Application of Quasi Monte Carlo methods and Global Sensitivity Analysis in finance



Similar documents
The Application of Fractional Brownian Motion in Option Pricing

Analysis of Premium Liabilities for Australian Lines of Business

Realistic Image Synthesis

Loop Parallelization

Risk-based Fatigue Estimate of Deep Water Risers -- Course Project for EM388F: Fracture Mechanics, Spring 2008

Hedging Interest-Rate Risk with Duration

Forecasting the Direction and Strength of Stock Market Movement

Lecture 3: Force of Interest, Real Interest Rate, Annuity

Recurrence. 1 Definitions and main statements

A hybrid global optimization algorithm based on parallel chaos optimization and outlook algorithm

ECONOMICS OF PLANT ENERGY SAVINGS PROJECTS IN A CHANGING MARKET Douglas C White Emerson Process Management

Pricing Multi-Asset Cross Currency Options

Level Annuities with Payments Less Frequent than Each Interest Period

What is Candidate Sampling

THE DISTRIBUTION OF LOAN PORTFOLIO VALUE * Oldrich Alfons Vasicek

NPAR TESTS. One-Sample Chi-Square Test. Cell Specification. Observed Frequencies 1O i 6. Expected Frequencies 1EXP i 6

The Development of Web Log Mining Based on Improve-K-Means Clustering Analysis

Credit Limit Optimization (CLO) for Credit Cards

L10: Linear discriminants analysis

Imperial College London

The Cox-Ross-Rubinstein Option Pricing Model

Lecture 3: Annuity. Study annuities whose payments form a geometric progression or a arithmetic progression.

Support Vector Machines

SPEE Recommended Evaluation Practice #6 Definition of Decline Curve Parameters Background:

Causal, Explanatory Forecasting. Analysis. Regression Analysis. Simple Linear Regression. Which is Independent? Forecasting

GRAVITY DATA VALIDATION AND OUTLIER DETECTION USING L 1 -NORM

Stock Profit Patterns

An Analysis of Pricing Methods for Baskets Options

Solution: Let i = 10% and d = 5%. By definition, the respective forces of interest on funds A and B are. i 1 + it. S A (t) = d (1 dt) 2 1. = d 1 dt.

An Integrated Semantically Correct 2.5D Object Oriented TIN. Andreas Koch

On the Optimal Control of a Cascade of Hydro-Electric Power Stations

Face Verification Problem. Face Recognition Problem. Application: Access Control. Biometric Authentication. Face Verification (1:1 matching)

Characterization of Assembly. Variation Analysis Methods. A Thesis. Presented to the. Department of Mechanical Engineering. Brigham Young University

A) 3.1 B) 3.3 C) 3.5 D) 3.7 E) 3.9 Solution.

Forecasting the Demand of Emergency Supplies: Based on the CBR Theory and BP Neural Network

Conversion between the vector and raster data structures using Fuzzy Geographical Entities

Optimal resource capacity management for stochastic networks

benefit is 2, paid if the policyholder dies within the year, and probability of death within the year is ).

Course outline. Financial Time Series Analysis. Overview. Data analysis. Predictive signal. Trading strategy

Prediction of Disability Frequencies in Life Insurance

Luby s Alg. for Maximal Independent Sets using Pairwise Independence

How To Understand The Results Of The German Meris Cloud And Water Vapour Product

Power-of-Two Policies for Single- Warehouse Multi-Retailer Inventory Systems with Order Frequency Discounts

A Novel Methodology of Working Capital Management for Large. Public Constructions by Using Fuzzy S-curve Regression

IMPACT ANALYSIS OF A CELLULAR PHONE

Abstract. 260 Business Intelligence Journal July IDENTIFICATION OF DEMAND THROUGH STATISTICAL DISTRIBUTION MODELING FOR IMPROVED DEMAND FORECASTING

Sensor placement for leak detection and location in water distribution networks

8 Algorithm for Binary Searching in Trees

Statistical Methods to Develop Rating Models

Adaptive Fractal Image Coding in the Frequency Domain

Vision Mouse. Saurabh Sarkar a* University of Cincinnati, Cincinnati, USA ABSTRACT 1. INTRODUCTION

A Hierarchical Anomaly Network Intrusion Detection System using Neural Network Classification

Improved SVM in Cloud Computing Information Mining

Risk Model of Long-Term Production Scheduling in Open Pit Gold Mining

Preventive Maintenance and Replacement Scheduling: Models and Algorithms

Portfolio Loss Distribution

INVESTIGATION OF VEHICULAR USERS FAIRNESS IN CDMA-HDR NETWORKS

"Research Note" APPLICATION OF CHARGE SIMULATION METHOD TO ELECTRIC FIELD CALCULATION IN THE POWER CABLES *

Production. 2. Y is closed A set is closed if it contains its boundary. We need this for the solution existence in the profit maximization problem.

