ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES

IN QUANTITATIVE ASSET MANAGEMENT

Electroniccopy
Electronic copy available
available at:
at: https://ssrn.com/abstract=3122237
https://ssrn.com/abstract=3122237
ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

2 © 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY

Electroniccopy
Electronic copy available
available at:
at: https://ssrn.com/abstract=3122237
https://ssrn.com/abstract=3122237
 ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

Artificial intelligence has recently experienced a remarkable increase Most machine learning methods have been developed outside of
in attention, following staggering achievements in applications such finance and built on well-known statistical models such as linear
as image, text and speech recognition, self-driving cars or chess and regression or clustering techniques. Still, machine learning allows
Go tournaments. It is therefore not surprising that also the financial for much more flexibility, for example, by allowing for nonlinearities
industry is ever more heavily trying to improve investment decisions and feedback effects, or by completely refraining from any prior
by incorporating self-learning algorithms into the investment process. knowledge about the problem under scrutiny. It can be applied to
For that matter, the application of quantitative tools and algorithms different kinds of problems, such as classification or regression anal-
in order to define systematic trading strategies has already a strong ysis. Classification algorithms group observations into a finite number
history in the hedge fund industry. Against this backdrop, quantitative of categories, whereas regression analysis estimates outcomes to
hedge funds may provide a fertile soil for the application of new problems that have an infinite number of solutions. While machine
machine learning techniques. But do all sectors of the asset manage- learning is a very broad field, it can be classified into three main
ment industry exhibit characteristics that can be exploited by artificial areas.
intelligence tools to uncover new patterns? What could be the espe-
cially relevant fields? Are there limits beyond which additional computing The currently most successful field is supervised learning, where
power and greater data availability have only marginal benefits? This algorithms learn based on training data that reveal known relation-
research note provides some initial answers. It shows that the adap- ships. Examples of supervised learning include tasks such as the
tivity and self-learning capability of machine learning tools could add detection of fraud in credit card transactions or the creditworthi-
value along the entire value chain of an asset manager. However, the ness of debtors. The simplest form of a supervised learning algorithm
inherently flexible nature of machine learning methods is also the is linear regression, which makes a prediction using a linear func-
biggest challenge. These methods must be applied thoughtfully and tion of the input features, by learning the relationship based on
in the right context. We start with a general overview of machine minimising the mean squared error between predicted and true
learning, then elaborate on specific applications in quantitative asset regression outputs. There is a general trade-off between optimis-
management, highlighting the limitations, challenges and possible ing the fit of a model on the in-sample training and the true
remedies before reaching our conclusions. out-of-sample prediction period. Given that all models tend to fit
the training data better the more input variables are used, it may
From Machine Learning in General ... be reasonable to penalise additional model complexity in order to
Machine learning refers to extracting knowledge from data by iden- maintain sufficient generalisation power for the prediction task.
tifying correlated relationships without receiving prior information Methods such as ridge or lasso regression help in automatically
about what causal dependencies to look for. It combines elements detecting the most relevant input variables by regularising model
from both statistics and computer science and has been in existence complexity to avoid overfitting. While the ridge regression relies on
for many years. As early as 1956, John McCarthy at a conference on minimising the importance of less relevant factors, the lasso regres-
the campus of Dartmouth College coined artificial intelligence as sion completely discards input parameters whose importance lies
“the science and engineering of making intelligent machines”. below a certain threshold. All three methods are, by nature, linear,
However, it is mostly due to recent significant advancements in com- but may account for non-linear relationships based on an appropri-
puting power and data availability that the application of artificial ate manipulation of the input variables, for example, by interacting
intelligence algorithms has become relevant in everyday life. themselves or by building polynomials on the original data. A simple
machine learning method that is not constrained to linear relations
Figure 1: Artificial intelligence and exemplary methods is the k-Nearest Neighbours algorithm. This model looks for the k
historical data points that come closest to the current situation and
Supervised Learning Unsupervised Learning predicts future values based on these historical “neighbours”. There
Linear Regression, Ridge, Lasso, Clustering (k-Means), Factor exist more complex non-linear supervised learning algorithms such
k-Nearest Neighbours, Decision Analysis (PCA, Manifold as decision trees or random forests, which in essence learn by a
Trees Learning) sequence of if/else rules that get to the true answer the fastest.
However, they are not able to extrapolate or make forecasts outside
Reinforcement Learning
of the range of the training data.
Deep Learning

