3. 3Qraft Technologies, Inc.
AT THE
CROSSROADS OF
FINANCE AND
ARTIFICIAL
INTELLIGENCE
We utilize ordinary data to develop extraordinary
products, automating and optimizing your work
process. We aim to create real products to reduce real
costs based on real use cases as opposed to trendy
hypothetical experiments.
We have accumulated a broad range of experiences
in order to enhance our capabilities, possessing a
wide client network and establishing references by
cooperating with various research institutions,
technology firms, financial companies, etc.
Our experienced engineers and consultants
build customized artificial intelligence
systems by organizing and integrating
clients’ data for large-scale progress from
deep learning technology.
4. 4Qraft Technologies, Inc.
QRAFT
TECHNOLOGIES, INC.
T E A M Q R A F T
J a n u a r y 2 0 1 6
F O U N D A T I O N
M A N A G E M E N T
B U S I N E S S
C O M P A N Y W E B S I T E
M r. H y u n g - S i k K i m ( C E O )
A I S o l u t i o n D e v e l o p m e n t fo r A s s e t M a n a g e m e n t & Tra d i n g
w w w. q ra f t e c . c o m
U S E T F W E B S I T E
H E A D C O U N T
L O C A T I O N
S e o u l , S o u t h Ko re a
3 0 ( i n c l . 2 5 e n g i n e e rs )
w w w. q ra f t a i e t f. c o m
Q R A F T A I E T Fs
L i s t e d H e re
5. 5Qraft Technologies, Inc.
H Y U N G - S I K K I M
C H I E F E X E C U T I V E O F F I C E R
O U R T E A M
Qraft Technologies provides institutions and companies with the AI total
solution for investment operations. With accumulations of various deep
learning-based AI products and services, we succeeded in establishing AI
OCIO(outsourced chief investment officer) platform. Qraft AI OCIO platform
will help you innovate your existing asset management model and improve
productivity dramatically. Qraft Technologies is the first mover in this area
and will do its best to provide an artificial intelligence partner that meets
your needs.
WE ARE THE FIRST MOVER
IN THIS NEW AREA
6. 6Qraft Technologies, Inc.
G I L - LY E O L J E O N G
C H I E F T E C H N O L O G Y O F F I C E R
O U R T E A M
In treating illness, it is reasonable to follow a scientifically and statistically
validated doctor’s advice. In terms of investment, a lot of pseudo-
investment methods flood and mislead individual investors. However, it
would be the most reasonable to follow scientifically and statistically
verified investment methods. Qraft Technologies is realizing these ideals
with AI ETFs, AI asset allocation, AI portfolio and AI order execution engine
so that everyone can make reasonable and efficient investments. It is such a
joy to do this with great staffs here in Qraft Technologies.
WE STAND ON VERIFIED
INVESTMENT THEORIES
7. 7Qraft Technologies, Inc.
H Y O - J U N M O O N
D I R E C T O R ,
H E A D O F A I R E S E A R C H
O U R T E A M
The value that AI technology adds to the asset management market is likely
to be more about reducing the inefficiency of the modeling and operational
process than about increasing the size of alpha. AI can be operated in
parallel and the marginal cost of cloning is near zero, which help us make
numerous AI agents. The cost of active operation can be dramatically
reduced in conjunction with the automated process. In practice, the
modeling manpower, time and administrative costs to make a financial
product continue to decline.
WE ADD VALUES TO
ASSET MANAGEMENT
8. 8Qraft Technologies, Inc.
A LV I N K W O N
D I R E C T O R ,
H E A D O F A I P R O D U C T S
O U R T E A M
Problems are everywhere. It is important how fast we find and solve them
in a reasonable way. The strength of Qraft Technologies is that we are well-
experienced at problem discovery and resolution, especially in the area of
finance and investment. I am convinced that the transition to AI in asset
management is only a matter of time. The question is how. With Qraft
Technologies, you will be one who determines when and how of this
transition.
WE FIND AND SOLVE
INVESTMENT PROBLEMS
10. 10Qraft Technologies, Inc.
R E F E R E N C E S
AI ETF LISTING
ON NYSE
In May 2019, Qraft listed two deep learning-based AI ETFs, QRFT US and AMOM US,
in NYSE. Both ETFs are fully managed by our own proprietary deep learning engine
with virtually no human intervention. And Qraft listed HDIV US in Feb, 2020.
Qraft will continue to launch our own AI ETF products to establish ourselves as a
global AI ETF house.
11. 11Qraft Technologies, Inc.
R E F E R E N C E S
AI ETF PERFORMANCE: QRFT
Despite its high management fee, QRFT is outperforming the benchmark and all
other US large cap multi-factor ETFs (DYNF of Blackrock, VFMF of Vanguard, GSLC of
Goldman, FCTR of FirstTrust, FLQL of Franklin Templeton) since inception.
12. 12Qraft Technologies, Inc.
