SOLUTIONS: Case Study
Predict Originations with Confidence
We used the LSTM (Long-Short Term Memory) neural network model to forecast future originations - the model uses past observations to predict future ones. LSTMs are very good at handling data with long term dependencies. Its predictions can easily be updated to incorporate new data as it comes in.
With this deep learning model - the plot shows the observed number of loan originations (per day) from Q1 2016 to Q2 2019 in blue. We use three variables to predict the number of loan originations in Q3 2019 - the 10 year treasury yield, the Chicago Federal Reserve Financial Conditions Index, and the (seasonally adjusted) quarter-over-quarter rate of change of GDP.
The LSTM's predictions are in red. They continue to remain elevated compared to the values prior to 2019. This matches what we would expect - the 10 year yield has fallen dramatically in 2018, and which led to more refinancing that is driving the increased loan activity. The 10 year yield has continued to fall in Q2 2019, so we expect refinancing activity to continue to increase when we update these predictions next quarter.