Xin Li Department of Information Systems, College of Business, City University of Hong Kong, Hong Kong, CHINA

Similar documents
we have E Y x t ( ( xl)) 1 ( xl), e a in I( Λ ) are as follows:

Research on Complex Networks Control Based on Fuzzy Integral Sliding Theory

Appendix for An Efficient Ascending-Bid Auction for Multiple Objects: Comment For Online Publication

Neural network-based athletics performance prediction optimization model applied research

A finite difference method for heat equation in the unbounded domain

Managing Capacity Through Reward Programs. on-line companion page. Byung-Do Kim Seoul National University College of Business Administration

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U)

COXREG. Estimation (1)

Supplementary material: Margin based PU Learning. Matrix Concentration Inequalities

MARKOV CHAIN AND HIDDEN MARKOV MODEL

Using T.O.M to Estimate Parameter of distributions that have not Single Exponential Family

Supplementary Material: Learning Structured Weight Uncertainty in Bayesian Neural Networks

On the Power Function of the Likelihood Ratio Test for MANOVA

The line method combined with spectral chebyshev for space-time fractional diffusion equation

Dr. Shalabh Department of Mathematics and Statistics Indian Institute of Technology Kanpur

Lecture Notes on Linear Regression

Machine learning: Density estimation

Predicting Model of Traffic Volume Based on Grey-Markov

MATH 5707 HOMEWORK 4 SOLUTIONS 2. 2 i 2p i E(X i ) + E(Xi 2 ) ä i=1. i=1

Unified Subspace Analysis for Face Recognition

Not-for-Publication Appendix to Optimal Asymptotic Least Aquares Estimation in a Singular Set-up

Nested case-control and case-cohort studies

A Local Variational Problem of Second Order for a Class of Optimal Control Problems with Nonsmooth Objective Function

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

Delay tomography for large scale networks

Quality-of-Service Routing in Heterogeneous Networks with Optimal Buffer and Bandwidth Allocation

8/25/17. Data Modeling. Data Modeling. Data Modeling. Patrice Koehl Department of Biological Sciences National University of Singapore

CS 2750 Machine Learning. Lecture 5. Density estimation. CS 2750 Machine Learning. Announcements

Assortment Optimization under MNL

Case Study of Markov Chains Ray-Knight Compactification

Regulation No. 117 (Tyres rolling noise and wet grip adhesion) Proposal for amendments to ECE/TRANS/WP.29/GRB/2010/3

The Minimum Universal Cost Flow in an Infeasible Flow Network

ON A DETERMINATION OF THE INITIAL FUNCTIONS FROM THE OBSERVED VALUES OF THE BOUNDARY FUNCTIONS FOR THE SECOND-ORDER HYPERBOLIC EQUATION

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

Assignment 5. Simulation for Logistics. Monti, N.E. Yunita, T.

Deriving the Dual. Prof. Bennett Math of Data Science 1/13/06

p(z) = 1 a e z/a 1(z 0) yi a i x (1/a) exp y i a i x a i=1 n i=1 (y i a i x) inf 1 (y Ax) inf Ax y (1 ν) y if A (1 ν) = 0 otherwise

Hidden Markov Model Cheat Sheet

Kernel Methods and SVMs Extension

The lower and upper bounds on Perron root of nonnegative irreducible matrices

More metrics on cartesian products

PROBLEM SET 7 GENERAL EQUILIBRIUM

Pricing and Resource Allocation Game Theoretic Models

Lecture 4. Instructor: Haipeng Luo

Greyworld White Balancing with Low Computation Cost for On- Board Video Capturing

Maximizing the number of nonnegative subsets

Modelli Clamfim Equazione del Calore Lezione ottobre 2014

Errors for Linear Systems

Chapter 12. Ordinary Differential Equation Boundary Value (BV) Problems

Approximate merging of a pair of BeÂzier curves

Nonextensibility of energy in Tsallis statistics and the zeroth law of

Reconstruction History. Image Reconstruction. Radon Transform. Central Slice Theorem (I)

