US-12626251-B2 - Systems and methods for predictive modelling of clearing messages

Abstract

A modelling platform including at least one processor in communication with a memory device and a payment processor is provided. The at least one processor is configured to retrieve subsets of data from a transaction history database, derive training data sets from the subsets, apply model input data fields of each training data set as inputs to one or more machine learning models, and apply a machine learning algorithm to adjust parameters of the one or more machine learning models. The at least one processor is also configured to upload at least one trained machine learning model to an operational predictive model module, apply a stream of real-time authorization request messages as inputs to the at least one trained machine learning model, and transmit, to the payment processor in real-time or near real-time, values of at least one output obtained by applying the stream.

Inventors

• Shanmuga Vidya Avudaiappan
• Matt Froidl
• Tracy McLaughlin
• Stephanie Detchemendy
• Dennis Hill
• Mariya Ivanyshyn Spaeth

Assignees

• MASTERCARD INTERNATIONAL INCORPORATED

Dates

Publication Date

20260512

Application Date

20230711

Claims (20)

1. 1 . A modelling platform comprising: at least one processor in communication with a memory device and a payment processor, wherein the at least one processor is configured to: retrieve subsets of data from a transaction history database, wherein the transaction history database stores transaction data extracted from authorization messages and clearing data extracted from clearing messages for a plurality of authorized transactions processed by the payment processor, derive training data sets from the retrieved subsets of data, wherein each training data set corresponds to one of the plurality of authorized transactions and includes: (i) the transaction data associated with one or more authorization request messages of the one of the plurality of authorized transactions, and (ii) the clearing data associated with one or more clearing messages associated with the one of the plurality of authorized transactions; train one or more machine learning models by applying the transaction data and the clearing data of at least a portion of the training data sets as inputs to the one or more machine learning models to predict one of: a clearing message latency for a subsequent transaction processed by the payment processor, or a clearing message amount for a subsequent transaction processed by the payment processor; return, from the one or more trained machine learning models and for each training data set, a first output prediction for clearing message latency or a second output prediction for the clearing message amount; derive a correction signal from: (i) a first result comparing the first output prediction to an actual clearing message latency, or (ii) a second result comparing the second output prediction to an actual clearing message amount; periodically re-train the one or more trained machine learning models using the correction signal; periodically re-upload the one or more re-trained machine learning models to further compare output predictions from the one or more re-trained machine learning models to at least one of actual clearing message latencies or actual clearing message amounts; adjust parameters of the one or more trained machine learning models by applying a machine learning algorithm to the one or more re-trained machine learning models, wherein the parameters are adjusted until an error between the first or second output prediction and at least one actual result falls below a threshold indicating that the first or second output prediction is predictive of the at least one actual result, the at least one actual result representing one of an actual clearing message latency or an actual clearing message amount; in response to the error falling below the threshold, upload at least one re-trained machine learning model of the one or more re-trained machine learning models to an operational predictive model module; apply a data stream associated with subsequent authorization request messages received from the payment processor as inputs to the at least one re-trained machine learning model; return, from the at least one re-trained machine learning model, at least one output obtained in response to applying the data stream to the at least one re-trained machine learning model, wherein the at least one output predicts values corresponding to at least a predicted time of issuance of a clearing message for each of the subsequent authorization request messages; and transmit, to the payment processor in real-time or near real-time, the returned at least one output, thereby providing an accurate prediction of at least one of: (i) an amount to be reserved for clearing purposes for each of the subsequent authorization request messages, or (ii) a period of time to reserve a corresponding amount before release for each of the subsequent authorization request messages.
2. 2 . The modelling platform of claim 1 , wherein the at least one processor is further configured to derive model input data fields from data fields in one of the subsets of data corresponding to historical authorization request messages processed by the payment processor.
3. 3 . The modelling platform of claim 2 , wherein the model input data fields include one or more of: an authorization date, a clearing date, a merchant identifier, a merchant category code, an account number entry code, a pre-authorization code, an acquirer code, or a country code.
4. 4 . The modelling platform of claim 1 , wherein the at least one actual result includes one or more delay results that represent an actual delay between authorization and receipt by the payment processor of a corresponding clearing message.
5. 5 . The modelling platform of claim 1 , wherein the at least one actual result includes one or more amount results that represent a change in a transaction amount between an initial authorized transaction amount and a clearing transaction amount in a corresponding clearing message.
6. 6 . The modelling platform of claim 1 , wherein the one or more machine learning models include a neural network.
7. 7 . The modelling platform of claim 6 , wherein the neural network comprises one or more layers of nodes, and wherein the parameters are adjusted based on respective weight values applied to one or more inputs to each of the nodes.
8. 8 . A method implemented by a modelling platform including: the modelling platform including at least one processor in communication with a memory device and a payment processor, the method comprising steps, performed by the at least one processor, of: retrieving subsets of data from a transaction history database, wherein the transaction history database stores transaction data extracted from authorization messages and clearing data extracted from clearing messages for a plurality of authorized transactions processed by the payment processor; deriving training data sets from the retrieved subsets of data, wherein each training data set corresponds to one of the plurality of authorized transactions and includes: (i) the transaction data associated with one or more authorization request messages of the one of the plurality of authorized transactions, and (ii) the clearing data associated with one or more clearing messages associated with the one of the plurality of authorized transactions; training one or more machine learning models by applying the transaction data and the clearing data of at least a portion of the training data sets as inputs to the one or more machine learning models to predict one of: a clearing message latency for a subsequent transaction processed by the payment processor, or a clearing message amount for a subsequent transaction processed by the payment processor; returning, from the one or more trained machine learning models and for each training data set, a first output prediction for at least one of the clearing message latency or a second output prediction for the clearing message amount; deriving a correction signal from: (i) a first result comparing the first output prediction to an actual clearing message latency, or (ii) a second result comparing the second output prediction to an actual clearing message amount; periodically re-training the one or more trained machine learning models to adjust using the correction signal; periodically re-upload the one or more re-trained machine learning models to further compare output predictions