Deep learning for credit controls

Patent metadata

Field	Value
Publication number	US-11625569-B2
Application number	US-201715467632-A
Country	US
Kind code	B2
Filing date	Mar 23, 2017
Priority date	Mar 23, 2017
Publication date	Apr 11, 2023
Grant date	Apr 11, 2023

Abstract

Systems and methods are provided to identify abnormal transaction activity by a participant that is inconsistent with current conditions. Historical participant and external data is identified. A recurrent neural network identifies patterns in the historical participant and external data. A new transaction by the participant is received. The new transaction is compared using the patterns to the historical participant and external data. An abnormality score is generated. An alert is generated if the abnormality score exceeds a threshold.

First claim

What is claimed is:

1. A computer implemented method comprising:
detecting a deviation from prior activity, inconsistent with current conditions, by a participant in a data transaction processing system in which data items are transacted by a hardware matching processor that matches electronic data transaction request messages for the same one of the data items based on multiple transaction parameters from different client computers over a data communication network by:
identifying, by an activity module, in a memory, historic participant transaction data for the participant and historic external market factor data, wherein the historic external market factor data includes data indicative of characteristics of previous transactions by other participants and characteristics of a financial derivative product for a time period that corresponds to the historic participant transaction data that occurs during the time period, wherein the financial derivative product is traded on an exchange;
identifying, by the activity module, one or more patterns in the historic participant transaction data and the historic external market factor data using a structured neural network, the structured neural network comprising a layered plurality of interconnected processing nodes, wherein each connection of the plurality of interconnected processing nodes to another is dynamically weighted, the one or more patterns indicative of historical normal activity by the participant in relation to the historic external market factor data;
receiving, by the activity module, from the participant, an electronic data transaction request message comprising data indicative of a current transaction specifying a product, quantity, and value, and storing the electronic data transaction request message in the memory;
calculating, by the activity module, current external market factor data;
comparing, by the activity module, the data indicative of the transaction and the current external market factor data with the one or more patterns;
generating, by the activity module, an abnormality score for the electronic data transaction request message based on the comparison; and
generating, by the activity module, an alert when the abnormality score exceeds a first threshold.

2. The computer implemented method of claim 1 , wherein identifying the one or more patterns comprises:
encoding, by the structured neural network, the historic participant transaction data and historic external market factor data using a plurality of first layers of the layered plurality of interconnected processing nodes;
decoding, by the structured neural network, the encoded data using a plurality of second layers of the layered plurality of interconnected processing nodes;
comparing, by the structured neural network, the decoded data with the historic participant transaction data and historic external market factor data; and
identifying, by the structured neural network, the one or more patterns in the layered plurality of interconnected processing nodes when the decoded data is within a predefined distance of the historic participant transaction data and historic external market factor data.

3. The computer implemented method of claim 2 , wherein the plurality of first layers comprise a decreasing number of nodes in each layer of the plurality of first layers, and the plurality of second layers comprise an increasing number of nodes in each layer of the plurality of second layers.

4. The computer implemented method of claim 3 , wherein only the outputs of the smallest layer of the plurality of first layers is connected to the largest layer of the plurality of second layers.

5. The computer implemented method of claim 1 , wherein the layered plurality of interconnected processing nodes comprises a plurality of long short term memory nodes.

6. The computer implemented method of claim 1 , further comprising:
updating, by the activity module, the historic participant transaction data and the historic external market factor data with the current transaction data and current external market factors.

7. The computer implemented method of claim 1 , further comprising:
prohibiting, by the activity module, the current transaction from being processed when the abnormality score exceeds a second threshold.

8. The computer implemented method of claim 1 , wherein the interconnected processing nodes comprise long short term memory units.

9. The computer implemented method of claim 1 , wherein the historic participant transaction data comprising data relating to a single or related set of products.

10. The computer implemented method of claim 1 , further comprising:
calculating, by the activity module, a participant risk profile including the data indicative of a current transaction; wherein the one or more patterns represent a historical participant risk profile; wherein the participant risk profile is compared with the historical participant risk profile and the current external market factor data.

11. The computer implemented method of claim 10 , wherein the participant risk profile is calculated as a function of a value at risk function that indicates a probability of occurrence of loss by a certain factor by a certain amount of time.

12. A computer implemented method comprising:
detecting a deviation from prior activity, inconsistent with current conditions, by a participant in a data transaction processing system in which data items are transacted by a hardware matching processor that matches electronic data transaction request messages for the same one of the data items based on multiple transaction parameters from different client computers over a data communication network by:
calculating, by a risk processor, a plurality of risk profiles for the participant for a plurality of time periods, the plurality of risk profiles based on a plurality of historical participant parameters;
calculating, by the risk processor, a plurality of external risk profiles for the plurality of time periods, the plurality of external risk profiles based on a plurality of historical external market parameters, wherein the plurality of historic external market parameters include data indicative of characteristics of previous transactions by other participants and characteristics of a financial derivative product that correspond to the plurality of historical participant parameters that occurs during the plurality of time periods, wherein the financial derivative product is traded on an exchange;
identifying, by a structured neural network a plurality of patterns between the plurality of risk profiles and the plurality of external risk profiles, the structured neural network comprising a layered plurality of interconnected processing nodes, wherein each connection of the plurality of interconnected processing nodes to another is dynamically weighted;
receiving, by the risk processor, electronic data transaction request messages comprising data for a new transaction;
calculating, by the risk processor, a current external factor risk profile as a function of current external market parameters;
generating, by the risk processor, a current risk profile for the participant comprising at least the data for the new transaction;
comparing, by the risk processor, the current risk profile, the current external factor risk profile, the plurality of risk profiles, and the plurality of external risk profiles using the plurality of patterns;
calculating, by the risk processor, an abnormality score based on the comparison; and
generating, by the risk processor, an alert when the abnormality score exceeds a threshold.

13. The computer implemented met

Assignees

• Chicago Mercantile Exchange Inc

Inventors

• Studnitzer Ari
• Geddes David
• Singh Inderdeep

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06Q40/04 Primary

  Trading; Exchange, e.g. stocks, commodities, derivatives or currency exchange · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/09

  Supervised learning · CPC title