SOME NOISELESS CODING THEOREM CONNECTED WITH HAVRDA AND CHARVAT AND TSALLIS S ENTROPY. 1. Introduction

Chapter 4: Root Finding

A generalization of Picard-Lindelof theorem/ the method of characteristics to systems of PDE

First day August 1, Problems and Solutions

Convexity preserving interpolation by splines of arbitrary degree

Associative Memories

An Upper Bound on SINR Threshold for Call Admission Control in Multiple-Class CDMA Systems with Imperfect Power-Control

Y. Guo. A. Liu, T. Liu, Q. Ma UDC

First Year Examination Department of Statistics, University of Florida

An LSB Data Hiding Technique Using Prime Numbers

PHY688, Statistical Mechanics

A A Non-Constructible Equilibrium 1

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

2-Adic Complexity of a Sequence Obtained from a Periodic Binary Sequence by Either Inserting or Deleting k Symbols within One Period

Modelli Clamfim Equazioni differenziali 7 ottobre 2013

LOW-DENSITY Parity-Check (LDPC) codes have received

Randić Energy and Randić Estrada Index of a Graph

REAL ANALYSIS I HOMEWORK 1

Computation of Higher Order Moments from Two Multinomial Overdispersion Likelihood Models

on the improved Partial Least Squares regression

The Order Relation and Trace Inequalities for. Hermitian Operators

Optimal Guaranteed Cost Control of Linear Uncertain Systems with Input Constraints

MATH 281A: Homework #6

3 Basic boundary value problems for analytic function in the upper half plane

Optimum Selection Combining for M-QAM on Fading Channels

Perron Vectors of an Irreducible Nonnegative Interval Matrix

GENERATIVE AND DISCRIMINATIVE CLASSIFIERS: NAIVE BAYES AND LOGISTIC REGRESSION. Machine Learning

CHAPTER III Neural Networks as Associative Memory

Problem Set 9 Solutions

Modelli Clamfim Equazioni differenziali 22 settembre 2016

STAT 405 BIOSTATISTICS (Fall 2016) Handout 15 Introduction to Logistic Regression

Counting Solutions to Discrete Non-Algebraic Equations Modulo Prime Powers

Outline. EM Algorithm and its Applications. K-Means Classifier. K-Means Classifier (Cont.) Introduction of EM K-Means EM EM Applications.

Hila Etzion. Min-Seok Pang

Image Classification Using EM And JE algorithms

Asymptotics of the Solution of a Boundary Value. Problem for One-Characteristic Differential. Equation Degenerating into a Parabolic Equation

Goodness of fit and Wilks theorem

Machine Learning. Classification. Theory of Classification and Nonparametric Classifier. Representing data: Hypothesis (classifier) Eric Xing

Bézier curves. Michael S. Floater. September 10, These notes provide an introduction to Bézier curves. i=0

HMMT February 2016 February 20, 2016

Lecture 12: Discrete Laplacian

Malaya J. Mat. 2(1)(2014) 49 60

Chapter 7 Channel Capacity and Coding

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Multispectral Remote Sensing Image Classification Algorithm Based on Rough Set Theory

DETERMINATION OF UNCERTAINTY ASSOCIATED WITH QUANTIZATION ERRORS USING THE BAYESIAN APPROACH

SELECTED SOLUTIONS, SECTION (Weak duality) Prove that the primal and dual values p and d defined by equations (4.3.2) and (4.3.3) satisfy p d.

