High-Performance Machine Learning for FinTech
Abstract: The use of machine learning techniques in the FinTech sector, such as in portfolio selection, requires
a high-performance compute engine to test a large number of candidate strategies. The higher the
runtime performance of the engine, the larger the number of strategies that can be considered and
the wider the scope of testing that could be conducted. This leads to better quality strategies. Thus,
runtime performance should directly impact strategy performance in the market.
This paper presents a compute engine for processing market data and implementing highly customis-
able trading strategies. We use it as the core of our differential-evolution-based machine learning
framework for portfolio selection. We then discuss a range of key techniques that improve its run-
time performance and show that it outperforms other extant engines.
Submission Length: Long submission (more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=sdIJYBULE3
Changes Since Last Submission: Comment from previous desk rejection:
Font appears modified from template default, please reviist and resubmit.
************
Solution:
Removed \usepackage{mathptmx} which seems to load Times New Roman font as a side effect. Font should now be correct and looks like the template PDF.
************
Revision on 04.02.2024:
2 Context and Background: (reference AutoML and talk about generalisability)
Kotthoff et al. (2017) with its AutoML capabilities is an inspiration to our research and our long-term goal is to
develop something similar for timeseries analysis in the financial trading domain. The concepts we develop
on the way and the final solution should be generalisable for other research domains, even though we initially
focus on one specific domain during the development.
3.2 Expression Language: (explain why this particular expression language is used)
The platform uses a domain-specific expression language to formalise building blocks for strategy gene-
ration (indicators and signals) and to automate portfolio management and other high-level decisions.
<anonymised> (2022b) discusses the design and choice to use this expression language in detail. In sum-
mary it is only possible to achieve a high performance by tightly integrating the expression language into
the surrounding context (in our case a trading platform). We also compare our expression language against
alternatives in the Java Virtual Machine in that publication. Our implementationn is significantly faster
while also being more feature-rich. Other contender expression languages are unsuitable because they are
prioprietary or written in a different programming language, which makes integration impossible or too slow
for high performance. Table 14 extends this analysis by comparing our integrated solution against propri-
etary solutions that also use expression languages. We hope that other and future expression languages can
be improved by applying our generalised concepts to make expression languages more efficient. For this
purpose we have open sourced our expression language implementation.
************
Revision on 09.02.2024: see commented answer to the second review by "9qwm" from same date.
Also modified the conclusion to summarize our recent advancements with scalability and alternative memory allocation techniques.
************
Revision on 09.02.2024: we have attempted to incorporate all comments by substantially rewriting, rephrasing, rationalising, and editing the manuscript. In particular, the introduction and motivation (Sections 1 and 2) now highlight the contribution of this paper more clearly.
Assigned Action Editor: ~Naigang_Wang1
Submission Number: 1987
Loading