Multilayer, Feed-Forward Neural Networks

Source: Aquila Capital Concepts GmbH

© 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY 3

Electronic copy available at: https://ssrn.com/abstract=3122237
ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

Contrary to the methods described in the section above, unsuper- source to validate or further guide the learning process. A reward
vised learning algorithms only receive input data to learn from, but feedback is required for the algorithm to learn a certain behaviour.
no information about the output data or relationships. These algo- For instance, self-driving cars can learn collision avoidance by learn-
rithms therefore detect patterns in the data by identifying clusters ing from the negative feedback received from crashing into obstacles,
of observations that depend on similar characteristics. Machine or computers can teach themselves the rules of games such as chess
learning can, for example, be used to identify the main topics in the or Go. Hence, the reward feedback need not necessarily be pro-
news flow for a given stock, or to look for a liquid security that vided by a human being.
mimics the characteristics of an illiquid asset. At the core of unsu-
pervised learning algorithms is the idea of reducing dimensionality Artificial intelligence literature also frequently refers to deep learn-
by clustering the data or by transforming it into simpler factor rep- ing or neural network algorithms. This kind of method, in a certain
resentations. Clustering methods partition the input data into subsets sense, mimics the function of the human brain by feeding informa-
that exhibit common characteristics, such that the data points within tion through different layers and nodes. It can be applied to all of
a cluster share some notion of similarity that decisively discriminates the three areas outlined above. The simplest form is called multi-
them from the data points in other clusters. Factor analysis, on the layer perceptron and can be seen as a generalisation of linear models
other hand, relies on transforming the original data into the most that perform multiple regression steps. Each node of a certain layer
relevant drivers or the most appropriate representation. The prin- of the network computes a weighted sum of the input information
cipal component analysis, for example, first finds the direction of received from the nodes of the foregoing layer and applies a non-
maximum dispersion in the data, assuming that this dispersion con- linear function to the result. This process is repeated multiple times
tains most of the information about the first factor driving the data. until the final output value is calculated. Given that in this setup the
It then finds the direction in the data that again features the highest information passes through each node only once, such neural net-
dispersion but, at the same time, is uncorrelated to the previously works are also designated feed-forward. There exist more advanced
identified factor. networks to deal with the challenges of simple networks which, for
example, also propagate information backwards through the network.
Combining methods of supervised and unsupervised learning results Given that even a sketchy synopsis of this field would exceed the
in so-called reinforcement learning, where an algorithm first detects scope of this research note, we refer the interested reader to cor-
patterns on its own, and then receives feedback from an exogenous responding abundant literature.

... To Specific Applications in Quantitative Asset Management ...

Figure 2: Artificial intelligence applications in quantitative finance

Investment Process Possible Machine Learning Applications Exemplary Methods

Investment Universe nn Identifying uncorrelated assets nn Dendrogram

nn Data mapping allowing for interpretation nn PCA, manifold learning
nn Proxy assets with more liquid instruments nn k-means

Signal Creation nn Create non-linear forecast nn Linear, lasso, ridge regression

nn Forecast based on predefined factors nn k-nearest neighbours
nn Forecast without prior factor knowledge nn Bag-of-words, term frequency-inverse document
nn Forecast based on new data sources frequency

Portfolio Construction nn Improve estimate of input parameters nn PCA

nn Dynamically maximize target value nn Lasso, ridge regression
nn Scenario and stress testing nn k-means

Trading nn Pricing based on sparse data nn k-means

nn Non-linear relations in transaction costs nn Lasso, ridge regression

Source: Aquila Capital Concepts GmbH

4 © 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY

Electronic copy available at: https://ssrn.com/abstract=3122237
 ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

