STM2016
Technical Program
Location: Conference Room I (2nd
floor), ISM
July 20 (Wednesday)
Time |
9:15 ~ 9:30am |
|
Opening |
Morning: Tutorial: Heavy Tailed
Time |
9:30 ~ 10:30am |
Speaker |
Prof. Nourddine Azzaoui (Université
Blaise Pascal) |
Title |
Construction of some special classes of stable processes that
generalizes spatial or temporal Gaussian processes |
Abstract |
In this talk, we first begin by detailing the motivation behind the
construction of these special classes. Indeed, the main incitation is
inspired from the second order case where characterizing spatial or temporal
processes by a family of sufficient functions, such as via covariance and mean
functions, forms the basis of a large number of statistical modelling
approaches. For instance, recently in the active area of Gaussian process
regression modelling, the mean and covariance function specify uniquely the
properties of the resulting statistical model. In order to allow for
non-stationarity, non-independent increments and heavy-tailedness, we present
a novel covariation spectral representation of some non-stationary and
non-independent increments symmetric a-stable
processes (SaS). Such a representation is based on a weaker
covariation pseudo additivity condition and should allow a very wide class of
statistical regression models to be subsequently developed. In this talk we
focus especially on the construction of some examples of these processes; We
present some concrete examples that allow general constructive
approaches to building models satisfying these pseudo additivity conditions. |
Time |
10:30 ~ 11:30am |
Speaker |
Dr. Malcom Egan (Université Blaise
Pascal) |
Title |
Simulation of a general class of stable processes with infinite variance |
Abstract |
Applications in engineering and finance are driving the development of non-Gaussian models that capture features such as heavy tails and tail dependence. In this talk, we consider one such class of these models based on symmetric alpha-stable processes characterized by bi-additive covariation, which has been introduced in earlier talks by Nourddine Azzaoui and Gareth Peters. In these talks, the process was constructed using the spectral measure representation. In contrast, we focus on the alternative bimeasure representation, where a key challenge is to characterize the class of valid bimeasures. In particular, we show how to construct valid bimeasures and simulate skeletons of the process. We provide some examples to illustrate key features. |
Time |
13:00 ~ 14:30pm |
Speaker |
Dr. Gareth W. Peters (UCL) |
Title |
New Functional and Matrix Variate Regression Approaches for Spatial
and Temporal Settings |
Abstract |
In this talk two new approaches to statistical regression modelling
will be discussed. The first involves developing new regression methods for
quantile function dynamics under a Symbolic Data Analysis (SDA) formalism.
The second approach is a joint temporal panel regression combining dynamic
basis function state space models with covariance regressions. New estimation
and calibration approaches will be developed. Then two applications of these
ideas will be explored, the first in quantile dynamics that link intra-daily
volatility modelling with daily volatility. The second example will involve
stress testing methodology for multiple yield curve models, with an
illustration of the Euro-region Gilts based on credit quality and liquidity
factors. |
(Slides)* (Video)* *after publishing |
Time |
14:30 ~ 15:00am |
Speaker |
Prof. Jennifer S. K. Chan (University of
Sydney) |
Title |
Quantile regression for conditional autoregressive range model |
Abstract |
To calculate value-at-risk (VaR) for risk management, we derive parametric quantile functions. The general technique is to first build a mean regression model and then estimate families of conditional quantile functions based on the mean regression model. Instead, we propose to regress directly on the quantiles of a distribution and demonstrate the method through the conditional autoregressive range (CARR) model which has increased popularity recently. Two flexible distribution families: the generalized beta type two on positive support and the generalized-t on real support are adopted for demonstration. Then, the models are extended to model the volatility dynamic and compared in terms of goodness-of-fit. The models are implemented using the module fminsearch in Matlab under the classical likelihood approach and applied to analyse the intra-day high-low price ranges from the All Ordinaries index for the Australian stock market to obtain value-at-risk forecasts. VaR are forecast using the proposed models. |
