Académique Documents
Professionnel Documents
Culture Documents
Air freight is inherently expensive shipping high-value cargo in a fast, safe and time-bound
manner comes at a high cost. Capacity comes at a premium. Specific handling requirements,
packaging instructions, expedited delivery etc. further hike the bill. Air freight is also heavily
impacted by aviation fuel price fluctuations. Given the high cost of air transportation, the survival
of a LSP depends upon its ability to strike a favorable rate agreement with the carrier and to
identify the best rate for a shipment.
Air shipments are moved either as loose cargo or as containerized cargo. The container in air
parlance is known as ULD. These are relatively small containers (compared to ocean shipments)
with pre-defined volume, weight and tare weight specifications. A ULD can be loaded either in the
upper deck or lower deck and this lends additional complexity to the loading. The specifications
depend upon this classification as well. The freight charge is calculated for the loose cargos and
ULDs in the booking and this is where the cookie crumbles
Freight charging in air freight industry follows complex rules for context-driven cost calculation.
The rates come in multiple variants ULD rates, rates for general cargo, commodity specific rates
etc. with multiple weight breaks. There may be rate dependent rates (e.g. class rates).
Furthermore, the applicability of a particular type of rate may be contingent to the cargo type and
shipment conditions in the shipment order. So a containerized cargo (or ULD in air parlance) may
be charged using ULD rate or loose cargo rate (GCR) depending upon the shipment conditions and
contractual agreements.
So any cargo over 100 kg would be charged 3 USD/ KG whereas a cargo exceeding 300 kg would
enjoy a lower rate of 2 USD/ kg. So this rate structure offers lower rate at higher weight and thus
promotes consolidation. But theres another aspect to it which is known as break-weight. This
determines whether it is more economical to charge the consignment on the actual chargeable
weight or the next weight break. So for example, if the consignment weight is 280 kg then it is
better to declare the weight as 300 kg as this would fetch a lower rate and subsequently a much
lesser charge.
The following table shows some representative loose cargo charge over a busy air route for a wide
range of weights. The numbers in bold show the effect of weight break. So freight charge works
out to be 2934.58 HKD for 125.94 KG if weight break is applied; which is considerably lower than
3351.31 HKD (the linear charge on 125.94 kg).
The accompanying graph brings this out more clearly. The peaks in the graph indicate potential
trade-off zones. For example, the linear freight charge beyond ~ 60 kg is costlier than the next
weight-break charge and that offers an opportunity for break-weight calculation
Fig1
ULD rate
ULD rating is more complex than loose cargo rating. The weight capacity of a ULD is rated in a
step-wise manner, i.e., ULD rating is a step function. Lets assume that a particular ULD has a net
capacity of 800 KG. The carrier offers a flat charge of 9000 HKD for any weight up to 460 kg loaded
in it. Any incremental weight after 460 kg is charged at 14.9 HKD per kg till the chargeable weight
reaches 727 kg. A flat charge of 11834 HKD is levied at 727 kg. A quick computation would show
that the flat charge at 727 kg is less than the linear charge and here lies the opportunity for
optimization. The following table and the accompanying graph validate this assumption. For
example, the chargeable weight of ~580 kg is the inflection point where the next weight-break
charge becomes more attractive than the linear charge.
Fig3
ULD vs. Loose cargo rating : optimization opportunity
Having examined loose cargo and ULD rating individually, lets now explore how they compare
over a relatively large weight distribution. A comparative charge snapshot is shown in the table
below. The accompanying graph indicates that loose cargo rating is more cost-effective for
relatively low cargo weights but ULD rates win by miles when it comes to high cargo tonnage. Not
surprisingly, consolidated cargo is generally charged with ULD rates.
580
60
Fig4
The inflection point where ULD rating becomes more economical than loose cargo rating comes at
about 400 KG. Lets now crystallize the key insights:
1. Weight breaks in loose cargo rating provide charge minimization opportunity locally
2. Weight breaks (Pivot Weights) in ULD rating also provide local charge minimization
opportunity locally
3. ULD rating becomes more economical than loose cargo rating beyond a cargo weight. Lets
call it the inflection point
A standard approach to arriving at the minimum freight charge for a ULD could be to compute
both loose cargo and ULD rates and take the minimum. But this is a wasteful approach,
considering that one of the calculations would be discarded. Furthermore, it does not make use of
the intuition that an inflection point exists.