from the one or more re-trained machine learning models to at least one of actual clearing message latencies or actual clearing message amounts; adjusting parameters of the one or more re-trained machine learning models by applying a machine learning algorithm to the one or more re-trained machine learning models, wherein the parameters are adjusted until an error between the first or second output prediction and at least one actual result falls below a threshold indicating that the first or second output prediction is predictive of the at least one actual result, the at least one actual result representing one of an actual clearing message latency or an actual clearing message amount; in response to the error falling below the threshold, uploading at least one re-trained machine learning model of the one or more re-trained machine learning models to an operational predictive model module; applying a data stream associated with subsequent authorization request messages received from the payment processor as inputs to the at least one re-trained machine learning model; returning, from the at least one re-trained machine learning model, at least one output obtained in response to applying the data stream to the at least one re-trained machine learning model, wherein the at least one output predicts values corresponding to at least a predicted time of issuance of a clearing message for each of the subsequent authorization request messages; and transmitting, to the payment processor in real-time or near real-time, the returned at least one output, thereby providing an accurate prediction of at least one of: (i) an amount to be reserved for clearing purposes for each of the subsequent authorization request messages, or (ii) a period of time to reserve a corresponding amount before release for each of the subsequent authorization request messages.
9. 9 . The method of claim 8 , further comprising: deriving model input data fields from data fields in one of the subsets of data corresponding to historical authorization request messages processed by the payment processor.
10. 10 . The method of claim 9 , wherein the model input data fields include one or more of: an authorization date, a clearing date, a merchant identifier, a merchant category code, an account number entry code, a pre-authorization code, an acquirer code, or a country code.
11. 11 . The method of claim 8 , wherein the at least one actual result includes one or more delay results that represent an actual delay between authorization and receipt by the payment processor of a corresponding clearing message.
12. 12 . The method of claim 8 , wherein the at least one actual result includes one or more amount results that represent a change in a transaction amount between an initial authorized transaction amount and a clearing transaction amount in a corresponding clearing message.
13. 13 . The method of claim 8 , wherein the one or more machine learning models include a neural network.
14. 14 . The method of claim 13 , wherein the neural network comprises one or more layers of nodes, and wherein the parameters are adjusted based on respective weight values applied to one or more inputs to each of the nodes.
15. 15 . At least one non-transitory computer-readable medium that includes computer-executable instructions embodied thereon that when the computer-executable instructions are executed by at least one processor of a modelling platform in communication with a memory device and a payment processor, the computer-executable instructions cause the at least one processor to: retrieve subsets of data from a transaction history database, wherein the transaction history database stores transaction data extracted from authorization messages and clearing data extracted from clearing messages for a plurality of authorized transactions processed by the payment processor; derive training data sets from the retrieved subsets of data, wherein each training data set corresponds to one of the plurality of authorized transactions and includes: (i) the transaction data associated with one or more authorization request messages of the one of the plurality of authorized transactions, and (ii) the clearing data associated with one or more clearing messages associated with the one of the plurality of authorized transactions; train one or more machine learning models by applying the transaction data and the clearing data of at least a portion of the training data sets as inputs to the one or more machine learning models to predict one of: a clearing message latency for a subsequent transaction processed by the payment processor, or a clearing message amount for a subsequent transaction processed by the payment processor; return, from the one or more trained machine learning models and for each training data set, a first output prediction for at least one of the clearing message latency or a second output prediction for the clearing message amount; derive a correction signal from: (i) a first result comparing the first output prediction to an actual clearing message latency, or (ii) a second result comparing the second output prediction to an actual clearing message amount; periodically re-train the one or more trained machine learning models using the correction signal; periodically re-upload the one or more re-trained machine learning models to further compare output predictions from the one or more re-trained machine learning models to at least one of actual clearing message latencies or actual clearing message amounts; adjust parameters of the one or more re-trained machine learning models by applying a machine learning algorithm to the one or more re-trained machine learning models, wherein the parameters are adjusted until an error between the first or second output prediction and at least one actual result falls below a threshold indicating that the first or second output prediction is predictive of the at least one actual result, the at least one actual result representing one of an actual clearing message latency or an actual clearing message amount; in response to the error falling below the threshold, upload at least one re-trained machine learning model of the one or more re-trained machine learning models to an operational predictive model module; apply a data stream associated with subsequent authorization request messages received from the payment processor as inputs to the at least one re-trained machine learning model; return, from the at least one re-trained machine learning model, at least one output obtained in response to applying the data stream to the at least one re-trained machine learning model, wherein the at least one output predicts values corresponding to at least a predicted time of issuance of a clearing message for each of the subsequent authorization request messages; and transmit, to the payment processor in real-time or near real-time, the returned at least one output, thereby providing an accurate prediction of at least one of: (i) an amount to be reserved for clearing purposes for each of the subsequent authorization request messages, or (ii) a period of time to reserve the corresponding amount before release for each of the subsequent authorization request messages.
16. 16 . The non-transitory computer-readable medium of claim 15 , wherein the computer-executable instructions further cause the at least one processor to derive model input data fields from data fields in one of the subsets of data corresponding to historical authorization request messages processed by the payment processor.
17. 17 . The non-transitory computer-readable medium of claim 16 , wherein the model input data fields include one or more of: an authorization date, a clearing date, a merchant identifier, a merchant category code, an account number entry code, a pre-authorization code, an acquirer code, or a country code.
18. 18 . The non-transitory computer-readable medium of claim 15 , wherein the at least one actual result includes one or more delay results that represent an actual delay between authorization and receipt by the payment processor of a corresponding clearing message.
19. 19 . The non-transitory computer-readable medium of claim 15 , wherein the at least one actual result includes one or more amount results that represent a change in a transaction amount between an initial authorized transaction amount and a clearing transaction amount in a corresponding clearing message.
20. 20 . The non-transitory computer-readable medium of claim 15 , wherein the one or more machine learning models include a neural network.