Time |
15:15 ~ 16:15pm |
Speaker |
Prof. Mario Wuthrich (ETH Zurich) |
Title |
Tutorial on Cash Flow Valuation |
Abstract |
We revisit the theory of cash flow valuation under the paradigm of no-arbitrage. For this purpose we introduce state-price deflators which can be viewed as stochastic discount factors. These state-price deflators provide a consistent valuation framework which leads to a financial market that is free of arbitrage. We describe how these state-price deflators can be constructed, and we show how they are related to equivalent martingale measures and risk neutral pricing. Moreover, we provide several applications including the pricing of zero-coupon bonds, defaultable coupon bonds or NatCat bonds. |
(Slides)* (Video)* *after publishing |
Time |
16:15 ~ 17:15pm |
Speaker |
Dr. Andrea Macrina (UCL) |
Title |
Multi-Curve
Discounting |
Abstract |
We present discounting
models for the pricing and hedging of financial and insurance instruments in
advanced markets and show how the proposed framework extends naturally to
asset valuation in emerging (and other developing) markets. The aim is to
“catch up” with the requirements of emerging markets when it comes to
suitability and usefulness of asset pricing and hedging models, which often
have not been developed with the typical properties of emerging markets in
mind. |
Time |
18:20 ~ 21:30pm |
|
Dinner &
discussion |
July 21 (Thursday)
Time |
8:30 ~ 9:30pm |
Speaker |
Prof. Yoshiki Yamagata, Dr. Daisuke
Murakami (National Institute for Environmental Studies) |
Title |
Climate change
risk management: CO2 and urban heat monitoring & modeling |
Abstract |
This paper
overview the studies regarding the climate change risk managements including
CO2 emission reduction and adaptation to climate change impacts. The “Paris
Agreement” last year requires countries to do massive emission reduction
efforts in the future. However, it is projected that the climate change
impact is un avoidable and we need to prepare for the extreme events such as
flooding and heat wave risks. Focusing on the GCP related project activities,
we report about the new research on the land use scenarios for adapting to
flooding risks, global urban carbon mapping and hear risk assessments using
rapidly developing big-data. These studies show important opportunities of
collaborations between the environmental researchers and the spatial and
temporal statistic researchers. |
Slides |
Time |
9:30 ~ 10:30pm |
Speaker |
Ms. Dorota
Toczydlowska (UCL) |
Title |
Regression Models for Yield Curves |
Abstract |
In
this study we develop a novel class of dimension reduction stochastic
multi-factor panel regression based state-space models for the modelling of
dynamic yield curves. This includes consideration of Principle Component
Analysis (PCA), Independent Component Analysis (ICA) in both vector and
function space settings. We embed these rank reduction formulations of
multi-curve interest rate dynamics into a stochastic representation for
state-space models and compare the results to classical multi-factor dynamic
Nelson-Seigal state-space models. We introduce new class of filter to study
these models and we also introduce new approaches to the extraction of the
independent components via generalisations of the classical ICA methods. This
leads to important new representations of yield curve models that can be
practically important for addressing questions of financial stress testing
and monetary policy interventions. We will illustrate our results on data for
multi-curve Libor interest rates in multiple countries. |
Time |
10:30 ~ 11:30am |
Speaker |
Prof. Guillaume Bagnarosa (ESC Rennes
School of Business) |
Title |
From Local to
Global Credit Risk exposure among Farmers |
Abstract |
We propose in
this paper a bottom up model of the credit risk associated to agricultural
business taking into consideration the impact of weather conditions on the
expected and unexpected losses of a loans portfolio. In order to build and
calibrate this model, a large French fertilizer company gives us access to
its highly detailed database. Combining these financial and farming data
about a set of Romanian farmers with weather and yields information we
managed to estimate what we call a local default risk, associated to the
region weather conditions uncertainty, and a global default risk, which is
rather related to the prices of the internationally traded agricultural
commodities. (This is a joint
