Approach

It’s an App

goMortgage is an app. You create a text file (*.gom) that instructs goMortgage what to do. Want to change the model structure? Change the source data? Change the features in the model? Change the sampling scheme? These are easy changes to the .gom file. There’s no need to hack into your code, create a bunch of new tests and hope that the changes won’t cause some unintended consequences.

The output of a goMortgage run is placed in a directory you specify. The directory includes the .gom file, a log file, the model, a summary from the data build, the cost curves from the model build, and all graphical assessments of the model. Other outputs are saved to ClickHouse tables.

One-stop Shopping

goMortgage handles the entire model building process. It will

• build the dataset
• build the model
• assess the model

Data

goMortgage employs three sets of data. Ideally they are, but do not have to be, disjoint. The three sets are:

• model. The data used to fit the coefficients of the model.
• validation. The data used to determine early stopping during the fit process.
• assessment. The data used to build graphical and numeric assessments of the model.

goMortgage will build these tables for you. It requires two source tables to do this: (1) loan data and (2) non-loan data. The non-loan data includes fields such as HPI, interest rates, and unemployment rates.

The structure of the data determines the type of model that is built. goMortgage is very flexible in the construction of the data. Hence, many types of models can be built.

See the examples page for some types that are possible.

Model

goMortgage fits neural net models. The target field may be either discrete or continuous.Two types of layers are supported: fully connected and dropout. Four activation functions are supported: relu, leaky relu, linear and softmax. There are additional options, such as L2 regularization and linearly decreasing learning rates.

Assessment

goMortgage produces a variety of model assessments, mostly graphical, to provide insight into the model performance and relationships of inputs to output.

Given the approach of goMortgage to model building, one can do something that’s very cool. It is straightforward to build the model on data up to time t and assess it on data after time t. In fact, one can march through time conducting this process to gain insight into how the model performs out of time period in a variety of economic environments. One can even perform experiments like restricting the assessment to loans, say, in California.

Extensible

goMortgage comes configured to work with Fannie and Freddie data but can be modified for other data sources. See the Bring Your Own Data guide for details on how to do this. In fact, it would not be difficult to apply goMortgage to other asset classes.