Each sector within the financial industry uses artificial intelligence regulatory and investor specific limits into account. Finally, the trading
methods differently and according to their specific needs. Insurance engine translates changes in the model portfolio positions into
companies, for example, use it among other applications to detect effective trades, while routing orders between different market
fraud in damage events. Banks count on it to assess credit quality venues, minimising transaction costs and avoiding slippage and
or employ robotic process automation for middle and back office market impact. In contrast to our structured approach with clearly
tasks. In wealth management, applications are only just emerging, defined tasks for the application of artificial intelligence, it would
with a current focus on automatically ensuring regulatory require- also be possible to more generally ask such an algorithm to com-
ments, client segmentation, chatbots to deal with repetitive client pletely choose an investment rationale on its own. However, we
requests or portfolio management based on robo advisors. But in cover that question in the section that discusses the challenges and
what areas of asset management can artificial intelligence provide limitations of machine learning models.
added value? What problems faced by investment managers may
not be solved by such tools? What then, in more detail, do these general descriptions adhere to?
Let us start with the investment universe, where machine learning
When we think about modelling the investment process – and about tools may assist in identifying uncorrelated assets that provide true
using machine learning algorithms to improve decision making – diversification benefits; or in the mapping of data into new rep-
we can subdivide the value chain into different steps. This allows resentations that allow for other interpretations, such as the detection
for a systematic application of models that are appropriate for a of style drifts in hedge fund strategies or, for instance, factor expo-
specific task. We will provide details of the various steps in the next sures such as momentum or value. An appropriate tool for the first
section, but start with an overview in order to facilitate the task would, for example, be a dendrogram analysis (see case study
synopsis. 1: Clustering the Investment Universe with Dendrograms); the second
goal could be achieved by relying on a principal component or man-
The first step consists of the defining of the investment universe, ifold learning analysis. In a similar manner, artificial intelligence
and of collecting and cleaning the corresponding data sets. Next, methods can be used to proxy valuation or even the actual invest-
the alpha engine or signal engine preprocesses the data in a proper ment of assets for which there is only sparse historical market data
way, calculates the signals for the various markets under scrutiny or that are not eligible due to liquidity issues. In the latter case this
based on the models used, and maps these signals into the portfo- can be achieved by substituting more liquid instruments that appro-
lio context. Then, the portfolio construction engine or risk management priately mimic the characteristics of the desired assets. A useful
engine builds the theoretical model portfolio based on a suitable variant to achieve that task would be the k-Nearest Neighbour
algorithm, such as mean-variance or maximum diversification, taking model.

© 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY 5

Electronic copy available at: https://ssrn.com/abstract=3122237
ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

Case Study 1: Clustering the Investment Universe with

Dendrograms

Dendrograms belong to methods of hierarchical clustering. The sub-clusters into a hierarchical order, depending on the correlation
algorithm iteratively clusters first individual data points and then structure.

Figure 3: Dendrogram for investment universe

Hierarchical clustering commodities by correlation
0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
Natural Soybeans Soybean Corn Wheat Gas Oil Heating Crude Oil Copper Aluminium Gold Platinum
Gas Meal Oil
Source: Aquila Capital Concepts GmbH

We use a dendrogram to structure a set of individual commodity two clusters are then joined to form a more general cluster of metals.
markets into more meaningful clusters. Ideally, it comes up with Similarly, heating oil and crude oil are merged before being clustered
well-known sectors, perhaps energy, precious metals and industrial together with gas oil to form an energy complex. The energy and metals
metals. clusters are then combined to form a cluster of commodities that are
highly dependent on changes in the business cycle. The soft commod-
At the bottom of the visual representation in Figure 3 are the single ities – soybeans, soybean meal, corn and wheat – are structured into
data points that are joined in first clusters. For example, the model a separate node that only consists of agricultural products. Interestingly,
groups copper and aluminium into a mutual cluster of industrial natural gas forms an individual cluster, most likely because of season-
metals, or gold and platinum into a cluster of precious metals. These ality factors that separate it from other energy commodities.

The aim of the alpha or signal engine in our context is to generate An elucidatory example would be the dynamic, selective weightings
forecasts about the direction and magnitude of future asset price of a given set of moving averages over various time windows,
movements or about the riskiness of assets, and to translate that depending on some historical pattern or exogenous factors. More
information into a meaningful signal for the portfolio construction involved is a third application, consisting of letting the model select
engine. Potential applications of machine learning methods in this relevant input signals on its own or access new data sources in order
field can be classified into three main blocks. First, an artificial intel- to extract additional information, for example, by clustering social
ligence algorithm may be helpful in creating a nonlinear forecast media posts or news announcements in order to construct alterna-
based on a single time-series (see Case Study 2: Analysing the Behav- tive sentiment indicators based on a bag-of-words or term
iour of the VIX Index with KNN). Second, machine learning methods frequency-inverse document frequency (TFIDF) algorithm.
may derive forecast value out of a predefined pool of relevant factors.

