Multi-layered self-calibrating analytics

1. A computer-implemented method executed by one or more processors for improving a fraud detection model, the method comprising:

providing one or more variables from a set of variables to a plurality of self-calibrating models, the one or more variables being generated from real-time production data related to the transaction data,

outlier values of at least one variable changing due to changing transaction dynamics over time, and

a first set of one or more variables being provided to a first self-calibrating model and a second set of one or more variables being provided to a second self-calibrating model;

processing the one or more variables according to the plurality of self-calibrating models implemented to produce at least two self-calibrating model outputs;

combining the at least two self-calibrating model outputs, the self-calibrating model outputs being tuned based on a first weight being assigned to a first self-calibrating model output and a second weight being assigned to a second self-calibrating model output, supervised training being used to limit amount of data needed to tune at least one of the first weight or the second weight; and

generating an output for the real-time production data from the self-calibrating model output, the output representing whether the transaction is fraudulent.

2. The method in accordance with claim 1 , further comprising assigning a weight to a self-calibrating model output.

3. The method in accordance with claim 2 , wherein the output model processes the output from the plurality of self-calibrating models according to the weight.

4. The method in accordance with claim 3 , wherein assigning a weight to the self-calibrating model output includes assigning a zero weight to a subset of one or more of the plurality of self-calibrating models to designate the subset as experimental self-calibrating models.

5. The method in accordance with claim 1 , wherein at least one of the plurality of self-calibrating models uses a linear combination of the one or more variables from a set of variables to determine a fraud score that is related to a number and a size of one or more outlier values associated with the one or more variables.

6. The method in accordance with claim 5 , wherein the one or more outlier values are based on a real-time computation of variable distributions of the one or more variables.

7. The method in accordance with claim 6 , wherein the one or more outlier values are updated recursively for a new set of transaction data.

8. A method for detecting fraud in transaction data, the method comprising:

executing, by one or more data processors, a plurality of self-calibrating models on one or more variables selected from a set of variables, outlier values of a variable changing due to changing transaction dynamics over time, and the one or more variables being generated from real-time production data related to the transaction data;

producing a self-calibrating model output for each of the plurality of self-calibrating models based on the selected one or more variables, each of the plurality of self-calibrating models using a linear combination of the one or more variables from a set of variables to determine a fraud score that is related to a number and a size of one or more outlier values associated with the one or more variables based on:

q

⁡

(

x

i

|

θ

)

=

x

i

-

θ

i

,

1

θ

i

,

2

∈

[

0

,

C

]

,

where ( (θ i,1 ,θ i,2 )∈θ ) are location and scale parameters respectively of a computed distribution of independent variable x i , wherein the fraud score is related to a number and a size of one or more outlier values associated with the one or more variables;

combining the self-calibrating model outputs from the plurality of self-calibrating models in an output model implemented by one or more data processors, the self-calibrating model outputs being tuned based on weights assigned to the self-calibrating model outputs; and

generating an output for the real-time production data from the self-calibrating model output, the output representing whether the transaction is fraudulent.

9. The method in accordance with claim 8 , wherein the one or more outlier values are based on a real-time computation of variable distributions of the one or more variables.

10. The method in accordance with claim 9 , wherein the one or more outlier values are updated recursively for each new set of transaction data.

11. The method in accordance with claim 8 , further comprising assigning a weight to each self-calibrating model output.

12. The method in accordance with claim 11 , wherein the output model processes the output from each of the plurality of self-calibrating models according to the weight.

13. The method in accordance with claim 12 , wherein assigning a weight to each self-calibrating model output includes assigning a zero weight to a subset of one or more of the plurality of self-calibrating models to designate the subset as experimental self-calibrating models.

14. The method in accordance with claim 8 , wherein combining the self-calibrating model output from each of the plurality of self-calibrating models in an output model includes combining the fraud score from the self-calibrating model to determine the score based on an inputted threshold.

15. A computer-implemented system comprising:

at least one programmable processor; and

a machine-readable medium storing instructions that, when executed by the at least one processor, cause the at least one programmable processor to perform operations comprising:

execute a plurality of self-calibrating models on one or more variables selected from a set of variables, the one or more variables being generated from real-time production data related to the transaction data;

produce a self-calibrating model output for the plurality of self-calibrating models based on the selected one or more variables, the plurality of self-calibrating models using a linear combination of the one or more variables from a set of variables to determine a fraud score that is related to a number and a size of one or more outlier values associated with the one or more variables;

combine the self-calibrating model output from the plurality of self-calibrating models in an output model implemented by one or more data processors; and

generate an output for the real-time production data from the self-calibrating model output, the output representing a score relating to whether the transaction is fraudulent.

16. The system in accordance with claim 15 , wherein the selection of one or more selected variables is based on factor group analysis to minimize correlation among selected variables for the plurality self-calibrating models.

17. The system in accordance with claim 15 , wherein the one or more outlier values are based on a real-time computation of variable distributions of the one or more variables.

18. The system in accordance with claim 17 , wherein the one or more outlier values are updated recursively for a new set of transaction data.

19. The system in accordance with claim 18 , further comprising assigning a weight to the self-calibrating model output.

20. The system in accordance with claim 19 , wherein the output model processes the output from the plurality of self-calibrating models according to the weight and assigning a weight to the self-calibrating model output includes assigning a zero weight to a subset of one or more of the plurality of self-calibrating models to designate the subset as experimental self-calibrating models.