allow better estimation of empty miles,
costs
City/St/Zip
Customer
Dray
Zone
CHI
City/St/Zip
Customer
Dray
Zone
21
Potential Problems with
More Specific “Locations”
•
Data sets grow –
– Increasing data management/data quality issues
– Increasing challenge of results interpretation
•
Data requirements grow –
– E.g., Lat/long of ramp, customer/dray area/city/state/zip
– Estimates of distance, dray costs for a large number of pairings
•
Assumptions surrounding feasibility grow –
– Customer1-Customer2 pairing in ‘plan’ – will it happen?
– Seasonality, day of week, delivery/pickup windows, etc.
•
Assumptions surrounding constraints grow –
– Can we ‘never’ dray two different box types in the same move?
More demanding data and modeling to get more ‘realistic’ outputs
Can create ‘brittle’ models that fall apart
when data is bad or assumptions are violated.
22
More detailed destination location
• For “Major” Ramps, Dray Zones, Customers:
City/St/Zip
Dray
Zone
City/St/Zip
Consignee
Dray
Zone
Consignee
CHI
LAX
City/St/Zip
Shipper
Dray
Zone
City/St/Zip
Dray
Zone
Shipper
• A Move is
Shipper/Ramp - Consignee/Ramp–
• 700 Ramp pairs can become 7000 Shipper-Consignee pairs
23
Considerations for level of specificity for location
• Volume
– If there is not high volume, then any anticipated triangularization is
suspect – need large numbers to be predictable
– Also, with low volume, potential impact on economics is lowered
• Proximity to Ramp
– For locations near ramp, empty moves are not as costly, and
potential for dray efficiency gains lower
• Geographic size of locations
– If a location encompasses a small geographic region, gains from
more specific location definition is diminished
• Proximity/Interaction of locations
– The level of proximity and actual or possible interaction between
locations drives the need for more specificity
24