Transcription:

RESEARCH ARTICLE MOELING FIXE OS BETTING FOR FUTURE EVENT PREICTION Weyun Chen eartment of Educatona Informaton Technoogy, Facuty of Educaton, East Chna Norma Unversty, Shangha, CHINA {weyun.chen@qq.com} Xn L eartment of Informaton Systems, Coege of Busness, Cty Unversty of Hong Kong, Hong Kong, CHINA {xn..hd@gma.com} ane Zeng eartment of Management Informaton Systems, Unversty of Arzona, Tucson, AZ 8572 U.S.A., and State Key Laboratory of Management and Contro for Comex Systems, Insttute of Automaton, Chnese Academy of Scences, Beng, CHINA {zeng@ema.arzona.edu} Aendx Theorem : If α ()= β (-), then f(x;,θ) = f(-x;-,θ). Proof: λ f( x;, θ ) = Beta x;, = λ = = ( α β ) λ = Beta x = λ ( - ; β, α ) λ = Beta ( - x; α, ) ( ;, ) β = f x θ = λ = # Theorem 2: E[Beta(x; α (),β ())] s strcty ncreasng wth. Proof: α E[Beta(x; α (),β ())] = = α + β β + α Snce s wthn (0, ), both α () and β () are ostve. MIS Quartery Vo. 4 No. 2 Aendx/June 207 A

Chen et a./modeng Fxed Odds Bettng β β α α Z Z h h α ( A) β α u uh β u + uh A = h= 2 h= 2 = α α β Ceary, α and the equa sgn hods ony f a u h =0 when s wthn range (0, ). 0 If a u h =0, both α () and β () equa 0, whch s n confct wth our assumtons. β () α Thus s strcty decreasng n, and s strcty ncreasng wth. # α() α + β Theorem 3: If arameters q,, ρ, γ, τ, o A, o B are a ostve, there exsts and ony exsts one beef vaue c 0 (0,), caed baance beef hereafter, satsfyng U(c, o A ) = U( c, o B ). Proof: When q,, ρ, γ, τ are ostve, both w + ( ) and w - ( ) are strcty ncreasng functons. Accordngy, the utty functon U(x, o) s a strcty ncreasng functon and U( x,o) s a strcty decreasng functon n beef x. Gven o A > 0 and o B > 0, [U(x, o A ) U(-x, o B )] s strcty ncreasng. It s easy to verfy that U(x = 0, o A ) < 0 < U(x =, o B ) and U(x =, o A ) > 0 > U(x = 0, o B ). Thus, [U(x, o A ) U(-x, o B )] < 0 for x = 0 and [U(x, o A ) U( x, o B )] > 0 for x =. As such, there must exst one and ony one baance beef x = c, satsfyng U(c, o A ) = U( c, o B ). # Theorem 4: For suffcenty arge m, maxmzng equaton 7 reduces to sovng PA(,θ) = s A. Proof: Lc(,θ) n equaton (7) s contnuous and dfferentabe. Snce 0 < <, the vaue of that maxmzes Lc(,θ), f t exsts, must satsfy the Lc(, θ) frst-order condton = 0. Lc(, θ ) PA(, θ) PA(, θ) = ms A + m( sa) + R + ( R) PA(, θ) PA(, θ) ( s ) ( R ) PA(, θ ) s A A R = m - + - PA(, θ) PA(, θ) ( θ ) θ ( θ ) PA(, ) (, ) - (, ) θ sa PA sa PA θ R R = m + PA(, ) PA(, ) PA(, θ) sa- PA(, θ) R = m + PA(, θ) PA(, ) Namey: [ PA θ s ] ( θ ) =0 ( θ )( ) ( ) PA(, ) (, ) (, ) θ PA θ PA R (, )- A = m A2 MIS Quartery Vo. 4 No. 2/June 207