work with Matthew Ames, Suikai Gao, Donatien Hainaut, Gareth W. Peters and
Tomoko Matsui.) |
Slides Video |
Time |
11:30 ~ 12:30am |
Speaker |
Mr. Matthew Ames (UCL) |
Title |
Oil Futures
Price Risk Factors: Short-Term, Medium-Term and Long-Term |
Abstract |
In this talk, we
investigate the influence of observable covariates, such as production,
inventory, freight costs, speculative positions etc., on the term structure
of oil futures prices. We would like to identify which covariates are
significant in explaining oil futures prices in the short-term, the
medium-term and the long-term. To this end we explore the following
approaches: introducing the instantaneous covariate value into the
observation equation; linking the instantaneous covariate value to parameters
in the spot price dynamics; exploring lagged covariates. A deeper
understanding of the driving covariates behind the dynamics of futures prices
is particularly useful for derivatives pricing, real options analysis, risk
management and speculative trading. |
Slides Video |
Time |
14:30 ~ 15:30pm |
Speaker |
Prof. Shiyong Zhou (Peking University) |
Title |
Description of
The Earthquake Hazard Insurance Frame in China |
Abstract |
|
Slides Video |
Time |
15:30 ~ 16:00pm |
Speaker |
Prof. Norikazu Ikoma (Nippon Institute of
Technology) and Mr. Hiromu Hasegawa (MegaChips Corporation, Japan) |
Title |
Likelihood with
Self-Organizing Map and Gaussian distributions in Hidden Markov Model for Human
Moving Behavior Estimation from 3-axis Accelerometer Signal |
Abstract |
Aimed at
practical estimation of human moving behavior by smartphone like appliances
with low computational power and small memory size, simple yet efficient
methodology using Self-Organizing Map (SOM) and Gaussian distributions as likelihood
in Hidden Markov Model has been proposed. Source signal obtained from 3-axis accelerometer
has been pre-processed by taking its length followed by Wavelet transform in
order to obtain temporal frequency factor and power factor. SOM performs
likelihood of the frequency factor while Gaussian distributions perform
likelihood of the power factor. Unsupervised learning of SOM generates
learned weights of the SOM, which will be used at the other parts’ SOMs in
the methodology. Supervised learning with labelled data set provides template
output maps of the SOM and Gaussian distributions. The template output maps
are masked mean of the SOM output for each category of the labelled data set,
where output neurons having large variance have been masked. Taking inner
products of the SOM output for current input signal with template SOM outputs
provide likelihood values for Hidden Markov Model (HMM), which constrains
behavior temporal changes via state transition probabilities. Experiments
with real smartphone data during actual commuting and labelled date collected
by co-author will show performance of the proposed methodology in comparison
with SOM only, SOM and power factor, and in addition with HMM. |
Time |
16:00 ~ 16:30pm |
Speaker |
Ms. Xina Cheng (Waseda University) |
Title |
Anti-occlusion
Observation Model and Automatic Recovery for Multi-view Ball Tracking in
Sports Analysis |
Abstract |
The 3D position
of the ball plays a crucial role in professional sport analysis. In ball
sports, tracking of ball’s precise position accurately is highly required,
whose performance is affected by inaccurate 3D coordinate and occlusion
problem. In this paper, we propose anti-occlusion observation model and
automatic recovery by 3D ball detection based on multi- view videos to track
the ball in 3D space accurately. The anti-occlusion observation model
evaluates each camera’s image and eliminates the influence of the camera in
which the ball is occluded. The automatic recovery method detects the ball’s
3D position by homography relation of the multi-video and generates a new
distribution to initiate the tracker when tracking failure is detected.
Experimental results based on the HDTV video sequences, which were captured
by four cam- eras located at the corners of the court, show that the success
rate of the 3D ball tracking achieves 99.14%. |
Slides |
Time |
17:00 ~ 18:00pm |
|
Speaker |
Prof. Arnaud Doucet (University of Oxford) |
|
Title |
On
pseudo-marginal methods for Bayesian inference in latent variable models |
|
Abstract |
The
pseudo-marginal algorithm is a popular variant of the Metropolis-Hastings scheme
which allows us to sample asymptotically from a target probability density
when we are only able to estimate unbiasedly an un-normalized version of it.