R E F E R E N C E S
AI ETF PERFORMANCE: QRFT
Since Inception US Large Cap Multi Factor ETF Comparison (2019-05-21~2020-03-25)
Name Ticker Manager
Management Fee
(bps)
Total Return(%) Alpha(%p, After fee) Fee Cost(%) Sharpe Ratio Alpha(%p,Before fee)
Qraft AI Enhanced US Large Cap ETF QRFT Qraft Tech 75 -5.93 6.15 0.63 -0.12 6.78
Goldman Sachs ActiveBeta US Large Cap Equity GSLC Goldman Sachs 9 -11.3 0.78 0.08 -0.31 0.86
SPDR S&P500 ETF Trust (Benchmark) SPY State Street 9 -12.08 0 0.08 -0.35 0.08
Blackrock US Equity Factor Rotation ETF DYNF Blackrock 30 -14.31 -2.23 0.25 -0.46 -1.98
Franklin LibertyQ US Equity ETF FLQL Franklin Templeton 15 -13.75 -1.67 0.13 -0.44 -1.54
iShares Edge MSCI Multifactor USA ETF LRGF Blackrock 20 -15.82 -3.74 0.17 -0.49 -3.57
First Trust Lunt US Factor Rotation ETF FCTR First Trust 65 -19.43 -7.35 0.55 -0.69 -6.80
Vanguard US Multifactor ETF VFMF Vanguard 18 -24.96 -12.88 0.15 -0.83 -12.73
-15
-10
-5
0
5
10
QRFT GSLC SPY DYNF FLQL LRGF FCTR VFMF
Alpha(%p, After fee)
-0.90
-0.80
-0.70
-0.60
-0.50
-0.40
-0.30
-0.20
-0.10
0.00
QRFT GSLC SPY DYNF FLQL LRGF FCTR VFMF
Sharpe Ratio
13. 13Qraft Technologies, Inc.
R E F E R E N C E S
AI ETF PERFORMANCE: AMOM
AMOM is outperforming its peers and SPMO (BM: S&P 500 Momentum index) since inception.
14. 14Qraft Technologies, Inc.
R E F E R E N C E S
AI ETF PERFORMANCE: AMOM
Since Inception US Large Cap Momentum ETF Comparison (2019-05-21~2020-03-25)
Name Ticker Manager
Management
Fee (bps)
Total Return(%) Alpha(%p, After fee) Fee Cost(%) Sharpe Ratio Alpha(%p,Before fee)
Qraft AI Enhanced US Large Cap Momentum ETF AMOM Qraft Tech 75 -8.78 3.3 0.63 -0.22 3.93
Invesco S&P500 Momentum ETF SPMO Invesco 13 -11.42 0.66 0.11 -0.29 0.77
SPDR S&P500 ETF Trust (Benchmark) SPY State Street 9 -12.08 0 0.08 -0.35 0.08
JP Morgan US Momentum Factor ETF JMOM JP Morgan 12 -13.37 -1.29 0.10 -0.44 -1.19
-2
-1
0
1
2
3
4
AMOM SPMO SPY JMOM
Alpha(%p, After fee)
-0.50
-0.45
-0.40
-0.35
-0.30
-0.25
-0.20
-0.15
-0.10
-0.05
0.00
AMOM SPMO SPY JMOM
Sharpe Ratio
15. 15Qraft Technologies, Inc.
R E F E R E N C E S
AI Robo-advisory
Bank
Public Funding sales (end
of ’18)
Note
Estimated sales Release date Source
Hana 113,882 ≥ $700 million July 2017 Qraft
Shinhan 113,901 ≥ $300 million Oct 2016 Qraft
Replaced on
Jan. 2019
Kookmin 142,858 ≤ $100 million Jan 2018 - Self-developed
Woori 113,657 ≤ $100 million May 2017 Fount Negotiating w/t Qraft
IBK 68,994 ≤ $100 million July 2018 Qraft
NH 76,639 ≤ $100 million Aug 2016 - Not adopted
Busan 15,193 ≤ $50 million Dec 2018 Qraft
Kyongnam 8,360 ≤ $50 million Dec 2018 Qraft
Over 70% of major banks in Korea are using Qraft’s AI robo-advisory solution.
17. 17Qraft Technologies, Inc.
AI EXECUTION
A X E
Brokerage firms receive transaction fees based on the trading volume from the
institutional investors and so forth.
To minimize the market distortions caused by massive trading orders, brokers
normally divide trading orders into pieces or utilize algorithm trading systems which
automatically break orders in parts.
Algo Trading System is a trading execution solution which designed to increase the
efficiency of trading when breaking large orders into pieces.
Algo Trading has algorithms such as TWAP, VWAP, MOC, POV, IS, Peg, etc.
These algorithms are predetermined and, therefore, cannot adapt to the fast-
changing market conditions instantly.
18. 18Qraft Technologies, Inc.
FUNCTIONS
A X E
AXE is a deep-learning-based execution system developed by Qraft Technologies
AXE is the second AI-driven trading execution system followed by JP Morgan’s LOXM
AXE is the world first commercialized AI-driven trading execution system (for
Shinhan Invest).
AXE is designed to discover the most efficient trading strategies and to learn and
evolve with fast-changing market conditions rather than just following
predetermined trading rules.
AXE was able to reduce 6bps of brokerage fees in terms of VWAP when it was
applied in the actual market (Oct. 2018 / Buy trades for KOSPI200 Equities)
20. 20Qraft Technologies, Inc.
CASES : AXE CHALLENGE
A X E
Compared with human professionals from institutions, AXE boasts superior performance in the
competion event. By winning the order execution competition, AXE has proven its performance.
