Mathematical Finance Seminar

Model Risk Management for Alpha Strategies created with Deep Learning

Speaker: Ben Steiner, Global Fixed Income, BNP Paribas Asset Management

Location: Warren Weaver Hall 1302

Date: Tuesday, November 12, 2019, 5:30 p.m.


Deep Learning has demonstrated spectacular success in domains outside finance and offers
tantalizing potential for developing trading strategies. This presentation reviews the basics
of Deep Learning and highlights when it should (or should not) be used.

Traditionally, Model Risk Management (MRM) consists of three elements:
(i) Conceptual Soundness – assessing the quality of the model design and construction;
(ii) Implementation Validation – confirming that the model is correctly implemented; &
(iii) Ongoing Monitoring - ensuring that the model is performing as intended.

Using deep learning to create trading strategies presents a number of challenges. Paramount
is the non-stationary nature of financial markets: out-of-sample data is most likely drawn
from a different distribution to training data. The key question is recalibration frequency:
recalibrating too fast results in fitting to noise, too slowly and a model is trained on stale
data. Either way, trading the sub-optimal strategy results in losses. A second challenge is
interpretation. Without knowing why a strategy is performing, limited information is
available for risk budgeting. The third challenge is ensuring deep learning is not simply an
expensive way of ‘rediscovering’ well-known risk factors.

In the presence of these three challenges, model risk management can be modified for
evaluating deep learning trading strategies. No simple test can discriminate between good
and bad strategies; rather a suite of analysis can be used to understand strategy behavior
and characteristics. Ongoing monitoring is then critical to understand when live trading is
not performing as intended. In this respect, the evaluation of deep learning strategies is
similar to the traditional evaluation of quant trading strategies. However, the increased ease
with which these strategies can now be created prompts even greater diligence in their
systematic evaluation and ongoing monitoring.

Bio – Ben Steiner
In his current role, Ben handles chief-of-staff and business management responsibilities
within the Global Fixed Income division.

Previous roles included: Head of Model Development team, Portfolio Manager & Quant
Researcher. This covered multiple asset classes ranging from the traditionally illiquid
(Private Debt and Real Estate) to more liquid markets (Non-traditional Bond; Managed
Futures; Global Macro and Equity Long/Short).

Prior to his current role, Ben was Head of Model Development at CIT where he managed
the team researching and implementing credit models. Earlier in his career, he was
Portfolio Manager and Senior Quant Researcher at BNP and, before that, Research
Manager at Aspect Capital in London. Ben started his career at Deutsche Bank in
quantitative research and portfolio construction.

He holds a BA in Economics from the University of Manchester and an MSc in
Mathematical Finance from Imperial College, London. In 2013, Ben was appointed to the
Board of Directors of the Society of Quantitative Analysts (SQA) and has given recent
guest lectures on machine learning and model risk management at Columbia & NYU.