It has found numerous applications in Bayesian statistics as there are many
latent variable models where the likelihood function is intractable but can
be estimated unbiasedly using Monte Carlo samples. In this talk, we will
first review the pseudo-marginal algorithm and show that its computational
cost is for many common applications quadratic in the number of observations
at each iteration. We will then present a simple modification of this
methodology which can reduce very substantially this cost. A large sample
analysis of this novel pseudo-marginal scheme will be presented. (This is joint
work with George Deligiannidis and Michael K. Pitt) |
|
Slides Video |
||
Time |
18:00 ~ 19:00pm |
|
Speaker |
Prof. Francois Septier (Telecom Lille 1) |
|
Title |
Sequential
Markov Chain Monte Carlo for Bayesian Filtering with Massive Data |
|
Abstract |
Advances in
digital sensors, digital data storage and communications have resulted in
systems being capable of accumulating large collections of data. In the
light of dealing with the challenges that massive data present, I will
present in this talk solutions to inference and filtering problems
within the Bayesian framework. Two novel Bayesian inference algorithms
are developed for non-linear and non-Gaussian state space models, able to
deal with large volumes of data (or observations). These are sequential
Markov chain Monte Carlo (MCMC) approaches relying on two key ideas: 1)
subsample the massive data and utilize a smaller subset for filtering and
inference, and 2) a divide and conquer type approach computing local
filtering distributions each using a subset of the measurements.
Simulation results highlight the accuracy and the large computational
savings, that can reach 90% by the proposed algorithms when compared
with standard techniques. (This is a joint
work with Allan De Freitas and Lyudmila Mihaylova) |
|
Time |
19:00 ~ 20:00pm |
|
Speaker |
Ms. Anna Zaremba (UCL) |
|
Title |
Modelling
Causality With Gaussian Process Models |
|
Abstract |
In this work we
are developing new classes of multivariate time-series models based on
Gaussian process functional non-parametric relationships between current
vector of observations and the lagged history of each marginal time series.
Then within these new classes of non-parameter time series model we consider
several families of causality testing that can be used to identify the conditional
dependence relationships that may arise in the non-parametric time series
structures. This can include structural causality relationships in the mean
function, covariance function and higher order information. Our main goal is to study the
ability to detect causality in such time series models, whilst allowing for
very general non-linear causal relationships not often considered in
classical parametric time-series models. To perform the testing for such
features we develop compound tests that first require estimation/calibration
of the mean and covariance functions parameterizing the non-parameteric
vector valued time series. We then provide a generic
framework that can be applied to a variety of different problem classes and
produce a number of examples to illustrate the ideas developed. |
|
Slides Video |
||
July 22 (Friday)
Time |
9:00 ~ 10:00am |
Speaker |
Prof. Laurent Clavier (Telecom Lille 1) |
Title |
Next
Generation of wireless communication: some questions concerning reliability |
Abstract |
In the coming
years, for sensing data, a massive number of low-cost, low-energy devices
will be deployed in the environment. This significantly modifies the way
transmissions have been designed in the last 50 years or more. One crucial
question is to have a good representation (model) of what the communication
environment will be and especially of the two essential limitation for
connectivity: the radio channel and interference. We can then re-visit the
fundamental questions (capacity, receiver design) for establishing robust
links. |
Time |
10:00 ~ 11:00am |
Speaker |
Dr. Ido Nevat (Institute for Infocom
Research, A-Star) |
Title |
Query-based Sensors Selection for Collaborative Wireless Sensor Networks with Stochastic Energy Harvesting |
Abstract |
We develop a statistical decision making framework to choose the optimal sub-set of sensors to activate from a set of all possible sensors, while meeting various Quality of Service (QoS) criteria specified by both the network’s operator as well as user’s queries. The sensor nodes are powered solely by energy harvested from the environment and should be activated in an efficient and economical manner based on the available battery energy, which may not directly observed by the decision maker. Our decision making framework consists of two aspects: estimation of the current available energy of each of the sensors; and a sensor activation policy. To this end, we develop a framework based on ruin theory, where the key metric is maximizing the probability that the network’s lifetime exceeds a fixed time period. We then develop an efficient algorithm to perform the sensor selection which is based on the Cross Entropy method. |