Event Sponsor: NVIDIA, Shinhan Bank, PwC
Participants : Three Professional Dealers vs. AXE
Awards : USD 100k for Human Dealers
Trading Universe : 50 Korea-listed Equities
(5 days * 10 Equities / Predetermined Pool of 50 Equities / Randomly Selected 10 Equities from the Pool)
Trading Accounts : USD 1m for Each Individual
Competition Period : 5 Days
Trading Hours : 09:00~15:20
21. 21Qraft Technologies, Inc.
AXE FOR OTHER AREAS
A X E
By utilizing the deep learning model, we provide FX management services such
as optimal currency exchange service and forex hedging service in B2B.
In addition, we provide deep learning models such as raw material procurement
& inventory management in B2B. FRESHEASY(mealkit startup whose revenue is
USD 200m) and POSCO LNG plants and multiple petro-chemical companies are
using our AI engines.
23. 23Qraft Technologies, Inc.
CONVENTIONAL vs.
AI-DRIVEN
I N D U S T R Y O U T L O O K
Value
Portfolio
Fund
Manager
Strategy
Analyst
PB / Sales
Customer
Financial
Products:
ETFs
Funds
Execution
Trader
Dealer
Conventional (current)
Asset Management Process
Human
Human
Human Human Human
Portfolio
Deep Portfolio
Strategy
Q-DNN
PB / Sales
Customer
Financial
Products:
AI ETFs
AI Funds
Execution
QRAFT AXE
AI-driven
Asset Management Process
Human
Human
AI AI AI
+
AI(Robo)-
Advisory
AI
World First Korea First
Korea Largest
24. 24Qraft Technologies, Inc.
EMERGENCE OF
ACTIVE INDEX
I N D U S T R Y O U T L O O K
Active fundPassive fund
Active index fund (Smart Beta)
Passive funds follow a market index. They are stable and passive type of investments
even with low returns.
Active funds are investing for more than the benchmark index returns.
Active index funds construct portfolios from benchmark index and add securities
unrelated to the underlying index driving performance higher.
Mutual funds and hedge funds fall into the category of “active funds”.
Index funds take up the majority of “passive fund” market.
Active index fund market has not yet been preempted.
25. 25Qraft Technologies, Inc.
EMERGENCE OF
ACTIVE INDEX
Although there are some differences between the definitions of smart beta and
factors and smart beta is more like a marketing term, the two terms, in practice, are
often used interchangeably. So we could say that style factor investing grows in use,
and this leads to more smart beta ETFs.
According to the 2019 Global ETF Investor Survey conducted by Brown Brother
Harriman with ETF.com, 92% of U.S. respondents are investing in at least on smart
beta ETF, and 26% of them are using it to replace an actively managed mutual fund.
Respondents said they were looking for more active and smart beta ETFs.
Ticker Fund Name AUM Expense Ratio
VTV Vanguard Value ETF $52bn 0.04%
IWF iShares Russell 1000 Growth ETF $46bn 0.19%
VUG Vanguard Growth ETF $43bn 0.04%
IWD iShares Russell 1000 Value ETF $40bn 0.19%
VIG Vanguard Dividend Appreciation ETF $39bn 0.06%
USMV iShares Edge MSCI Min Vol U.S.A. ETF $36bn 0.15%
IVW Vanguard High Dividend Yield ETF $27bn 0.06%
SDY iShares S&P 500 Growth ETF $24bn 0.18%
DVY SPDR S&P Dividend ETF $21bn 0.35%
IVE iShares Select Dividend ETF $18bn 0.39%
Largest 10 Smart Beta ETFs as of Nov 1, 2019
* Source: ETF.com
I N D U S T R Y O U T L O O K
26. 26Qraft Technologies, Inc.
MONEY FLOW
I N D U S T R Y O U T L O O K
Fund flow shows that there is an overwhelming move from active funds to passive
funds.
Passive fund market & active index fund market are the ones that show growth.
Among the two, the growth rate of active index fund market is much higher than
that of the passive fund market.
27. 27Qraft Technologies, Inc.
AI-DRIVEN
ACTIVE INDEX
I N D U S T R Y O U T L O O K
There is no evidence that AI is good ar doing things that human managers can’t do,
such as predicting stocks that will rise tremendously. But, there are strong evidences
that AI can find optimal measure that weigh stocks. We have proved that AI can find
more optimal measures than marketcap-weighting that S&P500 index uses in US
large cap category.
Active index fund market is a very area where AI has many advantages such as
cheap costs, high productivity, and consistent alpha.
28. 28Qraft Technologies, Inc.
COST of
ACTIVE MANAGEMENT
I N D U S T R Y O U T L O O K
Most quant hedge funds employ more than 1,000 people with high salary, and their
AUM per employee is about $30M on average.
Assuming management fee rate of 1%, management fee per employee is calculated to
be at around $300K. This is just enough to cover the quant analysts’ base salary, which
is why quant hedge fund houses charge management fee rate of greater than 1% to
make the business profitable.
29. 29Qraft Technologies, Inc.
QRAFT’S
TARGET MARKET
I N D U S T R Y O U T L O O K
As the active index fund market has not been preempted and is expected to show the
greatest growth in the coming years, Qraft’s foremost priority is becoming the market
leader in this area.
With the advantage over cost structure and compatible performance, Qraft will also
target the quant hedge fund market.
30. 30Qraft Technologies, Inc.
I N D U S T R Y O U T L O O K
The limits of static factor calculation
mean that we could enhance an
investment strategy with variable
methodologies to measure factors.
