Taran Decision Manager

One - click deployment of any model

Standard decision managers

Usually only logistic regression models are supported
Predictor binning or transformations need to be manually coded into the decision manager
Deployment of other techniques (XGBoost, Neural Networks) is very difficult of impossible, external modules are often needed

TaranDM

Direct support for one-click deployment of any predictive model (logistic regression, gradient boosting, neural networks)
Model specifications, predictor bins, and transformations are saved during the modelling phase to a deployment file and are automatically processed by TaranDM

Flexible & and modular architecture

Standard decision managers

Usually need to receive lots of data from core banking systems Common issues: IT says that it…
is hard to add new data aggregations
will slow down the other system
cannot be on real-time basis

TaranDM

Works with a minimal set of input data, which is client identification and application form data

All other data, such as antifraud predictors, behavioural preditors, and blacklists are calculated and stored in separate modules
Each module can use the technology that is most appropriate for its task

Unified and automated testing framework

Standard decision managers

Frequent IT releases influence the data integrations and input data calculations used by decision manager system
Due to mismatched testing and production environments, bugs and issues are often found in production only once they influence performance metrics (approval rate, default rate, ...)

TaranDM

Any change in the decision strategy, data integration or predictor calculation is automatically tested:

Assessment and reporting of impact on production data
Integration tests for all data sources
Unit tests for predictors, core system, hard checks, scorecards, decision tables, expected results, etc.

Single environment (analytics & production)

Standard decision managers

Analytical tasks are performed on data marts in the data warehouse
Usually only a subset of all predictors that are used in analytics is available in the production layer
No access to the raw data that serves as the base for building data marts

TaranDM

Is capable of working with large input objects like XML or JSON files
Contains one unified predictor catalogue that serves both the analytical layer ('calculate all') and the production layer 'calculate only those needed by the current strategy')
All raw data are stored and any new aggregation can be coded, tested, and deployed

Instant deployment of new predictors

Standard decision managers

When a new predictor is identified by data scientists:

A new IT request is needed to get it to production
Implementation can take months
There is often a mismatch between implementation and specifications

TaranDM

All predictors are Python-based and can be developed by analysts or data scientists
An automated testing framework to validate new predictors
Little or no dependence on IT

Smart data source caller

Standard decision managers

No data source caller included
IT has to integrate every data source, implement caching logic, and perform upgrades to any new versions

TaranDM

Contains a data source caller where:

Data are collected in multiple iterations and cached to save costs
Templates allow for rapid implementation of new data sources, and the OpenAPI format is used to define protocols
An automated solution upgrades stored data in older formats to the latest one

Automated reporting layer

Standard decision managers

When a new model is deployed it has to be added to the reporting framework manually
Various A/B tests of hard checks, predictive models, and decision
rules need to be evaluated

TaranDM

TaranDM includes a reporting module

Detailed data about every iteration of the decision process are generated and used as input for reporting
Files with models contain all information for reporting, allowing reports to be updated automatically
Reporting includes evaluation of A/B tests

Open code, no vendor lock-in

Standard decision managers

Making changes, for example adding new functionality to the decision manager:

has to be done by the vendor
is often impossible
if it is possible, is protracted and costly

TaranDM

Built on a carefully selected technology stack with its core and main modules in Python
Emphasizes optimum performance, open source, and scalability
A licensing option allows modification of the TaranDM core or modules

Workflow management included

Standard decision managers

Are usually called from core IT systems in specific phases of the approval process
Need an orchestrating system that gathers all data for the decision-making system
Any change in the workflow is complicated and takes months to implement

TaranDM

Includes persistent storage of all data used in the decision-making process
No need for an orchestrating system, any additional data can be sent to the system alone
Workflow is managed by TaranDM; there is no limit on the number of iterations or need to pre-define phases

No code duplication, no copying and pasting

Standard decision managers

Implementation of a Champion/Challenger test requires copying and modifying the entire strategy or a relevant part thereof
Subsequent changes to the Champion strategy are not passed on to the Challenger strategy

TaranDM

Uses a template system to compile strategies:

Base templates for various scenarios are included
In Challenger, you can change only the parts that you want to change
Parts shared by both Champion and Challenger remain consistent

TaranDM Components

Core Scoring

Strategies are easy to configure and saved in human-readable JSON files. Default strategy templates for various business problems are pre-configured and re-usable. Their adjustment and customization by business owners is user-friendly with effortless maintenance.

Data Sources Adapter

The Data Source Adapter provides a platform for two-way communication between the Core Scoring component and any external/internal data source. it uses an OpenAPI standard interface specification and can thus handle structured and unstructured data types. The Data Source Adapter has cache functionality that enhances speed and reduces costs (for example avoiding duplicate requests to credit bureaus).

AntiFraud

AntiFraud is a separate module that has been created to mitigate the risk of fraud in real-time on-line processes. A credit applicant's personal data, device information, fingerprint, and IP address are collected and multiple concentration and cross-checks are run to prevent fradulent applications, which often come as a series of requests from a single IP address or device. As AntiFraud is a stand-alone module, concentration and cross-checks can be built on a technology stack that is suitable for performing real-time aggregation tasks.

Data Lake with Analytical Layer

All inputs, outputs, and other unstructured and structured data generated during a TaranDM run are stored in a highly scalable Hadoop Data Lake. The Analytical Spark layer, which is a part of this component, enables data scientists to work in the familiar environment of Jupyter notebooks and develop new models using raw data directly from the Data Lake (as well as other predictors and components of TaranDM). Such models can easily be propagated to the Core Scoring component, including their predictors, and used in production.

Reporting

The TaranDM Reporting component uses Kafka to gather structured and unstructured data from various data sources such as a Data Lake, scoring runs, decisioning, external data sources, etc. The data... Another part of the Reporting component are predefined BI tool templates (e.g. Tableau) that are used to visualize stored data. Reporting allows users to access information about decision process outcomes, KPIs, strategy comparisons, model stability and performance, benefits of external data sources, etc.

TaranDM White paper (pdf)

Taran Decision
Manager

The ultimate competitive advantage for your business.

Key advantages:

One - click deployment of any model

Flexible & and modular architecture

Unified and automated testing framework

Single environment (analytics & production)

Instant deployment of new predictors

Smart data source caller

Automated reporting layer

Open code, no vendor lock-in

Workflow management included

No code duplication, no copying and pasting