Chen et a./modeng Fxed Odds Bettng ( θ )( ) ( ) PA(, ) (, ) (, ) θ PA θ PA R m [ PA(, θ )- sa] = m =0 + m + m m PA(, θ) Accordng to Theorem 4, >0, we obtan m [ PA(, θ )- sa] =0. # m + Theorem 5: I c (α (),β ()) s strcty decreasng n, where I c (α (),β ()) s the reguarzed ncomete Beta functon 0 c B eta ( x; α, β ) dx. Ic( α, β ) Ic( α, β ) α Ic( α, β) β Proof: Based on the chan rue of mutvarabe cacuus, = +. α β k I c ( α, β ) Γ( α) Γ ( α + β ) α ( α) k( α) k( β ) kc = [og( c) ϕ ( α) + ϕ ( α + β )] Ic ( α, β ) c α Γ( β ) k! Γ ( k + + α ) Γ ( k + + β ) where ( ) k s the Pochhammer symbo secfed as (x) 0 = ; (x) n = x(x + )(x + 2) (x + n-). Snce ϕ( α) = ( ) r, ϕ ( α + β) ϕ ( α) = ( ) < 0 when α > 0 and β > 0. k = k k + α k = k + α + β k + α Snce c<, og (c) < 0 and [og(c) φ(α) + φ(α + β)] < 0. Snce I c (α,β) > 0 f 0 < c <, we have [og(c) φ(α) + φ(α + β)]i c (α, β) < 0. k = 0 Snce Γ(x)>0 f x>0, we have k Γ( α) Γ ( α + β) α ( α) k( α) k( β) kc c > 0 Γ( β) k! Γ ( k + + α) Γ ( k + + β) k = 0 I Thus, ( c α, β ) < 0 when 0 < c<. α I Smary, we can rove ( c α, β ) > 0 when 0 < c <. β α β It s cear > 0 and < 0 when 0 < <. Ic( α, β ) Thus, < 0,.e., I c (α(),β()) s strcty decreasng n. # Formua for AIC The vaue of AIC crtera s comuted as AICc = 2t + 2 t( t + )/( N t ) 2Lc θ where N denotes the number of data nstances and t denotes the number of arameters, whch s ( + Z) > t = Z = accordng to equatons () and (2). MIS Quartery Vo. 4 No. 2 Aendx/June 207 A3

Chen et a./modeng Fxed Odds Bettng Inut: {o A, o B, R, s A, o B,}, 0 {,, H}, k max, ε. Outut: estmated otma arameter θ*.. Comute baance beef c, c 2,, c H for each bettng game accordng to (9), usng a root fndng agorthm. 2. Intaze a smex SX whch conssts of J + sets of arameters θ n the arameter sace Θ, where J s the dmenson of the arameter sace. 3. Whe: 4. For each θ : 5. Sove A from each bettng accordng to (0) usng a root fndng agorthm. 6. Comute the Lc( ) accordng to (2). ( max( Lc ) mn( ) ) * θ Lc * θ ε 7. If or IteratonCount < k max : 8. Udate vertces n SX usng Refect, Exand, Outsde contracton, Insde contracton, or Shrnk oeratons. 9. If oeraton <> Shrnk : IteratonCount = IteratonCount + 0. Ese: *. Return the arameter set corresondng to max( Lc ( θ )) as θ*. Fgure A. Maxmum Lkehood Estmaton Usng a Neder Mead Method etaed Beef strbuton Estmaton Procedure Sna 2008 Oymc Games ataset For each settng of and Z, varyng from to 3, resectvey, we numercay obtaned the otma arameters θ* 0 Θ. Tabe A reorts the og kehood and AIC vaues for these modes. Generay, the mode s kehood converges when and Z are arger than 2. The mode wth = 2 and Z = 2 s the mode wth the maxmum kehood. The estmated beef dstrbuton functon s gven as: f( x; ) = 0.74* Beta( x;3.76* +0.*,3.76*( )+0.*( ) ) + 0.26* Beta( x;0.* + 66.32*,0.*( ) + 66.32*( ) ) For the AIC crtera, we combned the three comonents wth smaest AICc ( =, Z = ; =, Z = 2; and =, Z = 3). The estmated beef dstrbuton functon s gven as f ( x; ) = 0.67* Beta( x;5.98*,5.98*( )) + + + + 0.09* Beta( x;5.95* + 0.02*,5.95*( ) + 0.02*( ) ) 0.24* Beta( x;5.69* 2 0.6*,5.69* 0.6*( 2 ) ) 3 3 Tabe A. Log Lkehood and AIC Vaues (Sna 2008 Oymc Games) Z = Z = 2 Z = 3 Log kehood AIC Log kehood AIC Log kehood AIC -0.04.26-0. 206.29-0.08.38 2-0.03 20.30-00.79 24. -0.05 29.0 3-0.03 24.59-00.80 220.74-00.827.67 A4 MIS Quartery Vo. 4 No. 2/June 207