Slides Video |
Time |
11:00 ~ 11:30am |
Speaker |
Prof. Tor Andre Myrvoll (NTNU/SINTEF) |
Title |
|
Abstract |
Time |
13:00 ~ 14:00pm |
Speaker |
Prof. Kazuya Takeda(Nagoya University) |
Title |
Data-centric approaches to driving behavior research: how can signal processing methods contribute to the development of autonomous driving? |
Abstract |
|
Slides Video |
Time |
14:00 ~ 14:30pm |
Speaker |
Prof. Jiancang Zhuang (ISM) and Prof. Jorge
Mateau (University Jaume I of Castellon) |
Title |
Semi-parametric
estimates of the long-term background trend, periodicity, and clustering
effect in crime data |
Abstract |
Past studies
have shown that crime behaviors are clustered. This study proposes a
spatiotemporal Hawkes-type point-process model, which includes a background
component with daily and weekly periodization and a clustering component that
is triggered by previous events, for describing the occurrences of violence
or robbery related to crimes in the city of Castellon, Spain, during 2012 and
2013. A nonparametric method, called stochastic reconstruction, is used to
estimate each component, including daily and weekly periodicity of background
rate, spatial background rate, long-term background trend, and the spatial
and temporal response function in the triggering component, of the
conditional intensity of the model. The results show that about 3 percents of
the crimes can be explained by clustering. Keywords: spatiotemporal
point process; Hawkes process; stochastic reconstruction; edge effect
correction |
Time |
14:30 ~ 15:30pm |
Speaker |
Prof. Kenji Fukumizu (ISM) |
Title |
Kernel Methods
for Topdogical Data Anaiysis |
Abstract |
|
Time |
16:00 ~ 17:00pm |
|
Speaker |
Prof. Taiji Suzuki (Tokyo Institute of
Technology) |
|
Title |
Statistical
Performance and Computational Efficiency of Nonparametric Low Rank Tensor
Estimators |
|
Abstract |
In this talk, we
consider statistical performances of some nonparametric tensor estimators. Nonparametric
tensor models have a wide range of applications such as recommendation
system, spatio-temporal data analysis, and multi-task learning. We show that
a Bayesian approach achieves the mini-max optimal predictive accuracy. On the
other hand, an alternating minimization approach has a computationally attractive
property. That is, the method converges linearly to a good local solution
that is mini-max optimal if we assume stronger conditions than the Bayes estimator.
We discuss the trade-off between the statistical performances and the computational
efficiency by showing some theoretical and numerical results. |
|
Slides Video |
||
Time |
17:00 ~ 18:00pm |
|
Speaker |
Prof. Jen-Tzung Chien (National Chiao
Tung University) |
|
Title |
Neural Network Learning |
|
Abstract |
In this talk, I will present a variety of learning strategies to deal with different issues in neural network model. In tensor factorized neural network, a tensor factorized error backpropagation algorithm is developed to preserve the structure of tensor inputs in layer-wise network during training a regression or classification network. We further present a semi-supervised learning for domain adaptation based on neural network model which jointly minimizes the divergence between the distributions from labeled and unlabeled data in source and target domains, the reconstruction errors due to an auto-encoder, and the classification errors due to the labeled data. Finally, a Bayesian unfolding inference is proposed to integrate the benefits from model-based method and neural network model. A Bayesian unfolded topic model is proposed to improve traditional topic model based on variational inference. A number of applications and future works will be addressed. |
|
Time |
18:00 ~ 18:30pm |
|
Speaker |
Prof. Konstatin Markov (Aizu University), Prof. Tomoko Matsui (ISM) |
|
Title |
Articulatory and
Spectrum Features Integration Using Generalized Distillation Framework |
|
Abstract |
It has been
shown that by combining the acoustic and articulatory information significant
performance improvements in automatic speech recognition(ASR) task can be
achieved. In practice, however, articulatory information is not available
during recognition and the general approach is to estimate it from the
acoustic signal. In this paper, we propose different approach based on the
