Derivatives Analytics with Python & Numpy

Dr. Yves J. Hilpisch

24 June 2011

EuroPython 2011

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

1 / 34

CV Yves Hilpisch

19931996 Dipl.-Kfm. (MBA) at Saarland University (Banks and Financial Markets) 19962000 Dr.rer.pol. (Ph.D.) at Saarland University (Mathematical Finance) 19972004 Management Consultant Financial Services & Insurance Industry 2005present Founder and MD of Visixion GmbH
management and technical consulting work DEXISIONDerivatives Analytics on Demand (www.dexision.com)

2 3 4

2010present Lecturer Saarland University

Course: Numerical Methods for the Market-Based Valuation of Options Book Project: Market-Based Valuation of Equity DerivativesFrom Theory to Implementation in Python

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

2 / 34

Derivatives Analytics and Python Data Analysis Time Series Cross-Sectional Data Monte Carlo Simulation Model Economy European Options American Options Speed-up of 480+ Times DEXISIONOur Analytics Suite Capabilities Technology

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

3 / 34

Derivatives Analytics and Python

What is Derivatives Anlytics about?

Derivatives Analytics is concerned with the valuation, hedging and risk management of derivative nancial instruments In contrast to ordinary nancial instruments which may have an intrinsic value (like the stock of a company), derivative instruments derive their values from other instruments Tyical tasks in this context are
simulation data analysis (historical, current, simulated data) discounting arithmetic operations (summing, averaging, etc.) linear algebra (vector and matrix operations, regression) solving optimization problems visualization ...

Python can do all this quite wellbut C, C++, C#, Matlab, VBA, JAVA and other languages still dominate the nancial services industry

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

4 / 34

Derivatives Analytics and Python

Why Python for Derivatives Analytics?

Open Source: Python and the majority of available libraries are completely open source Syntax: Python programming is easy to learn, the code is quite compact and in general highly readable (= fast development + easy maintenance) Multi-Paradigm: Python is as good at functional programming as well as at object oriented programming Interpreted: Python is an interpreted language which makes rapid prototyping and development in general a bit more convenient Libraries: nowadays, there is a wealth of powerful libraries available and the supply grows steadily; there is hardly a problem which cannot be easily attacked with an existing library Speed: a common prejudice with regard to interpreted languagescompared to compiled ones like C++ or Cis the slow speed of code execution; however, nancial applications are more or less all about matrix/array manipulations and other operations which can be done at the speed of C code with the essential library Numpy

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

5 / 34

Derivatives Analytics and Python

What does the nancial market say about Python?

in the London area (mainly nancial services) the number of Python developer contract oerings evolved as follows (respectively for the three months period ending on 22 April)
142 in year 2009 245 in year 2010 644 in year 2011

these gures imply a more than fourfold demand for the Python skill in 2011 as compared to 2009 over the same period, the average daily rate for contract work increased from 400 GBP to 475 GBP1 obviously, Python is catching up at a rapid pace in the nancial services industry ...

Source: all gures from www.itjobswatch.co.uk/contracts/london/python.do on 24 April 2011.

Y. Hilpisch (Visixion GmbH) Derivatives Analytics EuroPython 2011 6 / 34

Data Analysis

In Derivatives Analytics you have to analyze dierent types of data

Fundamental types of data to be analyzed

time series cross sections

Python libraries suited to analyze and visualize such data

xlrd (www.python-excel.org): reading data from Excel les Numpy (numpy.scipy.org): array manipulations of any kind Pandas (code.google.com/p/pandas): time series analysis, cross-sectional data analysis2 matplotlib (matplotlib.sourceforge.net): 2d and 3d plotting

Notably, this library was developed by a hedge fund.

Y. Hilpisch (Visixion GmbH) Derivatives Analytics EuroPython 2011 7 / 34

Data Analysis

Time Series

DAX time seriesindex level and daily log returns3

8000 7000 DAX Daily Quotes 6000 5000 4000 3000 2000 1992 0.10 DAX Daily Log Returns 0.05 0.00 0.05 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 1994 1996 1998 2000 2002 2004 2006 2008 2010

Source: http://finance.yahoo.com, 29 Apr 2011

Y. Hilpisch (Visixion GmbH) Derivatives Analytics EuroPython 2011 8 / 34

Data Analysis

Time Series

Reading data from Excel le ...

from from from from from

xlrd import open_workbook pandas import * datetime import * matplotlib . pyplot import * numpy import *

