GRNN with Small Samples

After a bank launches a new product or acquires a new portfolio, the risk modeling team would often be faced with a challenge of how to estimate the corresponding performance, e.g. risk or loss, with a limited number of data points conditional on business drivers or macro-economic indicators. For instance, it is required to project the 9-quarter loss in CCAR, regardless of the portfolio age. In such cases, the prevalent practice based upon conventional regression models might not be applicable given the requirement for a sufficient number of samples in order to draw the statistical inference. As a result, we would have to rely on the input of SME (Subject Matter Expert), to gauge the performance based on similar products and portfolios, or to fall back on simple statistical metrics such as Average or Median that can’t be intuitively related to predictors.

With the GRNN implemented in the YAGeR project (https://github.com/statcompute/yager), it is however technically feasible to project the expected performance conditional on predictors due to the fact that the projected Y_i of a future case is determined by the distance between the predictor vector X_i and each X vector in the training sample, subject to a smoothing parameter namely Sigma. While more samples in the training data are certainly helpful to estimate a generalizable model, a couple data points, e.g. even only one or two data points in the extreme case, are also conceptually sufficient to form a GRNN that is able to generate sensible projections without violating statistical assumptions.

Following are a couple practical considerations.

  1. Although normalizing the input data, e.g. X matrix, in a GRNN is usually necessary for the numerical reason, the exact scaling is not required. Practically, the “rough” scaling can be employed and ranges or variances used in the normalization can be based upon the historical data of X that might not be reflected in the training data with only a small sample size.
  2. With limited data points in the training data, the Sigma value can be chosen by the L-O-O (Leave-One-Out) or empirically based upon another GRNN with a similar data structure that might or might not be related to the training data. What’s more, it is easy enough to dynamically fine-tune or refresh the Sigma value with more data samples becoming available along the time.
  3. While there is no requirement for the variable selection in a GRNN, the model developer does have the flexibility of judgmentally choosing predictors based upon the prior information and eliminating variables not showing correct marginal effects in PDP (https://statcompute.wordpress.com/2019/10/19/partial-dependence-plot-pdp-of-grnn).

Below is an example of using 100 data points as the training sample to predict LGD within the unity interval of 1,000 cases with both GLM and GRNN. Out of 100 trials, while the GLM only outperformed the simple average for 32 times, the GRNN was able to do better for 76 times.