DEFINING %COMPLETE IN MICROSOFT PROJECT

On the pricing of illiquid options with Black-Scholes formula

Module 2 LOSSLESS IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Quantization Effects in Digital Filters

Feature selection for intrusion detection. Slobodan Petrović NISlab, Gjøvik University College

A frequency decomposition time domain model of broadband frequency-dependent absorption: Model II

PAS: A Packet Accounting System to Limit the Effects of DoS & DDoS. Debish Fesehaye & Klara Naherstedt University of Illinois-Urbana Champaign

Calendar Corrected Chaotic Forecast of Financial Time Series

The OC Curve of Attribute Acceptance Plans

Implementation of Deutsch's Algorithm Using Mathcad

Trade Adjustment and Productivity in Large Crises. Online Appendix May Appendix A: Derivation of Equations for Productivity

ANALYZING THE RELATIONSHIPS BETWEEN QUALITY, TIME, AND COST IN PROJECT MANAGEMENT DECISION MAKING

WORKING PAPERS. The Impact of Technological Change and Lifestyles on the Energy Demand of Households

How To Calculate The Accountng Perod Of Nequalty

Binomial Link Functions. Lori Murray, Phil Munz

Time Value of Money. Types of Interest. Compounding and Discounting Single Sums. Page 1. Ch. 6 - The Time Value of Money. The Time Value of Money

An Alternative Way to Measure Private Equity Performance

An Evaluation of the Extended Logistic, Simple Logistic, and Gompertz Models for Forecasting Short Lifecycle Products and Services

Time Domain simulation of PD Propagation in XLPE Cables Considering Frequency Dependent Parameters

RELIABILITY, RISK AND AVAILABILITY ANLYSIS OF A CONTAINER GANTRY CRANE ABSTRACT

Regression Models for a Binary Response Using EXCEL and JMP

Estimation of Dispersion Parameters in GLMs with and without Random Effects

Return decomposing of absolute-performance multi-asset class portfolios. Working Paper - Nummer: 16

STATISTICAL DATA ANALYSIS IN EXCEL

Comparison of Control Strategies for Shunt Active Power Filter under Different Load Conditions

Vasicek s Model of Distribution of Losses in a Large, Homogeneous Portfolio

Ring structure of splines on triangulations

THE METHOD OF LEAST SQUARES THE METHOD OF LEAST SQUARES

NMT EE 589 & UNM ME 482/582 ROBOT ENGINEERING. Dr. Stephen Bruder NMT EE 589 & UNM ME 482/582

A DYNAMIC CRASHING METHOD FOR PROJECT MANAGEMENT USING SIMULATION-BASED OPTIMIZATION. Michael E. Kuhl Radhamés A. Tolentino-Peña

Pricing Overage and Underage Penalties for Inventory with Continuous Replenishment and Compound Renewal Demand via Martingale Methods

CS 2750 Machine Learning. Lecture 3. Density estimation. CS 2750 Machine Learning. Announcements

How Sets of Coherent Probabilities May Serve as Models for Degrees of Incoherence

v a 1 b 1 i, a 2 b 2 i,..., a n b n i.

Transcription:

Applcaton of Quas Monte Carlo methods and Global Senstvty Analyss n fnance Serge Kucherenko, Nlay Shah Imperal College London, UK skucherenko@mperalacuk Daro Czraky Barclays Captal DaroCzraky@barclayscaptalcom 1

Outlne Applcaton of MC methods to path dependent ntegrals MC and QMC smulaton of opton prcng Is the Brownan brdge dscretzaton always more effcent than the Standard scheme? Global Senstvty Analyss and Sobol Senstvty Indces MC/QMC smulaton of the Cox, Ingersoll and Ross nterest rate model Comparson of dfferent Sobol sequence generators 2

Applcaton of MC methods to path dependent ntegrals I = F[ x( t)] d x, (1) C W xt ( ) contnuous n 0 t T, x(0) = x I = E( F[ W( t)]), W( t) random Wener processes (a Brownan moton) Monte Carlo approach: to construct many random paths W( t), evaluate functonal and average results 0 (1) can be reduced to r r I[ f] = n f( x) dx H 3