# DAX Open Workbook , Read xls = open_workbook ( ' D A X _ D a i l y _ 1 9 9 0 _ 2 0 1 1 . xls ' ) for s in xls . sheets (): datesDAX = []; quoteDAX = [] for row in range ( s . nrows -1 ,0 , - 1 ): year = int ( s . cell ( row , 0 ). value ) month = int ( s . cell ( row , 1 ). value ) day = int ( s . cell ( row , 2 ). value ) datesDAX . append ( date ( year , month , day )) quoteDAX . append ( float ( s . cell ( row , 8 ). value )) print

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

9 / 34

Data Analysis

Time Series

... and plotting it with Pandas and matplotlib

DAXq = Series ( quoteDAX , index = datesDAX ) DAXr = Series ( log ( DAXq / DAXq . shift ( 1 )) , index = datesDAX ) DAXr = where ( isnull ( DAXr ) , 0 .0 , DAXr ) # Data Frames for Quotes and Returns data = { ' QUO ' : DAXq , ' RET ' : DAXr , ' RVO ' : rv } DAX = DataFrame ( data , index = DAXq . index ) # Graphical Output figure () subplot ( 211 ) plot ( DAX . index , DAX [ ' QUO ' ]) ylabel ( ' DAX Daily Quotes ' ) grid ( True ); axis ( ' tight ' ) subplot ( 212 ) plot ( DAX . index , DAX [ ' RET ' ]) ylabel ( ' DAX Daily Log Returns ' ) grid ( True ); axis ( ' tight ' )

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

10 / 34

Data Analysis

Time Series

DAX time series252 moving mean return, volatility and correlation between both4

0.8 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 1.0 1992 0.50 0.45 0.40 0.35 0.30 0.25 0.20 0.15 0.10 1992 1.0 0.5 0.0 0.5 1.0 1992

Return (252d Mov)

1994

1996

1998

2000

2002

2004

2006

2008

2010

Vola (252d Mov)

1994

1996

1998

2000

2002

2004

2006

2008

2010

Corr (252d Mov)

1994

1996

1998

2000

2002

2004

2006

2008

2010

Source: http://finance.yahoo.com, 29 Apr 2011

Y. Hilpisch (Visixion GmbH) Derivatives Analytics EuroPython 2011 11 / 34

Data Analysis

Time Series

Pandas provides a number of really convenient functions

# mean return , volatility and correlation ( 252 days moving = 1 year ) figure () subplot ( 311 ) mr252 = Series ( rolling_mean ( DAX [ ' RET ' ] , 252 )* 252 , index = DAX . index ) mr252 . plot (); grid ( True ); ylabel ( ' Return ( 252d Mov ) ' ) x , y = REG ( mr252 , 0 ); plot (x , y ) subplot ( 312 ) vo252 = Series ( rolling_std ( DAX [ ' RET ' ] , 252 )* sqrt ( 252 ) , index = DAX . index ) vo252 . plot (); grid ( True ); ylabel ( ' Vola ( 252d Mov ) ' ) x , y = REG ( vo252 , 0 ); plot (x , y ); vx = axis () subplot ( 313 ) co252 = Series ( rolling_corr ( mr252 , vo252 , 252 ) , index = DAX . index ) co252 . plot (); grid ( True ); ylabel ( ' Corr ( 252d Mov ) ' ) x , y = REG ( co252 , 0 ); plot (x , y ); cx = axis () axis ([ vx [ 0 ] , vx [ 1 ] , cx [ 2 ] , cx [ 3 ]])

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

12 / 34

Data Analysis

Cross-Sectional Data

DAX cross-sectional dataimplied volatility surface5

21 20 19 Implied Volatility 18 17 16 15 14 13 12 7000 7200 7400 Strike 7600 7800 8000

maturities: 21 (red dots), 49 (green crosses), 140 (blue triangles), 231 (yellow stones) and 322 days (purple hectagons)
5

Source: http://www.eurexchange.com, 29 Apr 2011

Y. Hilpisch (Visixion GmbH) Derivatives Analytics EuroPython 2011 13 / 34

Monte Carlo Simulation

Model Economy

Model economyBlack-Scholes-Merton continuous time

economy with nal date T, 0 < T < uncertainty is represented by a ltered probability space {, F , F, P } for 0 t T the risk-neutral index dynamics are given by the SDE dSt = rdt + dZt St (1)

St index level at date t, r constant risk-less short rate, constant volatility of the index and Zt standard Brownian motion the process S generates the ltration F, i.e. Ft F (S0st ) a risk-less zero-coupon bond satises the DE dBt = rdt Bt the time t value of a zero-coupon bond paying one unit of currency at T with 0 t < T is Bt (T ) = er(T t) (2)

