For new customers (retail/wholesale), traditional models have low accuracy. AI counters by scoring subjective dimensionsthrough NLP or image models. Mapping text/image vectors to intuitive latent factors, enables business to take action #deeplearning#creditrisk#decisionmodel
With advances in NLP, we use deep learning in our framework to
1. capture intangibe aspects of borrower
2. detect and predict fraud in loan application
psychometric information when available gives much superior performance
#creditrisk#NLP#deeplearning#retail#pdmodel#AI
LIME and SHAP two common model-agnostic interpretability metrics - can be very wrong. Perturbations generated can have large deviation from actual data distribution and can be shown by PCA on actual data vs perturbations. #explainableAI#interpretability#blackbox#deeplearning
Regulators strive for model interpretability, for transparency in the lending market, modellers are more focussed on accuracy as provided by more complicated models. Interpretability of "Black-Box" deep learning models can bridge this gap. #deeplearning#explainableAI#creditrisk
Text data is ubiquitous now and the text data can act as an early indicator to defaults. We can use text data from customer calls, social profiles (public data) and news to enhance the models and create a early monitor for such events #creditrisk#creditmonitor#risk#ML#AI#DL
Credit risk models, built periodically to learn the changing market dynamics, gives fixed threshold for acceptance criterion of decision models. Reinforcement learning adapts the cut off dynamically to live data. #creditrisk#reinforcementlearning#machinelearning#deeplearning