6 © 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY

Electronic copy available at: https://ssrn.com/abstract=3122237
 ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

Case Study 2: Analysing the Behaviour of the VIX

Index with KNN

Figure 4: KNN for predictive power of moving averages

Positive
Long Moving Average

0
Negative

Negative 0 Positive
Short Moving Average
Source: Aquila Capital Concepts GmbH

The VIX Index measures market expectations for the volatility of the negative, and the green points indicate future positive directional
S&P 500 Index over the coming month, based on index option prices. changes. While it is difficult to extract a meaningful interpretation
Given that volatility can neither become negative nor grow bound- from this scatter plot, a k-nearest neighbour analysis reveals further
lessly, theory suggests a mean reverting behaviour. Additionally, the information. Based on this estimator, Figure 4 shows the decision
distribution of changes in volatility is commonly skewed, mimicking boundaries for the two states of future directional movements in
the fact that spikes in volatility often occur very quickly, while a reduc- separate colors. The area coloured in blue represents states where
tion in volatility normally takes more time and tracks a bumpier road. the two moving averages indicate falling VIX levels, whereas the red
area stands for scenarios in which the two moving averages predict
Given that backdrop, we analyse the predictive power of two moving a rising VIX. Clearly, positive values for the moving averages are
averages of past index movements for the future direction of the related to negative future VIX price movements, confirming a mean
VIX Index based on a k-nearest neighbour classification algorithm. reverting behaviour after an increase in volatility levels. The picture
Figure 4 plots the values of the short moving average on the x-axis for negative moving average values is more ambiguous, overall
and the values of the long moving average on the y-axis. Conditional upholding the thesis of mean reversion, but also showing some signs
on the value of these two moving averages, the blue points repre- of momentum. That makes intuitive sense, as volatility tends to trend
sent moments in time where the future VIX Index movement was lower after a sudden spike.

Based on these forecasts, the portfolio construction or risk manage- returns to maximum drawdown are maximised. Finally, enhanced
ment engine calculates target positions, taking regulatory and internal scenario analysis tools may improve model validation and stress
restrictions into account. In this step, artificial intelligence methods testing applications.
may be helpful in improving estimates for input variables such as
expected returns or the variance-covariance matrix for large portfo- The trading engine finally translates target positions into effective
lios in the context of limited historical data. This could, for example, market orders. This step is especially relevant for large asset manag-
be achieved by reducing the dimensionality of the dataset based on ers, as an estimated two thirds of gains on trades are lost due to
clustering algorithms such as a principal component analysis. Instead market impact costs when trading into and out of large position
of optimising the portfolio with respect to a predefined objective blocks. In this area, artificial intelligence may serve to obtain addi-
function and specific constraints, machine learning tools may also tional information from sparse historical data or help identify nonlinear
be asked to tweak the portfolio in a more general way. For instance, relationships in order flow to calibrate trading algorithms based on
by dynamically weighting the portfolio components such that risk-ad- both nonlinear regression tools and clustering algorithms.
justed returns as measured by Sharpe ratios or the ratio of average

© 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY 7

Electronic copy available at: https://ssrn.com/abstract=3122237
ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

... and Challenges and Limitations

In the previous section, we highlighted various steps along the value Artificial intelligence aims to extract relevant knowledge from pos-
chain of a well-structured investment process that, in our view, are sibly unstructured data on a self-learning basis. It works especially
suitable for further enhancement by machine learning applications. well for tasks with precisely defined rules and stable probability dis-
Does that mean we are unconditional believers in the merits of arti- tributions, such as mastering demanding games like chess or Go.
ficial intelligence? Far from it! There are a number of challenges and Machine learning tools also perform excellently in piloting self-driv-
limitations that are not necessarily new to quantitative investment ing cars, where the necessity to follow roads and traffic rules regulate
managers, which may be aggravated by the flexibility of new the set of possible actions. However, the stochastic nature of finan-
techniques. cial markets with their lack of stable rules and probability distributions may
challenge the validity of relationships that are learned from the past.
Accordingly, models should always be applied to clearly defined
problems and validated against sound theoretical assumptions.