31. 31Qraft Technologies, Inc.
ENHANCING STYLE FACTORS
USING DEEP NEURAL NETWORKS
Our basic idea is to find an optimal factor calculation methodology using deep neural
networks and enhance factor exposures. Although the seasonality of factors themselves is
still unavoidable, we believe we could boost long-term returns by optimizing exposure to
each factor.
For instance, the most common way to measure momentum is to calculate the increase of a
price one month before against a price twelve months ago. However, this static calculation
methodology can be relevant or irrelevant, depending on market situations. As we have
seen in the previous slide, it causes inconsistent short-term performances.
I N D U S T R Y O U T L O O K
32. 32Qraft Technologies, Inc.
Single-Factor
Enhancement
Using DNN
Non-Linear
Multi-Factor
Combination
Factor Analysis of
An Active Investment
Strategy
ACTIVE, QUANT
& AI ENHANCEMENT
D E E P A P P R O A C H
We disassemble the return of an active investment strategy to find which risk factors
attributed to it. Then, taking advantage of deep neural networks, we enhance
exposure to each risk factor with variable calculation methodologies. Lastly, we
combine all the enhanced factors non-linearly.
34. 34Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
Applying deep learning technologies to financial data is a different task from
developing deep learning technologies. Because financial input features (i.e., input
data) are extremely noisy and output labels (i.e., output data) from them are highly
uncertain. Besides, the size of financial data is relatively smaller than other types of
data.
Therefore, sophisticated skills to extract core information from small data with high
noise are vital to developing deep neural networks for financial investment tasks. We
have abundant experience with designing the architecture of deep neural networks
for financial investment tasks with proprietary applied deep learning technologies.
Ultimately, the competitive edge of our artificial intelligence technologies should lie
in our abilities to understand the nature of financial data, define deep learning tasks
for financial investments and develop applied deep learning technologies to
complete tasks.
APPLIED DEEP LEARNING
TECHNOLOGIES
34Qraft Technologies, Inc.
36. 36Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
The Qraft Deep Neural Networks (the “Q-DNN”) is an underlying architecture of our
deep neural networks specialized in extracting time-series and cross-sectional
information from financial data and work as a framework for most of our artificial
intelligence systems, including an asset allocation system and a factor allocation
system.
A vital feature of the Q-DNN is that it could extract information from data with fewer
parameters and strengthen its forecasting power mitigating an overfitting problem.
In some aspects, deep neural networks to extract information from data with fewer
parameters can be said to be very similar to the way human portfolio managers build
up quantitative investment strategies. So we are modeling deep neural networks for
each specific task based on the Q-DNN.
QRAFT DEEP
NEURAL NETWORKS
36Qraft Technologies, Inc.
37. 37Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
The Deep Learning Framework for Portfolio Optimization is a framework for end-to-
end deep portfolio optimization, including the pre-processing of input data, the
stable generation of output labels, the modulization of deep neural networks, and
other functions that are required to solve a portfolio optimization problem.
It trains the sub-modules of deep neural networks through pre-processing input data,
setting target volatility and a reward function, and generating output labels to
maximize a reward. This framework is based on supervised learning, but from the
optimization perspective, it is like policy-based optimization rather than value-based
optimization.
DEEP LEARNING FRAMEWORK
FOR PORTFOLIO OPTIMIZATION
37Qraft Technologies, Inc.
38. 38Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
The Uncertainty Quantification is a technology to estimate the heteroscedastic
uncertainty of outputs and help entire networks to make a more precise decision. A
financial investment task may require volatility information and the uncertainty of
outputs. But the uncertainty of outputs from financial data has a different
distribution every time.
We have deep neural networks to learn the heteroscedastic uncertainty of outputs
using an ensemble deep learning model and heteroscedastic loss. For example, the
Q-DNN extracts information from input data, and the Uncertain Quantification
technology estimates the uncertainty of outputs from the data.
UNCERTAINTY
QUANTIFICATION
38Qraft Technologies, Inc.
39. 39Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
Conventional denoising methods such as moving average, bilateral filter, Kalman filter
are useful to denoise input time-series data, but more hyperparameter needs to be
set to control the level of denoising.
Rather than setting more hyperparameter, which causes overfitting, our Deep Time-
series Denoising module concatenates first loss with denoising loss and automatically
setting a level of denoising through learning both original task and denoising task.
Our denoising loss is defined using a stacked 1-dimensional convolutional
autoencoder.
DEEP TIME-SERIES
DENOISING
39Qraft Technologies, Inc.
40. 40Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
The Multi-Task and Multi-Label Learning is a technology to remove noise in output
data. Different from input data, noise in output data cannot be removed because it
could cause a look ahead bias problem on simulation, and the structural information
of output data should be maintained.
To solve these problems, our Multi-Task and Multi-Label Learning also predicts
related features with the original target function and a more extended period
simultaneously through formulizing multi losses.
MULTI-TASK AND
MULTI-LABEL LEARNING
40Qraft Technologies, Inc.
42. 42Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
The Bayesian Neural Network-Based Hyperparameter Optimization (the
“Hyperparameter Optimization”) is an automated machine learning technology to
optimize the hyperparameters of a deep neural network using Bayesian neural
networks. For example, the Q-DNN is deep neural networks to extract features from
financial data and includes hyperparameters that determine the structure and the
performance of the networks.
