=Paper=
{{Paper
|id=Vol-3276/SSS-22_FinalPaper_35
|storemode=property
|title=Change Explanation in Financial Markets by
Graph-Based Entropy and Inter-regional Interactions
|pdfUrl=https://ceur-ws.org/Vol-3276/SSS-22_FinalPaper_35.pdf
|volume=Vol-3276
|authors=Yosuke Nishikawa,Teruaki Hayashi,Takaaki
Yoshi,Toshiaki Sugie,Yoshiyuki Nakata,Kakeru
Itou,Yukio Ohsawa
}}
==Change Explanation in Financial Markets by
Graph-Based Entropy and Inter-regional Interactions==
https://ceur-ws.org/Vol-3276/SSS-22_FinalPaper_35.pdf
Change Explanation in Financial Markets by Graph-Based Entropy and
Inter-regional Interactions
Yosuke Nishikawa1*, Teruaki Hayashi1, Takaaki Yoshino2, Toshiaki Sugie2, Yoshiyuki Nakata2,
Kakeru Itou2, Yukio Ohsawa1
The University of Tokyo1, Nissay Asset Management Corporation.2
utpurihime@g.ecc.u-tokyo.ac.jp*
Abstract when an external factor affects a group of nodes. Nishikawa
Explaining the causes of changes in complex financial mar- et al. (2020) used GBE and detected changes in markets.
kets can be helpful for future investment decisions. In this
study, we focus on an anomaly in the financial market when
the price movements of the Japanese stock index (TOPIX-17) : Percentage of nodes belonging to the J cluster
and the Japanese government bond (JGB) interest rate are
Inter-regional Interactions is a linear combination of the
aligned. We analyze the changes in the financial markets
from 2019 to 2020 in a descriptive manner by creating a variability in the existence region of the data belonging to
graph with the data of the two regions of stocks and interest each region cluster, weighted by the size of the cross-re-
rates as nodes and using Graph-based Entropy (GBE) and In- gional cluster across regions, and is an indicator that in-
ter-regional Interaction. GBE is an index for detecting creases as nodes in different regions are connected.
changes in the graph’s structure, and Inter-regional Interac-
In this study, we will improve the explanatory power of
tion is the variation in the regions where the data exist. The
novelty of this study is to propose the indicator calculated GBE change detection by dividing the data into regions and
from GBE and the Inter-regional Interaction to detect analyzing them from a finer perspective.
changes in financial markets considering the characteristics
that before the price movements of stocks and interest rates
are largely aligned, and the overall entropy goes down, but : The i-th region in layer
there is no interaction between regions. The result suggests : the set of cross-regional clusters in across lower regions
that our proposed indicator can detect changes in the early
stages that have not yet gone beyond regions but have been
beginning to occur within regions, which lead to significant
changes. Experiment
Purpose and hypothesis
Introduction The purpose of this study is to detect changes in the finan-
Financial market movements are complex. To detect mar- cial markets when the price movements of Japanese stock
ket changes, explaining the state of the market at the time of indices (TOPIX-17) and Japanese government bond (JGB)
those changes can be a great help in making investment de- interest rate are aligned. Under normal conditions, the price
cisions. The problem of change detection using only time- movements are not so aligned between TOPIX-17, where
series data is difficult to explain the causes and structure of the price movements are industry specific, and JGB interest
change, and this is approached by using graphs and regions rate, where the price movements are country specific. When
in this study. an anomaly occurs, the price movements of stocks and
bonds start to match each other. In a completely abnormal
state, these price movements will be aligned in the same di-
Related Research rection. In the late stage of the anomaly, the price move-
ments will return to normal state. This study detects the
In this study, to capture changes in the relationship be-
point when the anomaly starts, no cross-regional impact yet
tween stock price indexes and interest rates, we used Graph-
(Low Inter-regional Interactions), but price movements
based Entropy (GBE) (Ohsawa 2018) and Inter-regional In-
within the region will be aligned (GBE starts to decrease).
teractions. GBE detects changes in the structure of the graph
___________________________________
In T. Kido, K. Takadama (Eds.), Proceedings of the AAAI 2022 Spring Symposium
“How Fair is Fair? Achieving Wellbeing AI”, Stanford University, Palo Alto, California,
USA, March 21–23, 2022. Copyright © 2022 for this paper by its authors. Use permitted
under Creative Commons License Attribution 4.0 International (CC BY 4.0).
98
Data
Focusing on COVID-19, we used daily data for a total of 17
industry-specific TSE stock price indices (TOPIX-17) and
six JGB interest rates (2, 5, 7, 10, 20, and 30 years) for the
period from 1/1/2019 to 11/20/2020 (494 days in total).
Graph
We created a graph for each day, which includes 23 nodes
as TOPIX-17 and six JGB interest rates. For each time series
data, we (1) divided the time series by a window width of
10 days, (2) standardized the time series within the window, Fig.2 Change detection indicators (threshold = 1.2)
and (3) calculated the distance between time series within
the window width for all combinations of nodes by DTW
(Sakoe, 1978). We edged for pairs of nodes whose distance Discussions
was closer than a uniformly set threshold. In Figure 1, A, along with the change finder, reacted to a
point in time prior to March 2020 when COVID-19 im-
Region pacted the market. This is a good indication of the major
To calculate Inter-regional Interactions., regions were set up changes that occurred in the market.
using historical data. The two regions were divided by k- In Figure 2, where the threshold was loosened, only A re-
means using the date and time data of the past five years acted in July 2019. By loosening the threshold, we were able
(2014-2018) for 23 nodes. One region consists of the 17 to capture smaller changes, and we believe that we detected
TOPIX indexes excluding real estate (hereinafter called the point in time when the market changed to the initial state
stock region) and the second is the six JGB interest rates and of abnormality that would not show up in other methods.
TOPIX real estate (hereinafter called bond region).
Conclusions
Results
In this study, we treated the stock price index and the interest
We applied our own index below. rate on government bonds as two domains, and by looking
at the interaction between the domains, we were able to de-
Eq. (3) is an indicator that captures the moment when the tect the state of change in the financial market.
state transitions from normal to early abnormal, so that the As future work, we would like to add further explanation
GBE drops while Inter-regional Interactions remains low. of the content of structural change from the graph.
We compared it with change point (Akoglu, 2010) and
change finder (Takeuchi, 2006), which are change detection
indicators. References
Ohsawa,Y. 2018. Graph-based entropy for detecting explanatory
signs of changes in market. The Revies of Socionetwork
Strategies, 12(2), 183-203.
Nishikawa, Y., Hayashi, T., Yoshino, T., Sugie, T. Takano, K.., Itou,
K., Nakata, Y., Ohsawa, Y. 2021. Analysis of the impact of
COVID-19 on financial markets by applying entropy to al-
ternative data and financial data. Special Interest Group on
Artificial Intelligence. Kumamoto, Japan.
Sakoe, H., & Chiba, S. 1978. Dynamic Programming Algorithm
Optimization for Spoken Word Recognition. IEEE Transac-
tions on Acoustics, Speech, and Signal Processing, 26, 43-49.
Akoglu, L., & Faloutsos, C. 2010. Event detection in time series
of mobile communication graphs. In Army science confer-
ence (Vol. 1).
Takeuchi, J. I., & Yamanishi, K. 2006. A unifying framework for
detecting outliers and change points from time series. IEEE
transactions on Knowledge and Data Engineering, 18(4),
Fig.1 Change detection indicators (threshold = 0.6) 482-492
99