Figure 5: Challenges and possible remedies

Challenges for ML in asset management Possible remedies

Lack of stable rules and probability distributions nn Apply models to clearly defined problems
nn Validate results against sound theoretical assumptions
Limited data availability nn Focus on parsimonious model structures
nn Account for mightiness of models to learn the past
Lack of causal reasoning and imagination nn Keep simplified transformation in mind
nn Diligent analysis before investing resources and modelling power
Future regulation and susceptibility to nn Check future accessibility of data sources
manipulation nn Validate contributors to data sources
Source: Aquila Capital Concepts GmbH

Similarly, self-driving cars can be driven along the same roads as brought about by the Swiss National Bank in early 2015? Did quan-
many times as is necessary to teach them all the relevant aspects titative models foresee the taper announcement by the US Federal
of a journey and AlphaGo can play with itself until it perfectly Reserve Bank in the spring of 2013? Most likely not. In a similar
masters the rules of the game. However, despite a seemingly abun- manner, machine learning algorithms may just find theories that are
dant access to data, there is only one historical price trajectory for already well-known and proven. While this confirmation may add
each financial market to train a model on. This limited data avail- comfort, it may also just be a waste of time and money. So, despite
ability restricts the complexity of the artificial intelligence model increasing computer power and data availability, it is necessary to
that can be applied and, therefore, the flexibility of the model’s keep in mind that quantitative models remain a simplified transfor-
output when forecasting future price movements. This problem is mation of the world and will only have forecasting ability that is
further deteriorated by the fact that the vast majority of data for limited to specific tasks. Additionally, the complexity of calibrating
financial markets has only been collected recently. As a conse- artificial models requires a diligent analysis as to where to allocate
quence, researchers should focus on parsimonious model structures resources and model power most effectively.
and not be misled by the mightiness of artificial intelligence models
to adaptively learn the past. Other more general potential limitations include future regulation
and susceptibility to manipulation. While the former may limit the
Next, machine learning models excel at identifying relationships in use of data due to calls to protect privacy rights or fears of market
data that may be unrecognisable to the human eye by revealing infrastructure instability caused by black box models and automated
complex correlation structures. Still, they lack the causal reasoning trading processes, the latter may, for example, threaten the validity
and imagination that would be necessary to anticipate events that of contaminated information retrieved from fake social media user
have not happened in the same way many times before. Would a accounts or blogs posted by manipulated chatbots. It is worth check-
trading model based on artificial intelligence have been able to ing the likely future accessibility of a data source and the way it is
predict the currency peg break between the Euro and the Swiss franc comprised before basing investment models on it.

8 © 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY

Electronic copy available at: https://ssrn.com/abstract=3122237
 ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

Conclusion
This research note is aimed at providing a framework for assess- methods is also their biggest challenge. It requires that the
ing the opportunities and challenges of artificial intelligence methods are put in the right context and thoughtfully applied
methods within a structured investment process. It highlights so as to solve questions in a way that produces meaningful out-
that machine learning tools, owing to their adaptivity and comes. It would be illusory to believe that artificial intelligence
self-learning capabilities, may add value along the entire value can develop a profitable investment rationale on its own. Accord-
chain of an asset manager in two specific ways. First, by more ingly, we are convinced that machine learning will most likely
effectively using currently available data based on algorithms not turn out to be a holy grail, but that it will help quantitative
that learn to reveal new, nonlinear relationships or by transform- investment managers in further improving their allocation pro-
ing data into representations with more interpretable meanings. cesses. Nevertheless, the use of artificial intelligence in asset
Second, by embracing new data sources that provide additional management is undoubtedly still at an early stage. This research
information, for example, based on news flow and blog posts. note, consequently, provides evidence based on first experiences,
However, the inherently flexible nature of machine learning but no final results. We are looking forward to an exciting future.

