Online and offline hybrid risk control modeling training and production publishing method and system

Abstract

The invention provides an online and offline hybrid risk control modeling training and production publishing method and system. The online and offline hybrid risk control modeling training and production publishing system comprises a decision service module, a score calculation module and a feature calculation module, The decision service module is used for receiving a calculation request sent byan external system and providing a unified calling interface for interaction with the external system for the scoring calculation module and the feature calculation module; The feature calculation module is used for feature management, configuration of feature calculation logic, a calculation process and a data source; And the scoring calculation module is used for configuring a calculation modeland characteristic parameters, and carrying out scoring and testing on request parameters in the calculation request input by the external system. According to the invention, it can be ensured that the offline data environment of model training is consistent with the online data environment, and the offline implementation code of the model is consistent with the online production environment code,thereby ensuring that the scoring model falls to the production environment without distortion.

Description

technical field [0001] The invention relates to the technical field of financial product models, in particular to a method and system for online and offline mixed risk control modeling training and production release. Background technique [0002] The implementation process of the current common scoring models can be roughly divided into two steps: offline modeling training and online production and release. Step 1: Data modelers use a variety of data warehouses (offline non-real-time data, follow the standard data warehouse architecture, take a relatively common situation as an example: small-scale structured data is stored in MYSQL, large-scale semi-structured data Structured and unstructured data are stored in HIVE) as the data source, based on their own experience and data mining algorithms, using modeling tools (R, SAS, PYTHON and some interface modeling tools, depending on personal habits), to train Scores with good data performance at all stages of the financial prod...

AI Technical Summary

Problems solved by technology

[0003] The above model landing process has the following disadvantages: 1. The modeling source data comes from the source layer table in the data warehouse (MYSQL\HIVE) and the table processed by the middle layer. These data are structurally different from the production environment (splitting tables, Data warehouse analysis lost number,

2. Due to the inconsistency of the data environment, the needs of model personnel need to be transformed into online and offline data difference correspondence, online and offline logic correspondence, which consumes a lot of communication costs and is extremely error-prone

3. Before the model goes online, it cannot be backtested with real data, so it has to be tested online

Repeatedly, it sometimes takes several months until the model is actually tested and accepted

At this time, the model may have become invalid and it is difficult to cope with the ever-changing business changes

4. The scoring model is an important asset of a financial company, and its confidentiality is particularly important. This process inevitably increases the risk of contacting too many people and data leakage

Images