EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
2838. Deep learning-based allocation for financial strategy replication in incomplete market.
Invited abstract in session TC-63: Advanced Options Strategies Using O.R. and Machine Learning, stream OR in Banking, Finance and Insurance: New Tools for Risk Management.
Tuesday, 12:30-14:00Room: S14 (building: 101)
Authors (first author is the speaker)
| 1. | Johan Macq
|
| PRISM Sorbonne, Université Paris 1 Panthéon-Sorbonne, Vivienne Investissement | |
| 2. | Yannick Malevergne
|
| PRISM Sorbonne, Université Paris 1 Panthéon-Sorbonne | |
| 3. | Marc Senneret
|
| Research Department, Vivienne Investissement | |
| 4. | Patrice Abry
|
| Signaux, Système & Physique, Ecole Normale Supérieure Lyon |
Abstract
We develop and implement a model-free, deep learning-based approach to tracking the performance of any chosen financial strategy in the presence of market frictions. Our method provides the optimal dynamic allocation over a - possibly time-varying - set of trading instruments. To achieve the optimal replication strategy, we introduce a novel architecture called Universal Allocators, inspired by Transformer networks. Its number of parameters is independent of the dimension of the investment universe and the length of the training path, allowing it to handle large portfolios without additional computational cost. We implement an adversarial training scheme using the Wasserstein-1 distance to solve the replication task. To the best of our knowledge, this is the first time that an adversarial training scheme using the Wasserstein-1 distance has been used to solve such a problem. To illustrate our method, we replicate a highly non-linear strategy, similar to a collar index, which provides the daily performance of an underlying asset whose values have been clipped below and above a given threshold. The replication set consists of a large number of vanilla options whose strike prices and expiration dates are defined over a discrete and sparse lattice, making the market incomplete. Using both simulated and S&P500 data, we show that our method provides financially meaningful and interpretable allocations for replicating such a strategy at moderate cost.
Keywords
- Finance and Banking
- Machine Learning
Status: accepted
Back to the list of papers