Description

BACKGROUND This disclosure relates generally to electronic payment networks, and more specifically to systems and methods for developing, training, and applying machine learning models to predict parameters of a clearing message corresponding to a particular authorization message. Electronic payment networks are in widespread use to communicate information needed to process transactions between a payment account holder, an issuer of the payment account (e.g., a bank that issued a payment account to the account holder) a merchant, and an acquirer (e.g., a bank that provides a merchant account for use on the payment network). The transaction may involve presentment at a point-of-sale terminal of a physical payment card (e.g., a credit card, debit card, or prepaid card) linked to the payment account, use of a device that includes payment account information and digital payment capability (e.g., a smart phone device including a digital wallet that holds payment account information), manually entered payment account information via another device such as a computer device interfacing with a merchant's online platform, and payment account information stored by a merchant platform and used to initiate recurring payments previously consented to by the account holder. Sophisticated multi-party electronic payment networks are known to process payment account transactions, confirm authorized charges, manage payments and transfer of funds (settlement), confirm payment status, and compute available balances. Account holders may easily retrieve information online to check their pending and historical account charges and available balances whenever desired. In many electronic payment networks, account issuers receive authorization request messages in real time, typically from a merchant, as transactions takes place. Each authorization request message identifies a payment account and a tentative transaction amount for transfer to the merchant from the payment account. If the transaction is authorized, the account issuer places a hold on funds in the payment account corresponding to the tentative transaction amount. However, the actual amount for transfer to the merchant from the payment account is not established, and settlement of funds is not initiated, until the account issuer receives a follow-up clearing message via the payment network. Until the clearing message is received, the account issuer's records reflect only a pending charge for the tentative amount. Yet, in some cases it can take anywhere from hours to days to several weeks after the authorization for the account issuer to receive a clearing message. Non-limiting examples include payment card authorizations used to reserve a hotel stay, where no clearing message is sent until a final cost is established upon completion of the stay; online orders for future delivery in which no clearing message is sent until the order fulfillment date; and instances in which a clearing message is simply delayed due to erroneous data input or software events occurring at the merchant or acquirer. Moreover, in some cases the actual amount of the transaction differs from the tentative amount cited during authorization. For example, the cost of ordered goods may change before the fulfillment date, or the incidentals purchased during a hotel stay may increase the actual billed amount at clearing. Clearing message latency can cause account holders to be confused by seeing incorrect amounts when they check their balances with the account issuer. In addition, account issuers are deprived of liquidity as they hold funds in reserve for satisfaction of an expected clearing message. An inability to predict when a clearing message will arrive for such authorized transactions limits an ability of account issuers to plan and adjust their operations on a daily basis. For example, but not by way of limitation, the information provided in authorization request messages to issuers of debit cards and issuers of prepaid cards may be too limited to enable the issuer to draw any conclusions about clearing message timing. In addition, in some circumstances a clearing message is never communicated for certain authorized transactions. Non-limiting examples include orders which are cancelled by the account holder or merchant after the initial authorization, but for which notice of the cancellation is not submitted properly to the electronic payment network. Moreover, in some instances a clearing message is simply never sent due to erroneous data input or software events occurring at the merchant or acquirer. It is difficult for account issuers to determine whether or when pending charges/held funds from a previously authorized transaction should be released on the expectation that no clearing message is forthcoming. BRIEF DESCRIPTION In one aspect, a modelling platform including at least one processor in communication with a memory device and a payment processor is provided. The at least one processor