Conventional Bayesian optimization that uses a Gaussian process has a problem that
the more hyperparameters, the higher the dimension of the search space of
hyperparameters is. It is unable to optimize a large number of hyperparameters.
Our Hyperparameter Optimization uses Bayesian neural networks or approximated
Bayesian neural networks that are good at solving a higher-dimensional problem
instead of a Gaussian process. The networks estimate the distribution of acquisition
function for each hyperparameter and quickly find an optimal combination of all
hyperparameters in a small number of attempts.
BAYESIAN NEURAL NETWORK-BASED
HYPERPARAMETER OPTIMIZATION
42Qraft Technologies, Inc.
43. 43Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
The Evolutionary Multi-Network Learning is a technology to train eight or more deep
neural networks simultaneously to avoid an overfitting problem. The performance of
a deep neural network extracting features from financial data depends on an initial
set of hyperparameters, and it often results in overfitting.
We train multiple deep neural networks at the same time, remove the bottom n
networks that show the lowest rewards at a designated interval of time, and input
new deep neural networks. After the competition process, we use the ensemble of
survivors as a final deep neural network, and it is instrumental in preventing an
overfitting problem.
EVOLUTIONARY
MULTI-NETWORK LEARNING
43Qraft Technologies, Inc.
45. 45Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
The Cooperative Multi-Agent Reinforcement Learning Model for Trading Order
Execution (the “Multi-Agent Model”) is the reinforcement learning model with neural
networks of three agents, each of which determines a trading order volume at a
designated interval of time, whether to cancel a portion of the current trading order
and an order allocation over the current bidding or asking prices respectively.
It, ultimately, aims to minimize trading costs responding to real-time market status
training the networks to maximize rewards such as trading order settlements and
saved trading costs compared to a volume-weighted average price.
COOPERATIVE MULTI-AGENT
REINFORCEMENT LEARNING MODEL
FOR TRADING ORDER EXECUTION
45Qraft Technologies, Inc.
47. 47Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
SERVER AND METHOD FOR PERFORMING
ORDER EXECUTION FOR STOCK TRADING (1/2)
This invention aims to provide a server and a method that derives a trading order
execution strategy for one or more stocks from trading data using deep neural
networks and perform real-time trading order executions for one or more stocks
based on the trading order execution strategy derived by the networks.
As an example of the implementation of the invention to achieve the technical
challenges outlined above, it could provide a server and a method that consist of:
(1) an input unit receiving trading data for one or more stocks;
(2) a trading indicator generation unit having supervised neural networks to generate
trading indicators for one or more stocks;
(3) a stock price prediction unit having supervised neural networks to predict a stock
price at a designated interval of time;
(4) a trading-order-execution-strategy generation unit having reinforced neural
networks to derive a trading order execution strategy for one or more stocks based
on the trading indicators and the stock prices predicted at the previous stages; and
(5) a trading-order execution unit performing real-time order executions for one or
more stocks based on the trading order execution strategy derived at the previous
stage.
An input unit receives trading data for one or more stocks. The data, for example,
may include a book of buy and sell orders, real-time prices and trading volume,
settlements and market time stamps.
A trading indicator generation unit has supervised neural networks to generate
trading indicators for one or more stocks, such as a volume curve of all trading orders
that should be settled during the market hours of a day reflecting an investor’s risk
profile.
A stock price prediction unit has supervised neural networks such as attention
networks and recurrent neural networks to predict the rate of change of a stock price
at a designated interval of time based on the trading data from the previous input
unit. It trains the networks to minimize the mean squared error between the
predicted rate of change of a stock price and an actual rate of change at each interval
of time.
48. 48Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
SERVER AND METHOD FOR PERFORMING
ORDER EXECUTION FOR STOCK TRADING (2/2)
A trading-order-execution-strategy generation unit has reinforced neural networks
including multiple actors for action policies and multiple critics for action values to
derive a trading order execution strategy for one or more stocks that could minimize
trading costs responding to the current market status based on the trading indicators
and the predicted stock prices from the previous trading indicator generation and
stock price prediction units.
A trading-order execution unit performs real-time order executions for one or more
stocks based on the trading order execution strategy corresponding to the current
market status from the previous deep learning model generation unit.
49. 49Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
DEVICE, METHOD AND COMPUTER PROGRAM
FOR GENERATING PORTFOLIOS
BASED ON DEEP LEARNING (1/2)
This invention aims to provide a device, a method, and a computer program that
creates a custom investment portfolio based on deep learning technologies. More
specifically, it intends to allow individual investors to have an investment portfolio
tailored to their financial status and investment profile without visiting a bank branch
or having a face-to-face meeting with a portfolio manager.
As an example of the implementation of the invention to achieve the technical
challenges outlined above, it could provide a device, a method, and a computer
program that consist of:
(1) an input unit receiving investment information;
(2) a fund selection unit selecting one or more funds from a pool of investable funds
based on the information collected at the previous stage;
(3) a portfolio generation unit generating one or more portfolios with the
combination of funds selected at the last step;
(4) a weight allocation unit adjusting fund weights in each portfolio made at the
previous stage based on each fund’s expected return estimated by deep neural
networks; and
(5) a final portfolio generation unit selecting a final portfolio from a pool of portfolios
generated at the previous stage.