An alternate approach would be to train the system in solving a classification problem given
certain characteristics, should a cargo be charged with loose cargo or ULD rate. The following
section discusses about a possible framework that builds upon this approach
The learning set is a collection of actual output based upon actual input values. The actual inputs
are feature values that can be extracted from a rate table. The actual output is the type of rating
(Loose cargo or ULD) that minimizes the calculated freight charge. So the learning set gives exact
values.
The training set is fed to the adaptive algorithm. The job of the algorithm is to generate a function
that would map the input features to an output value. Ideally this function would help us to
predict whether loose cargo or ULD rating would be more economical for new input sets.
Fig5
We define the behavior of f(x) as the probability of y= 1 (Rating = ULD rating) for x parameterized
on
To find out the parameter value, we bring in a cost function. The cost is defined as the penalty in
making a wrong prediction of output value (Y) for some given values of X. So minimization of the
cost function would potentially lead us to the optimum parameter values.
Lets define our primary cost function as:
So if f(x) = 1 when y = 1 then cost = 0. But if f(x) 0 when y = 1 then cost . So this indicates
that if f(x) = 0 (which implies P(y=1) = 0) but y = 1 then the algorithm is penalized with a very high
cost
Finally, we define the logistic regression cost function as
Once we minimize the cost function using some advanced optimization technique line gradient
descent, the parameter values () would be calculated. These parameters can then be used to
classify new input sets. Based upon the classification, the freight calculation logic would directly
calculate the charges.
2.5
2 Tact Rate GCR
1.5
Tact Rate GCR
1
0.5
Contract GCR
0
0
10
20
30
45
60
80
100
120
140
Fig6
Step2: Backward Recursive method to find the lowest cost option. Basic issue is that at any stage,
user would like to find the lowest possible cost for the booking. This is the cost which can be the
combination of various rates for different weight ranges, see fig7
Fig7
Problem can be formulated as network flow problem where each node (state) is the location and
distance is the cost to move to the next location (next node) as in Fig8.
Fig8
Formulation:
Stage: as defined above stage are the key inflection points in the rate combinations
State: is the node in a stage represented by Ny;
Decision variable: XN is the node where to go for lowest cost
Objective function: Minimize total Cost
Fn(S,Xn)=Cs,Xn +F*n-1(Xn)
F*n(S)=Min Fn(S,Xn) and Xn* is the corresponding node ( or Vector Ny or ijTR N (Ws))
Step3: Save all the possible moves from the node in the value table.
Step4: Order comes with different chargeable weight and attributes. Different value of Dijb and
dijc.
Step 4.1: Only relevant cost functions are activated ( ijTR ) based on available attributes in the
order. All the relevant nodes Ny-will and stage will be selected for evaluation.
Step4.2: assign the current Dijc to stage S-1, find the nearest stage S-1 from already available set
such that WS-2 <Dijc. Regenerate the nodes for stage S-1.
Step4.3: Find the lowest cost demand mix. Check if the best value (minimum cost values are still in
the set of each stage) then do forward recursion from Stage S-2 to S. Otherwise go to
Step4.4: This requires already generated DP run. We call this a sub run as few of the values will not
be available as those attributes are not applicable for this run. This will give quick minimum cost
for the booking.
Conclusion:
Operational air freight cargo rating depends on many attributes (e.g. service levels, commodity
code and handling code), rating also depends on whether cargo is packed in ULD or it is treated
loose. Above strategies impacts planning as well rating in a booking.
As mentioned in the paper, forwarder can optimize the cost in booking by using various strategies
e.g. inclusion and exclusion on various attributes, consolidation etc.
Paper provides the depth and breadth of the problem and gives practical insights which can be
used by forwarders.
Forwarders can use either simple analytical tool or advance solution to find the optimal cost.
We have also experimented with machine learning models and dynamic programming which can
come handy for multi-attribute and context sensitive pricing.
References: