IFTA2020 Kei Nakagawa

Innovation Lab,
Nomura Asset Management Co., Ltd.
Kei Nakagawa
Copy Right© 2020 Kei Nakagawa

Contents
1. Introduction
2. Research Examples
- Time series prediction
- Cross sectional prediction
3. Conclusion

What is AI/Machine Learning?
・Cerebral Cortex：Unsupervised Learning
・Basal Ganglia：Reinforcement Learning
・Cerebellum：Supervised Learning
Machine Learning is one of the ways we expect to achieve AI.
- With the recent development of ML, ML is used in the same way as AI.
ML is an algorithm that imitates the function of the brain.
Learning means finding a model that represents the data well.

𝑋：Feature
???Model：CNN
𝑌：Label
In case of image recognition, the goal is to identify the image
… Problem
Formulation
… Data
𝑓
In supervised learning, output is called label and input is called feature.
Supervised learning learn the relationship between inputs and outputs.
Cat
Deciding what model, features, and labels to use is called
problem formulation.

Model：RNN
In case of translation, the goal is to translate English into Japanese
… Problem
Formulation
… Data
This is a pen.
???
This is a pen in Japanese.
𝑓
𝑋：Feature 𝑌：Label
これはペンです

Model：ML
In case of finance, the goal is to predict future stock prices.
… Problem
Formulation
… Data
???
+10%𝑓
𝑋：Feature 𝑌：Label

A. Firstly, we need to decide what data we use and set up an
appropriate problem formulation.
Problem Formulation
• Time Series
• Cross Section
Data
• Technical
• Fundamental
• Alternative
AI/Machine Learning
How to predict stock prices w/ AI and ML?

Problem Formulation
クロスセクションで予測
Stock A Stock B ・・・
Today
Yesterday
・
・
・
・
Buystockswithabsolutely
highreturns. Buy stocks with relatively high returns.
Investment Universe
TimeSeries
Cross SectionOrthogonal

Problem Formulation：Time Series
Past (known) data at time t
Predict a stock future price/return as time series data.
・ 𝑟𝑡+1: (Absolute) return at time t+1 (Label)
・ 𝑟𝑡+1 = 𝑓(𝑟𝑡, … , 𝑥 𝑡, … )
・ 𝑟𝑡, 𝑥 𝑡: Past returns and external variables （Feature）
Short term forecast (shorter than monthly) → trading strategy
Problem Formulation: Learn the following model 𝑓

Problem Formulation：Cross Section
Predict multiple stocks that are relatively better in a universe.
・ 𝑠𝑐𝑜𝑟𝑒𝑖: (Relative) return of each stock in the universe (Label)
・ 𝑠𝑐𝑜𝑟𝑒𝑖 = 𝑓(𝑿𝒊)
・ 𝑿𝒊: Criteria for describing the score: Factor (Feature)
Long term forecast (longer than monthly) → portfolio construction
Time information is omitted
Problem Formulation: Learn the following model 𝑓

Investment
Horizon
Minutes Hour Day Week Month Year
Problem
Formulation
Time Series
Cross Section
Data Technical
Fundamental
Alternative Micro(Corporate)→Macro
News、PoS、etc,…
AI/Machine Learning
Problem Formulation：Summary

Time Series × Technical × Supervised
“Stock Price Prediction with Fluctuation Patterns Using Indexing
Dynamic Time Warping and k∗-Nearest Neighbors”.
Nakagawa et.al. (2018)
New Frontiers in Artificial Intelligence
https://doi.org/10.1007/978-3-319-93794-6_7

Motivation
Following nearest trend Following similar trend
Momentum Patterns
Predict the future price using the past price fluctuation patterns.
Refine the “formation analysis” by machine learning.
Nakagawa et.al. (2018) “Stock Price Prediction with Fluctuation Patterns Using
Indexing Dynamic Time Warping and k∗-Nearest Neighbors”.
New Frontiers in Artificial Intelligence
https://doi.org/10.1007/978-3-319-93794-6_7

Method - Outline
Current pattern
Umm...
Similar!!
1. Measure 2. Predict
・How to measure the similarity?
Indexing Dynamic Time Warping（IDTW）
・How to predict with extracted fluctuations?
k*-Nearest Neighbors （k*-NN）
We propose a method to predict future stock prices based on
the shape of the past price fluctuations similar to the current.

frequencymeasuring period
1. Decide the measuring period and frequency.
2. Consider the difference of price levels.
Near?
Distance
3. Define the distance between time-series.
✔ Daily closing prices within a month.
✔ Dynamic Time Warping (DTW) method
How to measure the similarity between stock price movements?
✔ We index prices and adjust price levels.
Method - IDTW

●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Correspondence
With Euclidean Distance
1
n
n
Sequences are aligned “one to one”.
●
●
● ● ●
●
●
●
●
●
● ●
●
●
●
●
●
● ●
Correspondence
With DTW
1
m
n
Nonlinear alignments are possible.
Fixed Time Axis “Warped” Time Axis
DTW is a method that warps the time axis to find the optimum
correspondence and calculates the similarity between two-time series.
What is DTW??

Why do we use DTW??
・ DTW can be applied to a pair of time-series of different lengths.
・ Similarity of DTW fits human intuition.
・ DTW is a standard method to measure the similarity among
various time series data (UCR Datasets)
What is DTW??
https://cran.r-project.org/web/packages/dtw/index.html
R language package dtw

Method - Outline
Current pattern
Umm...
Similar!!
1. Measure 2. Predict
・How to measure the similarity?
Indexing Dynamic Time Warping（IDTW）
・How to predict with extracted fluctuations?
k*-Nearest Neighbors （k*-NN）
We propose a method to predict future stock prices based on
past price fluctuations similar to the current.

Method - k-NN
✔ 𝑘-NN algorithm is one of the most fundamental non-parametric methods.
Similarity
𝑘 = 1 𝑘 = 3
Prediction:
+2.5%
Prediction:
(𝛼12.5 + 𝛼21.1 − 𝛼35.1)%
Return on
Next month
(Labels)
…
+2.5%
𝑦1
+1.1%
𝑦2
−5.1%
𝑦3
−3.2%
𝑦4
0.15𝛽1 0.24𝛽2 0.41𝛽3 0.63𝛽4 𝛽𝑖 = 𝐼𝐷𝑇𝑊 𝑥0, 𝑥𝑖
Fluctuation
pattern
(Features)
𝑥0
𝑥1 𝑥2 𝑥3 𝑥4
？%
✔ Prediction is the average of the values of 𝑘 nearest neighbors.

Method - k*-NN
✔ 𝑘-NN algorithm is one of the most fundamental non-parametric methods.
𝑘 = 1 𝑘 = 3
Prediction:
+2.5%
Prediction:
(𝛼12.5 + 𝛼21.1 − 𝛼35.1)%
…
+2.5%
𝑦1
+1.1%
𝑦2
−5.1%
𝑦3
−3.2%
𝑦4
？%
✔ Prediction is the average of the values of 𝑘 nearest neighbors.
1. How many k should we use to predict?
2. How to decide the optimal weight alpha?
𝑘∗
-NN is a simple approach for calculating optimal 𝒌 and weight 𝜶.
https://cran.r-project.org/web/packages/ksNN/index.html
R language package ksNN

Empirical Study
✔ To demonstrate the advantages of IDTW+k*NN,
we analyze its performance using following major world stock indices.
✔ In-sample period : January 1989 ～ December 2005.
Out-of-sample period : January 2006 ～ August 2017.
✔ Compared six methods combining DTW, DDTW, or IDTW with k-NN or k*-NN.
Ticker Name Country Constituents
TPX TOPIX Japan Companies listed on the First Section of the Tokyo Stock Exchange
SPX S&P500 U.S. 500 large companies listed on the NYSE or NASDAQ
CAC CAC40 France 40 companies listed on the Euronext Paris
DAX DAX German 30 companies trading on the Frankfurt Stock Exchange
UKX FTSE100 U.K. 100 companies listed on the London Stock Exchange

Empirical Study
Step 1. Calculate the similarity between month 𝑡 and all months from 1 to 𝑡 − 1
with DTW, DDTW, or IDTW.
𝑡
1 𝑡 − 1
Calculate each similarity
t－3 t－2 t－1 t21
DTW
DDTW
or
IDTW
or

Empirical Study
Step 2. Predict the next month's return using k-NN or k*-NN.
6
120
t－2
27
11
Sorting
by distance
+1.5%
0.02
0.03
0.04
7.99
8.82
+1.9%
-0.5%
+3.0%
-2.7%
𝛽
𝑦Next month’s
returnsdistances
k-NN
k*-NN
Predict
or
+𝟐. 𝟔%

Empirical Study
Step 3. If the predicted next month's return is positive, buy one unit of Index at
the end of the month. If negative, sell one unit and calculate the return.
Step 4. Proceed to the next month: 𝑡 = 𝑡 + 1.
𝑡
t t+1
𝑡 + 1
Index Price
(Month End)
100 105
Predict
+𝟐. 𝟔%
+𝟓. 𝟎%Actual

Results
✔ Cumulative returns for the six methods in TOPIX.
The out-of-sample period is from January 2006 to August 2017.
Red Line:Our method
(IDTW+k*NN)
Blue Line:
Buy & Hold Index

Results
Table 1. The average accuracy of all years for each method.
Table 2. The total return of each method.

Results
Table ３. Comparison with Momentum Strategy(Return)
Table ４. Correlation with Momentum Strategy
Here, Prediction in Step 2 is replaced by simple 1Mom: 1 month return and
12 Mom: 12 month return excluding the last month (12 Mom).

Summary
✔ We proposed k*-NN with IDTW, a highly accurate stock prediction
method. This method is inspired by formation analysis.
✔ An empirical analysis was conducted with major world indices and
confirmed the following results:
・ IDTW is superior to DTW and DDTW
・ k*-NN is superior to k-NN.
・ IDTW-k*NN is the best prediction method.

Time Series × Technical × Unsupervised
“Stock price prediction using k‐medoids clustering with
indexing dynamic time warping”.
Nakagawa et.al. (2019)
Electronics and Communications in Japan, 102(2), 3-8.
https://doi.org/10.1002/ecj.12140

Motivation
Price movements
that are likely to rise？
✔ Can we visualize good patterns for forecasting by clustering price
fluctuation patterns?
Price movements
that are likely to fall？
✔ So how should we cluster the price fluctuation patterns?
Nakagawa et.al. (2019). “Stock price prediction using k‐medoids clustering with
indexing dynamic time warping”.
Electronics and Communications in Japan, 102(2), 3-8.

・ Always looks for medoids in the data (Needs only similarity).
✔ Since the similarity was given by IDTW, we should decide the clustering
method.
k-means method
k-medoids method
How to Cluster Price Fluctuation?
・ Most well-used clustering method
・ Clustering data based on centroid
・ Easy to extent (x-means, k-means++)
・ Clustering data based on medoids
(Data point closest to any data point in the cluster)

2007/1
2006/12
Clustering with k-Medoids
2004/5
Step1: Initialize medoids
(𝑘: number of medoids)
✔ Randomly select 𝑘 = 3 medoids in the data

2007/1
2006/12
2004/5
IDTW
Step2: Clustering
✔ Assign each data point to the cluster
to which the closest medoid belong
Step1: Initialize medoids
(𝑘: number of medoids)

2007/1
2006/12
Step3: Update medoids
✔ recalculate and update
medoids in each cluster
2004/5
Step2: ClusteringStep1: Initialize medoids
(𝑘: number of medoids) Repeat until the clusters
do not change

Empirical Study
✔ Using the TOPIX, we cluster similar price fluctuation patterns and visualize
good patterns for forecasting.
✔ In-sample period : January 1989 ～ December 2005.
Out-of-sample period : January 2007 ～ March 2017.
✔ IDTW with k-medoids is compared to DTW with k-medoids
✔ The number of k are from 2 to 12.
To decide k, we evaluate profitability of the following virtual trading.

k-medoids is performed based on
monthly daily price fluctuations up to
the previous month.
Identify the cluster for
the current month.
If there is more positive
(negative) in the cluster,
buy (sell).
Step 2:
Identify the cluster
Step 3:
Decide buy or sell
Step1:
Clustering
Empirical Study

✔ From the viewpoint of profitability, the number of k is 𝟓
Table 5. Profitability and accuracy
by number of clusters
Results

✔ Visualization of 5 clusters (black line: Medoids, x axis: days, y axis: return)
Reversal Momentum Momentum
Results
*The percentage in parentheses indicates the probability that stock price rises in the next month.
*

FlatReversal?
Results
✔ Visualization of 5 clusters (black line: Medoids, x axis: days, y axis: return)

✔ Using k-Medoids + IDTW, we found that the number of clusters in
the price fluctuation pattern was about 5 in TOPIX.
✔ Consider price fluctuation pattern clustering method (k-Medoids +
IDTW). This is unsupervised approach.
Summary
✔ Large fluctuations are tend to continue.

Time Series × Technical × Supervised
“Time-Series Gradient Boosting Tree for Stock Price Prediction”.
Nakagawa and Yoshida (2020)
International Journal of Data Mining, Modelling and Management
(IJDMMM)
To be appear
Fundamental

Motivation
✔ We improve a time series decision tree (TSDT) that uses only time
series data.
Nakagawa and Yoshida (2020)
“Time Series Gradient Booting Tree for Stock Price Prediction”.
International Journal of Data Mining, Modelling and Management (To be appear)
(Time series) decision trees
classify the (time series) data
by sorting them down the tree
and leaves provides the labels.
Decision tree is
easy to understand!
Current

Motivation
We propose a time-series gradient boosting tree that uses a TSDT
that adds features other than time-series.
(1) Add usual features other
than time series to TSDT
(2) “Boost” this TSDT with
gradient boosting technique

Method – Time Series Gradient Boosting Tree
✔ Boosting is one of the methods called ensemble learning.
✔ Gradient boosting tree is most used in various data analysis contests (Kaggle)
and has the best results.
strong model
weak model
Ensemble learning is a method of building a strong (accurate) model by
using multiple models called weak learners.

Method – Time Series Gradient Boosting Tree
Fit a new tree to the error
Prettenhofer and Louppe(2014)
Illustration of gradient boosting tree

Empirical Study
✔ To demonstrate the advantages of time series gradient boosting tree,
we analyze its performance using following major world stock indices.
✔ In-sample period : 10 years.
Out-of-sample period : 10 years later until 2018/6
Ticker Name Country Data period
TPX TOPIX Japan 1993/8～2018/6
SPX S&P500 U.S. 1990/1~2018/6
CAC CAC40 France 2001/6~2018/6
DAX DAX German 2001/7~2018/6
UKX FTSE100 U.K. 2000/1~2018/6

Empirical Study
✔ Feature:
Other Data ：As of the end of month in each index, we use
✔ Model： Compared four methods combining Time Series Decision Tree
(TSDT) and Time Series Gradient Boosting Tree (TSGBT) with and
without index data.
Data Description
PER Net Profit/Market Value
PBR Net Asset/Market Value
MOM Stock returns over the past 12 months except for past month
DIV Dividends paid in a year / Market Value
ROE Net income/Net Assets
Time Series Data：Daily price data within a month
✔ Teacher：1 month ahead return

Results
Table 6. The average accuracy of all years for each method.
Table 7. The total return for each method.
Improvement
by adding features
Improvement
by boosting

Summary
✔ We build a time series decision tree (TSDT) by introducing features
other than time series data.
・Adding features other than time series data improves accuracy and profitability.
・Gradient boosting also improves accuracy and profitability
✔ Furthermore, we propose a time series gradient boosting tree
(TSGBT) that “boosts” time-series decision tree.
✔ An empirical analysis was conducted with major
world indices and confirmed the following results:

1. Introduction
2. Research Examples
- Time series prediction
- Cross sectional prediction
3. Conclusion
Contents

Cross Section × Technical × Supervised
“Deep Factor Model”.
Nakagawa et.al. (2018) 3rd Workshop on MIDAS@ECML-PKDD 2018
([preprint] arXiv:1810.01278)
“Cross-sectional Stock Price Prediction using Deep Learning for
Actual Investment Management”
Abe and Nakagawa (2020) AIBC2020
([preprint] :arXiv:2002.06975)
“Deep Learning for Multi-factor Models in Regional and Global Stock
Markets”
Abe and Nakagawa (2020) New Frontier in AI
https://link.springer.com/chapter/10.1007/978-3-030-58790-1_6
Fundamental

Cross Sectional Prediction
✔ Over 300 factors until 2012 [Harvey et al. 2017]
✔ In practice, we predict future relative stock returns (scores) by
combining various factors
✔ Criteria for describing the score is called Factor in Finance
For example, Market Value is one of the most famous factors, and this factor combined with
simple sorting portfolio yields a positive return.
Cross-sectional Investment Strategy
✔ finds stocks with relatively high scores based on certain criteria.

ROE
1 Month Return
●
●
●
Score
Value
Growth
Quality
Momentum
・Linear Relationship
Factor
Candidates
Factor Classification
by human
Calculate
Relative Score
Both buying and selling are
evaluated equally.
Factors are evaluated monotonically.

ROE
1 Month Return
●
●
●
Score
Value
Growth
Quality
Momentum
・Linear Regression
Factor
Candidates
Factor Classification
by human
Calculate
Relative Score
Both buying and selling are
evaluated equally.
Factors are evaluated monotonically.
I would like to capture the non-
linear relationship of factors from a
large number of candidates by
deep learning.

Cross-sectional Investment Strategy w/ DL
ROE
1 Month Return
●
●
●
Factor
Candidates
・
・
・
・
・
・
・
・
・
・
・
・
・
・
・
Extraction of effective factors in deep
learning (without human judgment)
Score
・Non-Linear Relationship

「Deep Factor Model」
Kei Nakagawa, Takumi Uchida and Tomohisa Aoshima,
3rd Workshop on MIDAS@ECML-PKDD 2018 ([preprint] arXiv:1810.01278)
✔ We check the profitability of cross-sectional investment strategy
with deep learning
✔ We predict relative return using 80 factors and invest monthly.
✔ Investment Universe:
TOPIX index constituents (from Apr 2006 to Mar 2016 monthly)
Empirical Study No.1

✔ We use the 16 factors that are used relatively often in practice.
Factor Candidates Formula
Risk
60VOL Standard deviation of stock returns in the past 60 months
BETA Regression coefficient of stock returns and market risk premium
SKEW Skewness of stock returns in the past 60 months
Quality
ROE Net income/Net Assets
ROA Operating Profit/Total Assets
ACCRUALS Operating Cashflow ‐ Operating Profit
LEVERAGE Total Liabilities / Total Assets
Momentum
12-1MOM Stock returns in the past 12 months except for past month
1MOM Stock returns in the past month
60MOM Stock returns in the past 60 months
Value
PSR Sales/Market Value
PER Net Income / Market Value
PBR Net Assets / Market Value
PCFR Operating Cashflow / Market Value
Size
CAP log(Market Value)
ILLIQ average(Stock Returns/Trading Volume)

✔ Use 16 x 5 = 80 features for the last 5 points each quarter of 16 factors.
✔ We Learn the model using the training data for 120 sets.
✔ Training data (1 set)
・Feature: 80 features of 16 factors
・Teacher: One month ahead relative returns

Model Description
Deep Learning(Deep 1) We use full-connected network. The hidden layers are (80-50-
10),(80-80-50-50-10-10). We use the ReLU as the activation
function and Adam for the optimization algorithm.Deep Learning(Deep 2)
Linear Model
(Linear Regression;LR)
“sklearn.linear_model.LinearRegression”
Support Vector
Regression(SVR)
“sklearn.svm.SVR”
Random Forest(RF) “sklearn.ensemble.RandomForestRegressor”
✔ We use linear and nonlinear models as benchmarks.
✔ We make a long/short portfolio strategy for a net-zero investment to buy
top stocks and to sell bottom stocks with equal weighting in quintile portfolios.
✔ For the quintile portfolio performance, we calculate the
annualized average return, and Sharpe ratio.

Results No.1
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
Deep 1 Deep 2 LR SVR RF
Return✔ This graph shows the (annualized) returns of each methods
The results of deep learning are better
than other nonlinear models.
The linear model is not bad either.

Results No.1
✔ This graph shows the Sharpe Ratio of each methods
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40
1.60
Deep 1 Deep 2 LR SVR RF
ShapeRatio
The results of deep learning are better
than other nonlinear models.
The linear model is not bad either.
Linear model is good if a poor
nonlinear model is used.

✔ We check the profitability of cross-sectional investment
strategy with deep learning
✔ We predict relative return using 33 factors and invest daily basis.
Cross-sectional Stock Price Prediction using Deep Learning for Actual Investment Management
Masaya Abe and Kei Nakagawa,
AIBC2020 ([preprint] arXiv: 2002.06975)
TOPIX500 index constituents (from Apr 2013 to Mar 2017 daily)

No. Factor
1 Return from previous day
2 Return from 2 days ago
9 Average trading volume over the past 60 days
10 Average trading volume over the past 5 days/60 days
13 Change in operating income forecast from 5 days ago
16 Change in target stock price forecast from 5 days ago
No. Factor
19 Book-value to Price Ratio
20 Earnings to Price Ratio
21 Dividend Yield
22 Sales to Price Ratio
23 Cashflow to Price Ratio
24 Return on Equity
25 Return on Asset
26 Return on Invested Capital
27 Accruals
28 Total Asset Turnover Rate
29 Current Ratio
30 Equity Ratio
31 Total Asset Growth Rate
32 Capital Expenditure Growth Rate
33 Investment to Asset
- -

Model Hidden Layer (Dropout) Epoch
DNN_1
500 – 200 – 100 – 50 – 10
(50% – 40% – 30% – 20% – 10%）
20
DNN_2
500 – 200 – 100 – 50 – 10
(50% – 40% – 30% – 20% – 10%）
30
DNN_3
200 – 200 – 100 – 100 – 50
(50% – 50% – 30% –30% – 10%）
20
DNN_4
200 – 200 – 100 – 100 – 50
(50% – 50% – 30% – 30% – 10%）
30
DNN_5
300 – 300 – 150 – 150 – 50
(50% – 50% – 30% – 30% – 10%）
20
DNN_6
300 – 300 – 150 – 150 – 50
(50% – 50% – 30% – 30% – 10%）
30
・ Deep Learning ・ Random Forest (RF)
Model # of feature # of tree depth
RF_1 11 1000 3
RF_2 11 1000 5
RF_3 11 1000 7
model regularization
RR_1 0.1
RR_2 1.0
RR_3 10.0
・ Linear Model (Ridge Regression;RR)
✔ We make a quintile long/short portfolio and
calculate the annualized average return, and Sharpe ratio.

For the epoch number, 20 is good.
In this study, 30 is overfitting.
✔ This graph shows the (annualized) returns of each methods
0.00%
2.00%
4.00%
6.00%
8.00%
10.00%
12.00%
Return
For the level of return,
the linear models(RR) are very good.

0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
ShapeRatio
For the epoch number, 20 is good.
In this study, 30 is overfitting.
For the Shape ratio,
DNNs are the very good.

✔ We check the profitability of cross-sectional investment
strategy with deep learning
✔ We predict relative return using 20 factors and invest monthly.
Deep Learning for Multi-factor Models in Regional and Global Stock Markets
Masaya Abe and Kei Nakagawa,
New Frontiers in Artificial Intelligence (To be appear)
World ： MSCI WORLD
North America： MSCI North Americas
Europe： MSCI Europe & Middle East
Asia Pacific： MSCI Pacific
(from 2005 to 2017 monthly)

No. Factor No. Factor No. Factor
1 Book-to-market ratio 8 Return on invested capital 15 EPS Revision(1 month)
2 Earnings-to-price ratio 9 Accruals 16 EPS Revision(3 months)
3 Dividend yield 10 Total Asset Growth Rate 17 Past stock return(1 month)
4 Sales-to-price ratio 11 Current ratio 18 Past stock return(12 months)
5 Cash flow-to-price ratio 12 Equity ratio 19 Volatility
6 Return on equity 13 Total Asset Turnover Rate 20 Skewness
7 Return on asset 14 CAPEX Growth Rate

Model Description
Deep Learning
We use full-connected network. The hidden layers are (150-
150-100-100-50-50). Dropout rate: (50% – 50% – 30% –
30% – 10% – 10%).We use the ReLU as the activation function
and Adam for the optimization algorithm.
Linear Model
(Ridge Regression)
“sklearn.linear_model.Ridge”
Support Vector Regression “sklearn.svm.SVR”
Random Forest “sklearn.ensemble.RandomForestRegressor”
Gradient Boosting Tree “XGBRegressor”
✔ We make a quintile long/short portfolio and
calculate the annualized average return, and Sharpe ratio.

Results No.3
✔ This graph shows the (annualized) returns of each methods
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00
DNN
RR
XGB
RF
DNN
RR
XGB
RF
DNN
RR
XGB
RF
DNN
RR
XGB
RF
NA EU PF WLD
Return
✓ DNN generally outperforms other
methods in all regions.
✓ The performance decreases in the
order of PF, EU, NA regardless of
the prediction model.

Results No.3
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40 DNN
RR
XGB
RF
DNN
RR
XGB
RF
DNN
RR
XGB
RF
DNN
RR
XGB
RF
NA EU PF WLD
ShapeRatio
✓ The highest Shape ratios are given
by DNN in all regions.
✓ The performance decreases in the
order of PF, EU, NA regardless of
the prediction model.

Summary
DNN is generally superior to GB, RF, and RR in Japan and global markets
→ DNN shows promise as a skillful machine learning method in the cross section.
The performance deteriorates in the order of Pacific,
Europe, North America regardless of the prediction model.
→ The higher the market efficiency, the more difficult to earn returns

Problem Formulation
• Time Series
• Cross Section
Data
• Technical
• Fundamental
• Alternative
AI/Machine Learning
Conclusion
If you have any question or comment, please feel free to Email me
kei.nak.0315@gmail.com

Disclaimer
The Information in this presentation is provided for education and informational purposes only, without any
express or implied warranty of any kind, including warranties of accuracy, completeness, or fitness for any
particular purpose. The Information contained in or provided from or through this material is not intended to be
and does not constitute financial advice, investment advice, trading advice or any other advice. The Information
in this material is general in nature and is not specific to you the viewer or anyone else.
If this document contains any information relating to the past performance of any securities or other investment
strategies and products, you should note that past performance is not indicative of future results. Any market or
investment view mentioned in this document is not intended to be investment research. Information contained in
this document is subject to change without any obligation on our part to notify you of any change. If this
document forms part of a presentation or is presented to you together with other documents and materials, this
document should not be read in isolation and may not provide a full explanation of all of the topics presented and
discussed. You should note that investment involves risk. The price of any securities or other investment product
may go down as well as up. The information does not accept any liability or legal responsibility whatsoever for
any decisions taken based upon the material.

∗ 𝑑(𝑥, 𝑦) = 𝑥 − 𝑦
∗
79
▷ Indexing
▷ DTW
Appendix Algorithm of IDTW

81
▷ Step 1
▷ Step 2
▷ Step 3
Appendix Algorithm of k-medoids with IDTW

82
▷ Time Series
▷ Data
Appendix Algorithm of TSDT with usual data

83
Appendix Algorithm of TSGBT with usual data

IFTA2020 Kei Nakagawa

More Related Content

What's hot (20)

Similar to IFTA2020 Kei Nakagawa (20)

More from Kei Nakagawa (20)

Recently uploaded (20)

IFTA2020 Kei Nakagawa