An input unit receives investment information to analyze the investment profile of an
individual investor. The information, for example, may include countries or regions
the investor does or doesn’t want to invest in, investment themes the investor does
or doesn’t want to invest in, the number of funds of a portfolio, and other
information necessary for an investment profile analysis.
50. 50Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
DEVICE, METHOD AND COMPUTER PROGRAM
FOR GENERATING PORTFOLIOS
BASED ON DEEP LEARNING (2/2)
A fund selection unit selects one or more funds from a pool of investable funds that
are preliminarily filtered by various fund evaluation criteria such as the combination
of a momentum factor index and a relative strength index. Traditional portfolio
offerings don’t require a fund screening process as they mainly use index funds to
construct a portfolio. However, because the invention is expected to use the pool of
funds distributed by a designated fund distribution channel, a fund screening process
is essential to create optimal portfolios with a random pool of funds.
A portfolio generation unit generates one or more portfolios with the combination of
funds from the previous fund selection unit.
A weight allocation unit takes advantage of deep neural networks to forecast the
expected return of each country, region, or asset class over the next investment
period, such as a quarter. Then, based on the forecasts, it allocates weights to the
funds of each portfolio from the previous portfolio generation unit complying with all
limits pre-set by a portfolio service provider and an investor. Besides, before deep
learning-drive weight allocation, it uses fund weights based on traditional
quantitative investment strategies such as a momentum factor strategy and a risk
parity strategy as a starting point. It could increase the reliability of portfolios
generated by the networks.
A final portfolio generation unit selects a final portfolio from a pool of portfolios from
the previous weight allocation unit based on the expected return of each portfolio or
the similarity of each portfolio with its benchmark.
51. 51Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
DEVICE, METHOD AND COMPUTER PROGRAM
FOR PREDICTING ASSET PRICES BASED ON
ARTIFICIAL INTELLIGENCE (1/2)
This invention aims to provide a device, a method, and a computer program that
predicts asset prices based on input data, including historical prices and other data
related to assets. More specifically, it intends to predict asset prices mapping the
probability distribution of input data as an image and using a tree model and
convolutional neural networks.
As an example of the implementation of the invention to achieve the technical
challenges outlined above, it could provide a device, a method, and a computer
program that consist of:
(1) an input unit receiving historical price data and other data related to assets;
(2) a first prediction unit predicting asset prices over a test period using historical
training data;
(3) a comparison unit comparing predicted asset prices with actual asset prices;
(4) an image generation unit generating a three-dimensional image map with
historical asset prices, predicted asset prices, and actual asset prices;
(5) a second prediction unit predicting a three-dimensional image map before a
specific time in the future using the image map generated at the previous stage; and
(6) a final prediction unit predicting asset prices at a particular time in the future
based on the image map created at the last step.
An input unit receives historical price data and other data related to assets that can
be measured by the number of time-series data and the length of each time-series
datum. The data, for example, may include equity market indices (e.g., S&P 500,
KOSPI), macroeconomic data (e.g., unemployment rate, money supply), and financial
indicators (e.g., price-to-earnings ratio, dividend yield).
52. 52Qraft Technologies, Inc.
D E E P T E C H N O L O G I E S
DEVICE, METHOD AND COMPUTER PROGRAM
FOR PREDICTING ASSET PRICES BASED ON
ARTIFICIAL INTELLIGENCE (2/2)
A first prediction unit predicts asset price over a test period (e.g., T+1 to T+24) using
the historical data of a training period (e.g., T-24 to T) and may employ a tree model
that can classify input data quickly.
A comparison unit calculates the difference between predicted asset prices and
actual asset prices over a test period and estimates the accuracy of the first
prediction.
An image generation unit generates a three-dimensional image map with historical
asset prices over a training period (e.g., T-24 to T), predicted and actual asset prices
over a test period (e.g., T+1 to T+24) and the difference between predicted asset
prices and actual asset prices over a test period (e.g., T+1 to T+24). For the
generation of a three-dimensional image map of price data, it generates a group of
two-dimensional images and then piles them in chronological order.
A second prediction unit predicts a three-dimensional image map before a specific
time in the future (e.g., T+24); it aims to predict asset prices to a particular time
using the image map from the previous image generation unit and convolutional
neural networks that are specialized in extracting features from images.
A final prediction unit predicts asset prices at a specific time in the future (e.g., T+24)
based on the image map from the previous second prediction unit.
54. 54Qraft Technologies, Inc.
S Y S T E M A R C H I T E C T U R E
A deep learning engine for investment model generation, first of all, needs a training
environment to learn big data. A model that completed data learning automatically
operates to create daily investment portfolios. All the processes should be recorded
and monitored by a graphical user interface.
FULLY AUTOMATED
INVESTMENT SYSTEM
MARKET DATA
MACRO DATA
UNSTRUCTURED DATA
KIRIN API DEEP LEARNING
DAILY OPERATION
USER INTERFACE
S&P Compustat
Thomson Reuters PostrageSQL
Fred
Quandl File System
MongoDB
Fred API
Text
Summary
NLP Engine
Historical
Universe
Sectors &
Industries
Adj. Price
Mkt. Cap
Etc.
Universe
Mask
Monthly
Beta
3 Factors &
Residual
Etc.
Tailored
Data
Derived
Data
Kirin Port
AI Model
Training
Portfolio
Generation
PostrageSQL
Back-up
Auto
Reporting
Weight
Calculation
Portfolio
Validation
PostrageSQL
Back-up
Auto
Reporting
Record
Board
Web
GUI
Cloud
The Architecture of Qraft AI-Enhanced Quantitative Investment System
55. 55Qraft Technologies, Inc.