generalized distillation framework, where acoustic-articulatory inversion is
not necessary. In this framework, it is assumed that during training, in
addition to the training data, some kind of "privileged"
information is available and can be used to guide the training process. This
allows to obtain a system which at test time outperforms those built on
regular training data alone. Our system uses hybrid DNN-HMM acoustic model
where neural networks provide HMM state probabilities during decoding. We trained
two DNN models: one called "teacher" learns from both acoustic and articulatory
features, and the other one called "student" is trained on acoustic
features only, but its training process is guided by the "teacher"
model. For experiments, we used the Wisconsin X-ray Microbeam Database (XRMB)
which consists of parallel articulatory and acoustic data. Results clearly
show that distillation framework is effective and allows to achieve
significant reduction in the phoneme error rate. |
|
Time |
19:20 ~ 22:30pm |
|
Dinner &
discussion |
July 23 (Saturday)
Time |
9:00 ~ 10:00am |
Speaker |
Prof. Hiroshi Nakagawa (University of Tokyo) |
Title |
Privacy
Protection:Overview |
Abstract |
Privacy
protection is a big issue for utilizing big data including various personal
data. In addition to rapid growth of information technologies, regulations on
personal data are changing recently. For instance, 14 April 2016, General
Data Protection Regulations is adopted by the European Parliament. In
Japan, Personal Information Protection Act is approved Sep. 2015. To cope
with or even utilize these circumstances, the overview of mathematical models
of privacy protection is described from the viewpoint of what is to be
protected for what purposes by what method. The topics included are
pseudonymization, k-anonymization, questioner’s privacy, differential privacy
and so on. |
Time |
10:00 ~ 11:00am |
Speaker |
Prof. Kana Shimizu (Waseda University) |
Title |
Privacy-preserving
genome sequence search |
Abstract |
The
state-of-the-art DNA sequencer generates 160 Giga bases per day, which is
hundreds of thousands times as large amount of data as the technology of 15
years ago can generate. The huge cost down in DNA sequencing has
encouraged large-scale personal genome sequencing, which eventually
spotlighted privacy issues in genomics. In our work, we assumed the frequent
case such that the user wishs to query the server while hiding the contents
of the query, and developed a novel algorithm that enables searching on DNA
sequences without leaking user’s query to the server. The proposed algorithm
combines a searchable string data structure such as (positional)
Burrows-Wheeler Transform and a cryptographic technique called oblivious
transfer, and allows variable length substring match. In an experiment using
the dataset created from 1000 Genome project, our algorithm was order of
magnitude efficient both in run time and data transfer overhead compared to
the base line exhaustive method. |
Time |
11:00 ~ 11:30am |
Speaker |
Prof. Kazuhiro Minami (ISM) |
Title |
Anonymization
and Location Privacy |
Abstract |
Location data is
useful for various analytic purposes in urban planning and marketing and
there is thus a strong demand for its secondary usage. However, to
anonymize time-series location data is notoriously difficult to achieve due to strong temporal and spatial
correlations among data points. In this talk, we introduce a statistical
framework for measuring location privacy and discuss possible measures
for improving the safety of anonymized location traces. |
Time |
11:30 ~ 12:00am |
Speaker |
Dr. Toshinao Yoshiba, Dr. Naoshi
Tsuchida (Bank of Japan) Dr. Toshiaki Watanabe (Hitotsubashi
University) |
Title |
The
Intraday Market Liquidity of Japanese Government Bond Futures |
Abstract |
We investigate the intraday market liquidity of the Japanese government bond (JGB) futures. First, we overview the movement of various market liquidity indicators during the past decade, classifying them into four categories: tightness, depth, resiliency, and volume. Second, using the data under the current trade time, we extract their intraday pattern and the autocorrelation. Third, we find that the announcement of economic indicator has a negative effect on these liquidity indicators while the monetary policy announcement and the surprise of economic indicator have a positive effect on volume indicators. Fourth, we show that the shock persistence in liquidity indicators rises around April 2013, and the increased persistence remains in some liquidity indicators even several months after April 2013. |
Slides |
Time |
12:00 ~ 12:15am |
|
Closing |