Monte Carlo ntegraton methods r I[ f] = E[ f( x)] N 1 r Monte Carlo : IN[ f] = f( z) N = 1 r { z } s a sequence of random ponts n H Error: ε = I[ f] I [ f] N 2 1/2 σ ( f ) εn = ( E( ε )) = 1/2 N Convergence does not depent on dmensonalty but t s slow n 4

Quas random sequences (Low( dscrepancy sequences) Dscrepancy s a measure of devaton from unformty: r r Q y H Q y = y y y n Defntons:: ( ), ( ) [0, 1) [0, 2) [0, n ), mq ( ) volume of Q D N N r Q( y) = sup m( Q) r n Q( y) H N D N N N 1/2 1/2 Random sequences: N (ln ln ) / ~ 1/ n (ln N) DN c( d) Low dscrepancy sequences (LDS) N Convergence: ε = I[ f] I [ f] V( f) D, n O(ln N) εqmc = N Assymptotcally ε ~ O(1/ N) much hgher than ε MC ~ O(1/ N) QMC N N QMC 5

Approxmatons of path dependent ntegrals wth Standard and Brownan brdge schems SDE: dw = z dt, z ~ N (0,1) Standard algorthm: Brownan brdge algorthm: W( t ) = W( t ) + tz, t = T / n, 0 n 1 + 1 + 1 t 0 T/2 T WT ( ) = W+ Tz, 0 1 1 1 W( T /2) = ( W( T) + W0) + Tz2, 2 2 1 1 W( T /4) = ( W( T /2) + W0) + T /2 z3, 2 2 1 1 W(3 T/ 4) = ( WT ( / 2) + WT ( )) + T/ 2 z4, 2 2 M 1 1 W(( n 1) T / n) = ( W(( n 2) T / n) + W( T)) + 2 T / nzn 2 2 6

Opton prcng Dscretzaton of the Wener process Share prce follows geometrcal Brownan moton: 1/2 ds =µ Sdt +σ SdW, dw = z( dt), z ~ N(0,1) Usng Ito's lemma 1 2 St ( ) = S0 exp[( µ σ ) t+σwt ( ))], Wt ( ) Wener path 2 For tme step t 1 2 St ( + t) = St ( )exp[( r σ ) t+σ ( Wt ( + t) W( t))] 2 For the standard dscretzaton algorthm a termnal asset value: 1 2 ST ( ) = S0exp[( r σ ) T+σ t( z1+ z2 + + zn)] 2 7

MC smulaton of opton prcng The value of European style optons rt ( ) C( KT, ) = e E P St ( ), K The payoff functon for an Asan call opton P A =max( S- K,0), For a geometrc average Asan call: S=( S ) There s a closed form soluton Q n 2 rt σ T 1 C( KT, ) = e max[0,( S0 exp[( r ) t +σ Φ ( uj)] n 1 2 n H = j= 1 K)] du du 1 n n =1 1/ n 1/ n 8

MC smulaton of opton prcng Dscretzaton In a general case N rt 1 ( () () C (, ) ) N KT = e P S0, S1, L, ST, K N = 1 For the case of European-style call N N 1 () rt 1 () CN ( K, T) = C = e max( ST K,0) N = 1 N = 1 9

MC and QMC methods wth standard and Brownan Brdge dscretzatons Asan Call (32 observatons) S=100, K=105, r=005, s=02, T=05, C=384 (analytcal) 45 4 Opton Value 35 QMC, Brownan Brdge QMC, Standard Approxmaton MC, Brownan Brdge 3 Analytcal value 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 N_path Call prce vrs the number of paths MC - slow convergences, convergence curve s hghly oscllatng QMC convergence monotonc Convergence s much faster for Brownan brdge 10

Asan call Convergence curves Asan Call wth geometrc averagng 252 observatons S=100, K=105, r=005, s=02, T=10, C=556 (analytcal) 10 K 1 k ε = ( I IN ) K k = 1 2 1/2 1 Log(RMSE) 01 ε ~ N α, 0< α < 1 001 0001 QMC, Brownan Brdge QMC, Standard Approxmaton MC, Brownan Brdge Trendlne -QMC, BB, 1/N^082 Trendlne - QMC, Stand, 1/N^056 Trendlne - MC, 1/N^05 10 100 1000 10000 Log(N_path) Log-log plot of the root mean square error versus the number of paths Brownan brdge much faster convergence wth QMC methods: ~1/N 08 11

