Presented at QuantCon Singapore 2016, Quantopian's quantitative finance and algorithmic trading conference, November 11th.
Deep learning is a subset of machine learning that draws on fields including applied mathematics, statistics, computer science and neuroscience. Currently it is experiencing tremendous growth due to the confluence of larger datasets, massive computational power and the development of new algorithms. While there is a lot of work on static data and some work on sequential data (such as text-based learning), less attention has been paid to (dynamic) time series data. In finance we are often interested in problems of prediction and classification, based on time series data.
In this talk, we introduce deep learning models and discuss their application to time series data. We do this in the context of using a trained model to make predictions from new data. After introducing the framework, we work through the application of deep learning to a number of areas in finance.
2. Introduction
Founder of Noviscient – a technology-led, research and proprietary trading firm.
We apply statistical and machine learning technologies to investment management
problems.
• Experience: 20 years in finance
• Learning: BEng, MBA, MQF, PhD candidate
• Programming: Python (Matlab, C)
• Trading: Asian markets
• scott.treloar@noviscient.com
November 2016 Noviscient Pte Ltd (www.noviscient.com) 2
3. Introduction
In this presentation we discuss a framework for predictive trading using deep learning
• The context
• General approach to systematic trading
• The challenge (prediction)
• Machine learning and deep learning
• An approach
• Some examples
• Wrap up
November 2016 Noviscient Pte Ltd (www.noviscient.com) 3
4. The context: problems in traditional asset management
November 2016 Noviscient Pte Ltd (www.noviscient.com) 4
5. Context: drivers of change
November 2016 Noviscient Pte Ltd (www.noviscient.com) 5
The North American asset management industry is on the brink of a once-
in-a-generation shift in competitive dynamics, due to five converging trends
McKinsey: Thriving in the New Abnormal (November 2016)
• The end of 30 years of exceptional investment returns
• A shake-up in active management - a large pool of benchmark-hugging active assets
(up to $8 trillion) will be up for grabs over the next several years
• The decline in average returns will also spur a third significant trend, a boost in the
steady stream of assets moving into alternative investments
• A true digital revolution in data and analytics – firms that can deliver both investment
and operational alpha will be industry leaders in the coming years
• An era of heightened regulation increasing legal and compliance costs and raising
barriers to entry
6. Context: the future of asset management
November 2016 Noviscient Pte Ltd (www.noviscient.com) 6
8. General approach to systematic trading
Necessary conditions for a successful systematic trading business:
1. Source of trading ideas
2. Strategy development framework to review the ideas and find the edge
3. Infrastructure to execute on the edge
November 2016 Noviscient Pte Ltd (www.noviscient.com) 8
9. Necessary Condition 1 – source of trading ideas
November 2016 Noviscient Pte Ltd (www.noviscient.com) 9
10. Necessary Condition 2 – strategy development framework
Trading is an information game. We are trying to input various traditional and
alternative data sources into one or more models and combine the output to produce
an attractive portfolio return distribution.
• Choice of market instruments and trading frequency
• Development of trading rules for entry and exit and any trailing stops or profit takes
• Position-sizing and target volatility calculation
• Method for incorporation into the broader portfolio
• Robust backtesting methodology
• Approach to forward testing
All within a hypothesis-testing framework to reduce over-fitting problems
November 2016 Noviscient Pte Ltd (www.noviscient.com) 10
11. Necessary Condition 3 – infrastructure
November 2016 Noviscient Pte Ltd (www.noviscient.com) 11
Principles
• Cloud-hosted
• Open-source
• Single language
• Modular
• APIs
• Messaging
12. Digression
November 2016 Noviscient Pte Ltd (www.noviscient.com) 12
Why are we here at this conference?
• I think it is because we don’t really want to be
employees in large institutions
• Maybe we want to be our own boss, or work in
small entrepreneurial teams where we control our
own fate?
• The infrastructure for institutional-level, systematic
trading is becoming available through groups like
Quantopian and Noviscient
• We are getting to the point where we can build a
career in systematic trading
13. Digression
Why now for systematic trading?
Quantitative trading segments:
• Low frequency risk premia traders
• Mid frequency systematic traders
• High frequency market makers
The scarce resource for systematic
trading is not capital or infrastructure.
It is alpha generating capability.
November 2016 Noviscient Pte Ltd (www.noviscient.com) 13
Capital
(risk premia)
Ideas
(alpha)
Infrastructure
(speed)
14. Strategy development - the challenge
Objective
• Prediction – we want to predict the future given only the information we have today
Problems
• Inputs are unknown
• Model linking the inputs to the outputs is unknown
• Noise / signal ratio is very high
• The data generating process generating the outputs is likely to be time-varying
November 2016 Noviscient Pte Ltd (www.noviscient.com) 14
Input (t) Output (t+1)Model (t)
16. Why now?
November 2016 Noviscient Pte Ltd (www.noviscient.com) 16
Machine Learning
Algorithms
Cloud
Computing
Big Data
17. Current applications
November 2016 Noviscient Pte Ltd (www.noviscient.com) 17
• http://www.deeplearningbook.org/
Often quite static and benefit from big data.
18. Strategy development framework
Systematic trading is a very competitive space with relatively low barriers to entry.
To build a sustainable business you need a robust and differentiated approach to
identify alpha opportunities.
This means being set up to take advantage of
November 2016 Noviscient Pte Ltd (www.noviscient.com) 18
Old New
Simple data Complex data
Linear models Non-linear models
Batch processing Online processing
Static Dynamic
Dynamic Adaptive
19. Daemon
Why?
• Complex data – sensors, data feeds, voice …
• Non-linear processing – filtering
• Online
• Dynamic
• Adaptive
November 2016 Noviscient Pte Ltd (www.noviscient.com) 19
20. Some of our research areas
Recursive least square filters
• an online filter that recursively looks for coefficients that minimize a weighted linear
least squares cost function
November 2016 Noviscient Pte Ltd (www.noviscient.com) 20
21. Some of our research areas
Extreme learning machines
• Feedforward neural network for classification (up /
down) with a single layer of hidden nodes, where the
weights connecting inputs to hidden nodes
• Very fast – only (?) matrix inversion
• Can use with recursive least squares
November 2016 Noviscient Pte Ltd (www.noviscient.com) 21
22. Some of our research areas
Use of autoencoding and deep learning
• Learn a representation of the data – similar
to dimensionality reduction
• Calculate the mutual information between
the stock an its encoded / decoded
representation
• Use deep learning on a ‘smart’ subset of the
stocks to achieve a certain portfolio return
objective
November 2016 Noviscient Pte Ltd (www.noviscient.com) 22
23. Wrap up
• The future is systematic
• Still early days – we are on or near the
ground floor for finance applications
• We also need to go towards online, non-
linear, dynamic and adaptive
• If you want to learn more from us at
Noviscient we have the first of series of
workshops being held in Singapore on 5-6
December (see our website for details).
Questions?
November 2016 Noviscient Pte Ltd (www.noviscient.com) 23