S Y S T E M A R C H I T E C T U R E
The Qraft AI-Enhanced Quantitative Investment System is fully automated, and
consists of five parts of (1) data feed, (2) data integration & API management, (3)
deep learning & model generation, (4) daily operation of generated models, and (5)
web user interface for end-users.
FULLY AUTOMATED
INVESTMENT SYSTEM
MARKET DATA
MACRO DATA
UNSTRUCTURED DATA
KIRIN API DEEP LEARNING
DAILY OPERATION
USER INTERFACE
S&P Compustat
Thomson Reuters PostrageSQL
Fred
Quandl File System
MongoDB
Fred API
Text
Summary
NLP Engine
Historical
Universe
Sectors &
Industries
Adj. Price
Mkt. Cap
Etc.
Universe
Mask
Monthly
Beta
3 Factors &
Residual
Etc.
Tailored
Data
Derived
Data
Kirin Port
AI Model
Training
Portfolio
Generation
PostrageSQL
Back-up
Auto
Reporting
Weight
Calculation
Portfolio
Validation
PostrageSQL
Back-up
Auto
Reporting
Record
Board
Web
GUI
Cloud
The Architecture of Qraft AI-Enhanced Quantitative Investment System
1
DATA FEED
3
DEEP LEARNING
4
DAILY OPERATION
5
USER INTERFACE
& REPORTING
2
DATA & API MANAGEMENT
57. 57Qraft Technologies, Inc.
S Y S T E M A R C H I T E C T U R E
Complete data is essential to deep learning. For this, we subscribe to various data
from several qualified data vendors such as S&P's Compustat, Thomson Reuters, and
the Federal Reserve Economic Data. In addition to structured numeric data, we also
mine non-structured text data taking advantage of natural language process engines.
Although it is one of the most fundamental and vital elements of deep learning to
secure flawless data, many financial institutions or AI technologies companies that
work on developing AI systems for financial services tend not to pay enough
attention to it. But we make hard efforts to collect high-quality data and pre-process
them for deep learning.
DATA FEED FROM
QUALIFIED SOURCES
MACRO DATA
UNSTRUCTURED DATA
S&P Compustat
Thomson Reuters
PostrageSQL
Fred
Quandl
File System
MongoDBFred API
Text
Summary
NLP Engine
MARKET DATA
58. 58Qraft Technologies, Inc.
Post-Masking
Pre-Masking
S&P 500
S Y S T E M A R C H I T E C T U R E
META DATA FOR
ACCURATE LEARNING
When simulating various investment strategies and having AI learn the simulation
results, we need vast metadata to provide an accurate historical investment universe
corresponding to each investment strategy. Survivorship bias and look-ahead bias
could hinder AI's deep learning, overstating the performance of simulated strategies.
We now use historical metadata for stocks including but not limited to:
a. GICS sectors and industries;
b. listing and delisting dates;
c. merger dates;
d. exchanges;
e. share classes;
f. announced financial data and announcement dates; and
g. revised financial data and revision dates.
* Source: Qraft Technologies, Inc.
Momentum Strategy Simulation: Pre-Masking vs. Post-Masking
-500%
0%
500%
1000%
1500%
2000%
2500%
3000%
3500%
4000%
4500%
5000%
59. 59Qraft Technologies, Inc.
S Y S T E M A R C H I T E C T U R E
DATA INTEGRATION &
API MANAGEMENT
Post-cleansing raw data are loaded onto the central database. From the raw data, the
system also derives additional data such as market beta, factor returns, and financial
ratios and saves them to the database. All data in the database can be called through
APIs and used for deep learning and daily portfolio management at later stages.
More data could be helpful to improve the learning rate of a deep learning engine,
but only if they are relevant to the other data simultaneously learned by the engine.
We continuously research to find additional raw and derived data for better learning
results.
Historical Universe
Sectors & Industries
Adj. Price & Mkt. Cap
Etc.
Universe
Mask
Monthly
Beta
3 Factors & Residual
Etc.
Tailored Data Derived Data
Kirin Port
KIRIN API
Fully Auto Installing from Qraft SVN Repo
Numba JIT, Vectorization, Key-Based Hashing
60. 60Qraft Technologies, Inc.
S Y S T E M A R C H I T E C T U R E
HIGH-SPEED PARALLEL
CALCULATION
Calculating metadata requires enormous computing resources. Our Kirin APIs are
structured parallel and can process metadata 40 times faster than conventional APIs.
It enables our researchers and AI systems to test a vast number of investment
strategies and models within a short time.
Historical Universe
Sectors & Industries
Adj. Price & Mkt. Cap
Etc.
Universe
Mask
Monthly
Beta
3 Factors & Residual
Etc.
Tailored Data Derived Data
Kirin Port
KIRIN API
Fully Auto Installing from Qraft SVN Repo
Numba JIT, Vectorization, Key-Based Hashing
61. 61Qraft Technologies, Inc.
S Y S T E M A R C H I T E C T U R E
DEEP LEARNING &
MODEL GENERATION
A deep learning engine that has a systematic goal to mimic or enhance a specific
investment strategy calls for pre-processed data from the central database through
the Kirin APIs and repeats to learn the data until finding the most optimal investment
model up to date. Automated machine learning (AutoML) runs this process
automatically.
