Wells Fargo Bank, N.a.

Bank of America Corporation

Harsh Singhal

Agus Sudjianto

Wayne B Shoumaker

Timothy John Breault

Jie Chen

William Anthony Nobili

Sandi Setiawan

Fernando Cela Diaz

Kaloyan Mihaylov

Naveen Gururaja Yeri

Vijayan Narayana Nair

Arun Pinto

Tarun Joshi

Aijun Zhang

Benhong Zhang

Zhe Zhang

Hungien Wang

Michael Chorba

Daniel Hudson

Mina Naghshnejad

Jocelyn Sikora

Angelina Yang

Kiran Vuppu

Ye Yu

Suhas Sreehari

Ningzhou Zeng

Ridhi Puppala

Harsh Vardhan Tiwari

Karan H Jeswani

Andrea Renee Turner

Bradford Timothy Winkelman

Magali F Van Belle

Wei Wei

Kai Arne Krueger

Systems, apparatuses, methods, and computer program products are disclosed for distillation of a natural language processing model. An example method includes receiving, by communications circuitry, a set of text data comprising a set of observations and predicting, by processing circuitry and using the NLP model, classifications for each observation in the text data. The example method further includes generating, by model training engine, a balanced sampled data structure based on the predicted classifications for each observation in the text data and training, by the model training engine, a surrogate model using the balanced sampled data structure. The example method further includes identifying, by an interpreter and from the surrogate model, a set of most-influential tokens in the text data.

Systems and methods for flexible regularized distillation of natural language processing models to facilitate interpretation

Disclosed is an example approach in which network and non-network features are used to train a predictive machine learning model that is implemented to predict financial crime and fraud. Graphical network features may be generated by applying financial entity risk vectors to a network model with representations of various types of networks. The network model may comprise transactional, non-social, and/or social networks, with edges corresponding to linkages that may be weighted according to various characteristics (such as frequency, amount, type, recency, etc.). The graphical network features may be fed to the predictive model to generate a likelihood and/or prediction with respect to a financial crime. A perceptible alert is generated on one or more computing devices if a financial crime is predicted or deemed sufficiently likely. The alert may identify a subset of the set of financial entities involved in the financial crime and present graphical representations of networks and linkages.

Network based features for financial crime detection

This disclosure describes techniques that include validation or other assessments of digital systems, such as machine learning models and other statistical models. In one example, this disclosure describes a method that includes receiving, by a validation computing system and from a development system, a request to perform a test on a model configured to execute on the development system; outputting, by the validation computing system to the development system and in response to the request, an instruction; enabling the development system to process the instruction; receiving, by the validation computing system, test response data; evaluating, by the validation system, the test response data.

Model validation as a service

Systems, apparatuses, methods, and computer program products are disclosed for screening data instances based on a target text of a target corpus. A screening device analyzes a target corpus to generate at least two term dictionaries for the target corpus. The screening apparatus, based on a frequency of a term in the target corpus, determines a term weight for the term; for each data instance, determines term scores for the data instance and the target text based on the term weights; filters the data instances based on the term scores, to generate a short list of data instances; determines term similarity scores between each data instance of the short list and target text based on the term weights; and provides a data instance determined to likely correspond to the target text and an indication of the corresponding term similarity score(s). A term is a word or an n-gram.

Systems and methods for screening data instances based on a target text of a target corpus

Methods, computer readable media, and apparatuses for evaluating models using forecast error attribution are presented. According to one or more aspects, one or more input values corresponding to one or more input variables may be forecast. One or more results of a modeling function may be calculated using the one or more forecasted input values. Thereafter, actual performance data corresponding to the modeling function may be received. One or more holdout values for the modeling function may be calculated using the actual performance data. Subsequently, a graph that includes the one or more results of the modeling function, the actual performance data, and the one or more holdout values for the modeling function may be plotted. In some arrangements, the one or more holdout values for the modeling function may be indicative of one or more assumption errors made with respect to the one or more forecasted input values.

Evaluating models using forecast error attribution

A data driven and forward looking risk and reward appetite methodology for consumer and small business is described. The methodology includes account level historical data collection for customers associated with accounts as part of a portfolio. The account level historical data is segmented into groups of customers with similar revenues and loss characteristics. Segmented data is decomposed into seasoning, vintage, and cycle effects. Statistical clusters are formed based upon the data and effects. A simulation is applied to the statistical clusters and prediction data is generated. A simulation strategy to forecast and simulate revenue and loss volatility is developed. Efficient frontier curves of risk (e.g., return volatility) and reward (e.g., expected return) are created for the current portfolio under various economic scenarios.

Risk and reward assessment mechanism

Embodiments of the present invention relate to methods and apparatuses for determining leading indicators and/or for modeling one or more time series. For example, in some embodiments, a method is provided that includes: (a) receiving first data indicating the value of a total income amount for a plurality of consumers over a period of time; (b) receiving second data indicating the value of a total debt amount for a plurality of consumers over a period of time; (c) selecting a consumer leverage time series that compares the total income amount to the total debt amount over a period of time; (d) modeling the consumer leverage time series based at least partially on the first and second data; (e) determining, using a processor, the value of the cycle component for a particular time; and (f) outputting an indication of the value of the cycle component for the particular time.

Determining leading indicators

A data driven and forward looking risk and reward appetite methodology for consumer and small business is described. The methodology includes customer segmentation to create pools of homogeneous assets in terms of revenue and loss characteristics, forward looking simulation to forecast expected values and volatilities of revenue and loss, and risk and reward optimization of the portfolio. One methodology used for modeling revenue and loss is a generalized additive effect decomposition model to fit historical data. Based on the model, a segmentation procedure is performed, which allows for creation of groups of customers with similar revenue and loss characteristics. An estimation procedure for the model is developed and a simulation strategy to forecast and simulate revenue and loss volatility is developed. Efficient frontier curves of risk (e.g., return volatility) and reward (e.g., expected return) are created for the current portfolio under various economic scenarios.

Growing community of inventors

Charlotte, NC, United States of America

Harsh Singhal