Chen et a./modeng Fxed Odds Bettng Sohu Entertanment ataset Tabe A2 shows the resuts on the Sohu entertanment event dataset, varyng and Z from to 3. The mode ( = 3 and Z = 3) s the mode wth the maxmum kehood. The estmated beef dstrbuton functon s as foows: f x Beta x 2 3 2 3 ( ; ) = 0.84* ( ;0.* +0.* + 000*,0.*( )+0.*( ) + 000*( ) ) + + + + + 2 3 2 3 0.6* Beta( x;7.54* 0.* 0.*,7.54* 0.* 0.* ) For the AIC crtera, we combned the three comonents wth smaest AIC ( =, Z = and =, Z = 2 and =, Z = 3). The estmated beef dstrbuton functon s gven as f ( x; ) = 0.62* Beta( x;2.28*, 2.28*( )) + 0.29* Beta( x;.29* + 3.3*,.29*( ) + 3.3*( ) ) 2 3 2 3 + 0.09* Beta( x; + 2.6* +,( ) + 2.6*( ) + ( ) ) Tabe A2. Lkehood and AIC Vaues (Sohu Entertanment) Z = Z = 2 Z = 3 Log kehood AIC Log kehood AIC Log kehood AIC -22.56 47.25-22.20 48.79-22.20 5.20 2-22.26 53.90-2.32 57.75-2.32 64.40 3-22.26 59.63-2.32 68.4-2.20 8.26 Sohu 204 FIFA ataset Tabe A3 shows the resuts on the Sohu 204 FIFA dataset, varyng and Z from to 3. The mode ( = and Z = ) s the mode wth the maxmum kehood. The estmated beef dstrbuton s gven as ( ; ) = ; 428., 428. ( ) f x Beta x For the AIC crtera, we combne the three comonents wth smaest AIC ( =, Z = ; =, Z = 2; and =, Z = 3). The estmated beef dstrbuton functon s gven as f ( x; ) = 0.67* Beta( x;42.8*, 42.8*( )) + 0.24* Beta( x;4.88*, 4.88*( )) + 0.09* Beta( x;4.98*,4.98*( )) Tabe A-3: Lkehood and AIC Vaues (Sohu 204 FIFA) Z = Z = 2 Z = 3 Log kehood AIC Log kehood AIC Log kehood AIC -9.36 240.74-9.36 242.79-9.36 244.86 2-9.36 246.95-9.36 25.2-9.36 255.56 3-9.36 25.2-9.36 257.78-9.36 264.59 MIS Quartery Vo. 4 No. 2 Aendx/June 207 A5

Chen et a./modeng Fxed Odds Bettng Wth the estmated beef dstrbuton functon, we can cacuate the ercentage of eoe who consder the event may haen f there are no odds (or equa odds on both sdes) n the redcton market, as ustrated n Fgure A2. In a three datasets, the bettors beefs are more extreme than the actua event robabty. If the event robabty s ess than 0.5, the fna bet rato w be ower than the event robabty. If the event robabty s hgher than 0.5, the fna bet rato s hgher than event robabty. Peoe wth a Beef >.5 Sna 2008 Oymc Games Sohu Entertanment Sohu 204 FIFA Fgure A2. Bettor Beef Wthout Odds Infuence A6 MIS Quartery Vo. 4 No. 2/June 207