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

14 / 34

Monte Carlo Simulation

Model Economy

Model economyBlack-Scholes-Merton discrete time

to simulate the nancial model, i.e. to generate numerical values for St , the SDE (1) has to be discretized to this end, divide the given time interval [0, T ] in equidistant sub-intervals t such that now t {0, t, 2t, ..., T }, i.e. there are M + 1 points in time with M T /t a discrete version of the continuous time market model (1)(2) is St Stt Bt Btt
r 2
2

= =

t+

tzt

(3) (4)

ert

for t {t, ..., T } and standard normally distributed zt this scheme is an Euler discretization which is known to be exact for the geometric Brownian motion (1)

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

15 / 34

Monte Carlo Simulation

European Options

Option valuation by simulationEuropean options

a European put option on the index S pays at maturity T h(ST ) max[K ST , 0] for a xed strike price K to value such an option, simulate I paths of St such that you get I values ST,i , i {1, ..., I } the Monte Carlo estimator for the put option value then is V0 = erT 1 I
I

h(ST,i )
i=1

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

16 / 34

Monte Carlo Simulation

European Options

Simulating the index level for European option valuation6

Average Drift
50

40

30

Jan 1, 2011

Jun 15, 2011

Dec 1, 2011

20 simulated index level paths; thick blue line = average drift

Source: analytics.dexision.com
Y. Hilpisch (Visixion GmbH) Derivatives Analytics EuroPython 2011 17 / 34

Monte Carlo Simulation

European Options

Numpy oers all you need for an ecient implementation (I)

# # Valuation of European Put Option # by Monte Carlo Simulation # from numpy import * from numpy . random import standard_normal , seed from time import time t0 = time () # # Parameters -- American Put Option S0 = 36 . # initial stock level K = 40 . # strike price T = 1.0 # time - to - maturity vol = 0 . 2 # volatility r = 0 . 06 # short rate # # Simulation Parameters seed ( 150000 ) # seed for Python RNG M = 50 # time steps I = 50000 # simulation paths dt = T / M # length of time interval df = exp ( - r * dt ) # discount factor per time interval

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

18 / 34

Monte Carlo Simulation

European Options

Numpy oers all you need for an ecient implementation (II)

# # Index Level Path Generation S = zeros (( M +1 , I ) , ' d ' ) # index value matrix S [0 ,:]= S0 # initial values for t in range (1 , M +1 , 1 ): # stock price paths ran = s ta ndard_normal ( I ) # pseudo - random numbers S [t ,:]= S [t -1 ,:]* exp (( r - vol ** 2 / 2 )* dt + vol * ran * sqrt ( dt )) # # Valuation h = maximum (K - S [ - 1 ] , 0 ) V0 = exp ( - r * T )* sum ( h )/ I # inner values at maturity # MCS estimator

# # Output print " Time elapsed in Seconds % 8 . 3f " %( time () - t0 ) print " - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - " print " European Put Option Value % 8 . 3f " % V0 print " - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - "

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

19 / 34

Monte Carlo Simulation

American Options

American optionssolving optimal stopping problems (I)

to value American options by Monte Carlo simulation, a discrete optimal stopping problem has to be solved: V0 = sup
{0,t,2t,...,T }

EQ 0 (B0 ( )h (S ))

(5)

it is well-known that the value of the American option at date t is then given by Vt (s) = max[ht (s), Ct (s)] (6)

i.e. the maximum of the payo ht (s) of immediate exercise and the expected payo Ct (s) of not exercising; this quantity is given as
r t Vt+t (St+t )|St = s) Ct (s) = EQ t (e

(7)

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

20 / 34

Monte Carlo Simulation

American Options

American optionssolving optimal stopping problems (II)

problem: given a date t and a path i, you do not know the expected value in (7)you only know the single simulated continuation value Yt,i solution of Longsta and Schwartz (2001): estimate the continuation values Ct,i by ordinary least-squares regressiongiven the I simulated index levels St,i and continuation values Yt,i (use cross section of simulated data at date t) their algorithm is called Least Squares Monte Carlo (LSM)

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

21 / 34

Monte Carlo Simulation

American Options

The LSM algorithmregression for American put option7

Source: analytics.dexision.com
Y. Hilpisch (Visixion GmbH) Derivatives Analytics EuroPython 2011 22 / 34

Monte Carlo Simulation

American Options

The LSM algorithmbackwards exercise/valuation8

Source: analytics.dexision.com
Y. Hilpisch (Visixion GmbH) Derivatives Analytics EuroPython 2011 23 / 34

Monte Carlo Simulation

American Options

Again, the Python/Numpy implementation is straightforward (I)