Y = f ( x) x ( x, x,, x k ) 0 x 1 Consder a model x s a vector of nput varables = Y s the model output 1 2 1 0 1 s 1 s k k ( ) ( ) ( ) Y = f ( x) = f + f x + f x, x + + f x, x,, x, ANOVA decomposton and Senstvty Indces x Ω ANOVA decomposton: 0 j j 1,2,, k 1 2 k = 1 j> f ( x,,, x ) dx = 0, k, 1 k s Varance decomposton: Model, f(x) 2 2 2 2, j 1,2,, n Y K σ = σ + σ + σ j Sobol SI: k = 1 + = 1 S + S j + S jl + S 1, 2,, < j < j < l k 12

Sobol Senstvty Indces (SI) Defnton: σ S σ σ 2 2 = / 1 s 1 s 1 2 2 ( ) 1 = 1,,,, s f x s x 1 s dx x 1 s 0 1 2 2 σ = ( f ( x) f0 ) dx 0 - partal varances - varance Senstvty ndces for subsets of varables: σ m 2 2 y = σ 1, K, s= 1 Κ ( ) Introducton of the total varance: 1 s s x ( tot ) 2 2 2 σ = σ σ y z = ( y, z) Correspondng global senstvty ndces: 2 2 tot S = / σ, ( ) 2 S = σ tot / σ y σ y y y 2 13

How to use Sobol Senstvty Indces? tot y y tot 0 S S 1 = 1 y S S accounts for all nteractons between y and z, x=(y,z) The mportant ndces n practce are and ( ) ( ) tot S = 0 f x does not depend on ; only depends on ; If y S tot S corresponds to the absence of nteractons between and other varables 1, then functon has addtve structure: f ( x) f f ( x ) Fxng unessental varables If S f x tot S = S n s= 1 S = ( ) S << 1 f x z tot z ( ) (, ) does not depend on so t can be fxed f x f y z n z complexty reducton, from to n n varables 0 x x x = + 0 14

Applcatons of Global Senstvty Analyss Global Senstvty Analyss can be used to dentfy key parameters whose uncertanty most strongly affects the output; rank varables, fx unessental varables and reduce model complexty; select a model structure from a set of known competng models; dentfy functonal dependences; analyze effcences of numercal schemes 15

Effectve dmensons Let u be a cardnalty of a set of varables u The effectve dmenson of f( x) n the superposton sense s the smallest nteger d such that 0< u< d S S u (1 ε), ε << 1 S It means that f( x) s almost a sum of -dmensonal functons d S The functon f( x) has effectve dmenson n the truncaton sense d f u {1,2,, d } T Example: d d S T S (1 ε), ε << 1 u ( ) f x n = = 1 x d = 1, d = n S T does not depend on the order n whch the nput varables are sampled, - depends on the order by reoderng varables d can be reduced T T 16

Classfcaton of functons 17

Global Senstvty Analyss of standard dscretzaton and Brownan Brdge Apply global SA to payoff functon PA ({Z })=max( S({Z })- K,0), {Z }, = 1, n 1 Standard Approxmaton Brownan Brdge 01 S_total 001 0001 00001 1 6 11 16 21 26 31 tme step number Log of total senstvty ndces versus tme step number Standard dscretzaton - S_total slowly decrease wth Brownan brdge - S_total of the frst few varables are much larger than those of the subsequent varables 18

Global Senstvty Analyss of two algorthms at dfferent n Standard approxmaton: the effectve dmenson d T n the effectve dmenson d S > 2 Brownan Brdge approxmaton: the effectve dmenson d T 2 the effectve dmenson d S 2 19

Opton prcng: Why Brownan Brdge s more effcent than standard dscretzaton n the case of QMC? The ntal coordnates of LDS are much better dstrbuted than the later hgh dmensonal coordnates Global SA: for the Brownan brdge dscretzaton the low ndex varables are much more mportant than hgher ndex varables For the Brownan brdge dscretzaton well dstrbuted coordnates are used for mportant varables and hgher not so well dstrbuted coordnates are used for far less mportant varables The standard constructon does not account for the specfcs of LDSs dstrbuton propertes Applcaton of QMC wth the Brownan brdge dscretzaton results n the 10 2-10 6 tme reducton of CPU tme compared wth MC! 20

Cox, Ingersoll and Ross nterest rate model ( ) dy = α+β y dt +σ y dw t t t t where α > 0, β > 0, σ > 0 The generalsed method of moments (GMM) estmaton ) ) ) s used to obtan α, β and σ estmates Data: the 9-month Eurbor nterest rate daly tme seres usng 250 daly observatons, startng at 30 Dec 1998 ) ) ) α= 1e-005, β= 01109, σ= 01929 Euler dscretsaton: ( ) y y = α+βy t+σ y tε t+ t t t t t 21