© 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY 9

Electronic copy available at: https://ssrn.com/abstract=3122237
ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

Authors and Investment Team at Aquila Capital

Aquila Capital’s Systematic Trading Group has considerable experi- The team members combine a variety of academic backgrounds,
ence in managed futures (CTA and multi-asset, long-only), equity including mathematics, physics, computer science, econometrics and
quant, statistical arbitrage and volatility strategies. This expertise economics, with a proven track record of successfully developing
forms the foundation and rigorous scientific background upon which and implementing managed futures strategies.
the company’s quantitative strategies have been developed, using
the latest technology.

The Aquila Systematic Trading Group

Urs Schubiger Egon Ruetsche, PhD Fabian Dori

Quantitative Strategist Quantitative Strategist Quantitative Strategist

Urs Schubiger has comprehensive experi- Egon Ruetsche has in-depth experience in Fabian Dori has more than ten years’
ence in the research, development and the development and modeling of quantita- experience in fund management and was
implementation of quantitative investment tive investment strategies. He was a partner previously Chief Investment Officer and
strategies as well as leading business of the Swiss subsidiary of Achievement AM member of the Management Board of La
development initiatives. Prior to founding LLC before joining AQ Investment AG. Prior Roche Private Bank. Prior to this, he was
prime Capital and Achievement AM’s Swiss to that, Mr Rütsche was a senior quantita- Head of Portfolio Management and a
subsidiary, he held senior positions at tive researcher and portfolio manager at member of the Management Board at 1741
leading companies including Vescore Ltd, Man AHL where he developed momentum Asset Management, as well as a portfolio
1741 Asset Management Ltd, Wegelin & and carry models and shared responsibility manager at Wegelin & Co. Private Bankers.
Co. Private Bankers and UBS O’Connor. Mr for managing volatility strategies. Egon Fabian Dori holds a Master’s degree in
Schubiger holds master’s degrees in Mathe- Ruetsche holds a master‘s degree in Mathe- Quantitative Economics and Finance from
matics from the ETH Zurich and in Law matics and a PhD in Arithmetic Geometry the University of St. Gallen and is Certified
from the University of Basel. from the ETH Zurich. Risk Manager and Chartered Financial
Analyst.

10 © 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY

Electronic copy available at: https://ssrn.com/abstract=3122237
 ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

© 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY 11

Electronic copy available at: https://ssrn.com/abstract=3122237
ARTIFICIAL INTELLIGENCE – CHANCES AND CHALLENGES IN QUANTITATIVE
ASSET MANAGEMENT

Contact Details | Aquila Group

Valentinskamp 70
20355 Hamburg
Germany
Tel.: +49 (0)40 87 50 50-199
E-mail: info@aquila-capital.com
Web: www.aquila-capital.com

Hamburg · Frankfurt · Munich · Luxembourg · London · Madrid · Oslo · Zurich · Tokyo

Important Risk Notice: Please note that all information have been collected and examined carefully and to the best of our knowledge; however,
the information is provided without any guarantee. All in-formation is believed to be reliable but we are not able to warrant its completeness
or accuracy. This document does not constitute an offer or solicitation for the purchase or sale of any financial instrument, is launched for
informational purposes only and is not intended to provide and should not be relied on for accounting, legal or tax advice, or investment
recommendations. Investors are alerted that future performance can vary greatly from past results. The value of investments
and income from them may rise as well as fall. You may not get back the amount initially invested. All investments involve risks
including the risk of a possible loss of principal. Emerging markets may be subject to increased risks, including less developed custody
and settlement practices, higher volatility and lower liquidity than non-emerging market securities. Some information quoted was obtained
from external sources we consider to be reliable. No responsibility can be accepted for errors of fact obtained from third parties, and this data
may change with market conditions. This does not exclude any duty or liability that Aquila Capital has to its customers under any regulatory
system. Investment regions, sectors and strategies illustrated in the document are for illustrative purposes only and should not be viewed as a
recommendation to buy or sell. The terms Aquila and Aquila Capital comprise companies for alternative and real asset investments as well as
sales, fund management and service companies of the Aquila Group. The respective responsible legal entities of the Aquila Group that offer
products or services are named in the corresponding agreements, sales documents or other product information. A publication of Aquila
Capital Concepts GmbH. As of 30.09.2017

12 © 2018 AQUILA CAPITAL – FOR PROFESSIONAL INVESTORS ONLY

Electronic copy available at: https://ssrn.com/abstract=3122237