Risk management with AAD and ML

[jbe456]

Description

Explore how Atoti can leverage risk management with AAD and ML

Resources

https://antoinesavine.com/2020/05/04/differential-machine-learning/

[jbe456]

A few extra comments from Georges/Anastasia:

The authors of the article discuss a way to improve the calibration quality of a neural network applied to pricing and valuation problems, by including not only the risk factors and expected result into the training set, but also including the derivatives of the expected result to risk factors.

• Why: resolve computational bottlenecks in trading and risk infrastructure
• What: use deep learning (neural networks) in pricing and valuation
• How: improve precision of the neural networks by including derivatives into the training set.

Because they have successfully implemented AAD at Danske and are proud of the fact that they can produce greeks as needed, at a low cost.

For instance, imagine we are trying to train an ML model to compute FRTB. We'd generate a set of portfolios and expected results to calibrate the model. Having trained a model, we can send any portfolio to the network - and it will compute FRTB. If we could also compute derivatives of FRTB to inputs - then the network will be able to extract factors more efficiently, and calibrate better.

Youtube on AAD:

• Executive summary: https://youtu.be/9CscP41z9ts
• Algorithm explained: https://youtu.be/twTIGuVhKbQ

Let's explore these two applications for atoti:

1. Machine-learning application for atoti: Demonstrate that a post processor can fetch ML model parameters from a TensorFlow calibrated model (for example, a linear classifier), and apply it to the data -> for further aggregation. For example, churn rate/customer classification for a telecommunication company. https://www.bedrockdata.com/blog/introduction-to-ai-machine-learning-part-iii We can subscribe and update model parameters in real-time mode! :) This is captured by https://github.com/atoti/notebooks/issues/35

2. Implement some sort of AAD as an algorithm: AAD can produce Greeks at lower cost and therefore open the possibility of

• generating more Greeks, either with finer granularity, second and cross derivatives, and how those derivatives can be made accessible to traders.
  • For IR we can look into multiple curve structures
  • For options, build multi-dimensional risk and stress P&L dashboard using more derivatives ( in particular second order and cross derivatives) rather than massive full reprising along multiple market data points.
• generating Greeks faster, therefore more intra-day incremental updates