5-days ahead backtest for the 9-month Eurbor nterest rate seres Forecast s accurate but can be prohbtvely expensve f ran over a large tme span or f a large number of MC runs s needed 22

CIR model: 50 day forecasts of the 9-month Eurbor rate QMC Euler QMC Mlsten 50-perod forecast 362 366 370 374 50-perod forecast 362 366 370 374 0 50 100 150 200 250 Number of smulaton runs (N) MC Euler 0 50 100 150 200 250 Number of smulaton runs (N) MC Mlsten 50-perod forecast 361 363 365 50-perod forecast 358 362 366 370 0 50 100 150 200 250 0 50 100 150 200 250 Number of smulaton runs (N) Number of smulaton runs (N) MC and QMC estmators wth usng standard Euler and Mlsten schemes QMC produces much smother and much faster convergence than MC 23

RMSE convergence for 250-days forecastng horzon K 1 k ε = ( I IN ) K k = 1 2 1/2 0-2 -4 MC Euler (H = 250) QMC Euler (H = 250) Trendlne MC (-053) Trendlne QMC (-079) log2(rm SE) -6-8 -10 ε ~ N α, 0< α < 1-12 -14-16 6 8 10 12 14 16 18 log2(n) QMC estmator dsplays much faster convergence compared to MC regardless of the forecastng horzon (dmenson) 24

RMSE convergence for 250-days forecastng horzon K 1 k ε = ( I IN ) K k = 1 2 1/2 ε ~ N α, 0< α < 1 There s no notable advantage of Brownan brdge over the standard Euler scheme 25

Global Senstvty Analyss of StandardS and Brownan Brdge dscretzatons (CIR model) Apply global SA to a functon yt (Z, Z,, Z ), n s a number of days 1 2 n 1 08 Standard Approxmaton Brownan Brdge S_total 06 04 02 0 1 6 11 16 21 26 31 tme step number Log of total senstvty ndces versus tme step number Standard dscretzaton - S_total are constant Brownan brdge - S_total of the frst varable s much larger than those of the subsequent varables 26

Global Senstvty Analyss of StandardS and Brownan Brdge dscretzatons (CIR model) n S Standard S BB 32 099 099 64 099 099 128 099 099 Standard approxmaton: the effectve dmenson d T n the effectve dmenson d S =1 Brownan Brdge approxmaton: the effectve dmenson d T 2 the effectve dmenson d S =1 The low effectve dmenson d S for both schemes =1, hence QMC effcency can not be further mproved by changng samplng strategy 27

What s the optmal way to arrange N ponts n two dmensons? Regular Grd Sobol Sequence Low dmensonal projectons of low dscrepancy sequences are better dstrbuted than hgher dmensonal projectons 28

Why Sobol sequences are so effcent? n O(ln N) Convergence: ε = for all LDS N n 1 O(ln N) k For Sobol' LDS: ε =,f N = 2, k nteger N Sobol' LDS: 1 Best unformty of dstrbuton as N goes to nfnty 2 Good dstrbuton for farly small ntal sets 3 A very fast computatonal algorthm 29

Sobol LDS Property A and Property A A Property A Consder n-dmensonal hypercube whch s cut by plans x j =1/2 nto 2 n subcubes Sequence of Sobol ponts satsfes Property A, f after dvdng the sequence nto blocks of 2 n ponts, each one of the ponts n any one block belong to a dfferent subcube Property A Consder n-dmensonal hypercube whch s cut by plans x j =k/4, j=1,,n, k=1,2,3 nto 4 n subcubes Sequence of Sobol ponts satsfes Property A, f after dvdng the sequence nto blocks of 4 n ponts, each one of the ponts n any one block belong to a dfferent subcube Property A Property A 30

Comparson of Sobol sequence generators SobolSeq generator: Sobol' sequences satsfy two addtonal unformty propertes: Property A for all dmensons and Property A' for adjacent dmensons I = [,1] s n 0 = 1 ( 1+ c ( x 05) ) dx 31

Summary Global Senstvty Analyss s a general approach for uncertanty, complexty reducton and structure analyss of non-lnear models It can be wdely appled n fnance Quas MC methods outperform MC regardless of nomnal dmensonalty for problems wth low effectve dmensons n ether truncaton or superposton sense The effcency of the Brownan Brdge or any other constructon wthn the framework of the Quas MC method depends on effectve dmensons of the ntegrand The Sobol Sequence generator satsfyng unformty propertes A and A' has superor performance over other generators 32

Acknowledgments Prof Sobol EPSRC grant EP/D506743/1 33