The most promising investment model generated by the engine is used in managing
investment portfolios until the next learning date (e.g., monthly, quarterly).
AI Model Training
Portfolio Generation
PostrageSQL Back-up
Auto Reporting
DEEP LEARNING
62. 62Qraft Technologies, Inc.
S Y S T E M A R C H I T E C T U R E
All the processes a deep learning engine learns data can be monitored through web
bulletins in real-time.MONITORING
LEARNING PROGRESS
63. 63Qraft Technologies, Inc.
S Y S T E M A R C H I T E C T U R E
DAILY OPERATION OF
GENERATED MODELS
A deep learning engine learns the comparison of the actual result with the last
forecast and new data on every investment portfolio rebalancing (e.g., monthly,
quarterly). Until the next learning time, the system provides daily target investment
portfolios with the latest investment model.
After market-hours, daily data such as closing prices are newly loaded onto the
central database. Then, the system automatically generates a new target investment
portfolio for the next trading day and delivers it to an administrator through a web
user interface.
Weight Calculation
Portfolio Validation
PostrageSQL Back-up
Auto Reporting
DEEP LEARNING
64. 64Qraft Technologies, Inc.
S Y S T E M A R C H I T E C T U R E
An administrator can access the central database through a web user interface and
check data learning processes, daily operation outputs, and so on. Download in excel
is also available.
WEB USER INTERFACE
FOR END-USERS
66. 66Qraft Technologies, Inc.
E X P E C T E D E F F E C T S
Our AI-Enhanced Quantitative Investment System can streamline and automate long
and complicated current investment product development and portfolio
management processes. Also, it can enhance investment performance through
optimization.
TRANSFORMATION OF
INVESTMENT PROCESS
The As-Is Investment Product Development Process
Kick-Off Launching
SCHEME DESIGN BACK TESTING OPERATION TESTING DAILY OPERATION
Investment Idea Discussion Quantitative Validation
Product Validation
Rebalancing
Performance Check
Operation Testing
Risk Management
Weight Adjustment
Reporting
67. 67Qraft Technologies, Inc.
E X P E C T E D E F F E C T S
The system can be applied to single investment products as well as the entire
investment management organization thanks to its high expandability. All processes
are automatically recorded and reported for easy management by a human
administrator.
TRANSFORMATION OF
INVESTMENT PROCESS
The To-Be Investment Product Development Process
Kick-Off Launching
SCHEME DESIGN AI OPTIMIZATION AUTOMATED OPERATION
Investment Idea Discussion Quantitative & Product Validation
Performance Check
Rebalancing & Reporting
69. 69Qraft Technologies, Inc.
Our AI-Enhanced Quantitative Investment System can reduce all these processes to
less than a month. The time it takes to back-test an idea is short, so you can quickly
validate a considerable number of investment ideas.
And the system can take into account a large number of variables from pre-processed
big data at the same time and find the best investment strategy.
A shorter development period can significantly reduce development costs. And as the
process of product operation is automated, so additional costs can be minimal even if
the number of products increases.
E X P E C T E D E F F E C T S
TIME SAVE &
COST REDUCTION
In general, it takes at least several months to nearly a year to develop a single
investment product after back-testing an investment idea, modifying it based on the
results, and repeating back-tests.
Besides, a typical back-test can only take into account a limited number of variables
simultaneously. So even if it is an investment strategy created through many back-
tests, it could still be far from the optimal one.
Also, the back-testing results are often inaccurate due to data bias.
69Qraft Technologies, Inc.
70. 70Qraft Technologies, Inc.
E X P E C T E D E F F E C T S
INTEGRATED RISK
MANAGEMENT
Conventional investment products continue to rely on individual portfolio managers
even after a long development process. It makes it challenging to maintain their
investment strategies as initially intended.
Besides, it is not easy for them to remain effective in the ever-changing market.
Unexpected errors often occur when rebalancing, and it leads to investors' losses.
Integrated risk management is even more challenging to achieve.
70Qraft Technologies, Inc.
Our AI-Enhanced Quantitative Investment System can computerize the capabilities of
portfolio managers and systematically manage their investment strategies regardless
of any organizational changes.
And the system can monitor investment performances in real-time, analyze them in
high dimensions using big data, and provide feedback so that the validity of
investment strategies is maintained and improved regardless of market conditions.
Automated portfolio validation also helps to prevent operational errors.
71. 71Qraft Technologies, Inc.
E X P E C T E D E F F E C T S
SUPPORTS FOR
HUMAN CREATIVITY
Even after product launches, simple and iterated tasks during operation require
considerable human resources. It not only makes it more challenging to maintain
team members but also often lead to operational errors due to poor concentration.
The amount of data that can be used to make investment decisions is rapidly growing.
But it is not a challenge to resolve by putting more people into the data analysis job.
A person's ability to calculate and analyze cannot be simply added.
By automating iterations, our AI-Enhanced Quantitative Investment System can not
only help portfolio managers focus on more creative and analytical tasks but also
significantly reduce operational errors.
And through learning big data and recognizing vast patterns that humans can never
catch up, the system can effectively assist portfolio managers in their analysis.
It can also help portfolio managers make timely decisions by providing quick
feedback on product development and operational processes.
71Qraft Technologies, Inc.