|
Cost-Benefit
Matrix
A cost-benefit
matrix is an input to the modeling process that allows predictive modelers
to describe the costs and the benefits associated with each possible prediction.
By default the cost-benefit matrix has a value of one (1.0) for correct
predictions and zero (0.0) for incorrect predictions. This configuration
asks that the predictive model optimize raw accuracy. In most real-world
situations, however, an incorrect prediction has a net monetary cost (less
than zero), and a correct prediction has a positive benefit. The correct
or incorrect values that are chosen affect the values chosen for the matrix.
The default cost matrix assumes no weighting for each output possibility.
When the cost-benefit matrix has new non-default values assigned, the
model optimizes the net benefit (profit) associated with each prediction.
The cost-benefit matrix input is essential for businesses that want to
optimize their return on investment. PredictionWorks supports the use
of a cost-benefit matrix.
The purpose
of the Cost-Benefit Matrix is to properly weigh each possible outcome
based on business value metrics. If there is a significant difference
in cost and benefit between the outcomes predicted by a model then model
selection should not be based on raw accuracy but on optimizing benefit.
In business the benefit is predicting profitability of campaigns. To keep normal weighting
of outcomes for the cost-benefit matrix, simply leave the default values
in place. The default settings will pick the most accurate model. If predicting
certain events whose outcomes are more important than others, then weighting
is necessary. For commercial use,weightings are measured in a monetary
value to the business. Certain events that have negative effects if wrongly
predicted can have negative values. For instance in predicting which customers
would respond to a direct mail campaign for opening a new savings account
with a financial institution, there are associated benefits and costs.
The company can either send or not send a mail-out. For each customer
that does open an account, that benefit could be $75. However, each mailout
has an associated cost of $5, and for each customer that would have opened
an account but was not on the mailing list has an associated cost of -$70.
The final Cost-benefit matrix would have the following values:
- If you do not send a customer the mail-out and they would not respond
- VALUE=0
- If you send a customer the mail-out and they would not respond - VALUE=
(-$5)
- If you do not send a customer the mail-out and they would have responded
- VALUE= (-$70)
- If you do send a customer the mail-out and they would respond - VALUE=
$70
Therefore,
we are optimizing on the total profits to determine the final result of which customers should receive a promotional mail-out."
Back to Glossary
|