# # Valuation of American Put Option # with Least - Squares Monte Carlo # from numpy import * from numpy . random import standard_normal , seed from matplotlib . pyplot import * from time import time t0 = time () # # Simulation Parameters seed ( 150000 ) # seed for Python RNG M = 50 # time steps I = 4 * 4096 # paths for valuation reg = 9 # no of basis functions AP = True # antithetic paths MM = True # moment matching of RN # # Parameters -r = 0 . 06 vol = 0 . 2 S0 = 36 . T = 1.0 V0_right = 4 . 48637 dt = T / M df = exp ( - r * dt ) American Put Option # short rate # volatility # initial stock level # time - to - maturity # American Put Option ( 500 steps bin . model ) # length of time interval # discount factor per time interval

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

24 / 34

Monte Carlo Simulation

American Options

Again, the Python/Numpy implementation is straightforward (II)

# # Function Definitions def RNG ( I ): if AP == True : ran = standard_no rm a l ( I / 2 ) ran = concatenate (( ran , - ran )) else : ran = standard_no rm a l ( I ) if MM == True : ran = ran - mean ( ran ) ran = ran / std ( ran ) return ran def GenS ( I ): S = zeros (( M +1 , I ) , ' d ' ) # index level matrix S [0 ,:]= S0 # initial values for t in range (1 , M +1 , 1 ): # index level paths ran = RNG ( I ) S [t ,:]= S [t -1 ,:]* exp (( r - vol ** 2 / 2 )* dt + vol * ran * sqrt ( dt )) return S def IV ( S ): return maximum ( 40 . -S , 0 )

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

25 / 34

Monte Carlo Simulation

American Options

Again, the Python/Numpy implementation is straightforward (III)

# # Valuation by LSM S = GenS ( I ) # generate stock price h = IV ( S ) # inner value matrix V = IV ( S ) # value matrix for t in range (M -1 , -1 , - 1 ): rg = polyfit ( S [t ,:] , V [ t +1 ,:]* df , reg ) # C = polyval ( rg , S [t ,:]) # V [t ,:]= where ( h [t ,:] > C , h [t ,:] , V [ t +1 ,:]* df ) # V0 = sum ( V [0 ,:])/ I # LSM estimator

paths

regression at time t continuation values exercise decision

# # Output print " Time elapsed in Seconds % 8 . 3f " %( time () - t0 ) print " - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - " print " Right Value % 8 . 3f " % V0_right print " - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - " print " LSM Value for Am . Option % 8 . 3f " % V0 print " Absolute Error % 8 . 3f " %( V0 - V0_right ) print " Relative Error in Percent % 8 . 3f " %(( V0 - V0_right )/ V0_right * 100 ) print " - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - "

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

26 / 34

Monte Carlo Simulation

Speed-up of 480+ Times

The Challengedozens of minutes in Matlab

realistic market models generally include multiple sources of randomness which are possibly correlated the simulation of such complex models in combination with Least Squares Monte Carlo is computationally demanding and time consuming in their research paper, Medvedev and Scaillet (2009) analyze the valuation of American put options in the presence of stochastic volatility and stochastic short rates Medvedev and Scaillet (2009) write on page 16: To give an idea of the computational advantage of our method, a Matlab code implementing the algorithm of Longsta and Schwartz (2001) takes dozens of minutes to compute a single option price while our approximation takes roughly a tenth of a second.

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

27 / 34

Monte Carlo Simulation

Speed-up of 480+ Times

The Resultsonly seconds in Python

Python is well-suited to implement ecent, i.e. fast and accurate, numerical valuation algorithms
MCS/LSM with 25 steps/35,000 paths: 180 megabytes of data crunched in 1.5 seconds MCS/LSM with 50 steps/100,000 paths: 980 megabytes of data crunched in 8.5 seconds

reported times are from my 3 years old notebook ... the speed-up compared to the times reported in Medvedev and Scaillet (2009) is 480+ times (1.5 seconds vs. 720+ seconds) to reach this speed-up, our algorithm mainly uses variance reductions techniques (like moment matching and control variates) which allows to reduce the number of time steps and paths signicantly

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

28 / 34

Monte Carlo Simulation

Speed-up of 480+ Times

Results from 3 simulation runs for the 36 American put options of Medvedev and Scaillet (2009)
---------------------------------------------------Start Calculations 2011-06-22 13:43:02.163000 ---------------------------------------------------Name of Simulation Base_3_25_35_TTF_2.5_1.5 Seed Value for RNG 150000 Number of Runs 3 Time Steps 25 Paths 35000 Control Variates True Moment Matching True Antithetic Paths False Option Prices 108 Absolute Tolerance 0.0250 Relative Tolerance 0.0150 Errors 0 Error Ratio 0.0000 Aver Val Error -0.0059 Aver Abs Val Error 0.0154 Time in Seconds 135.7890 Time in Minutes 2.2631 Time per Option 1.2573 ---------------------------------------------------End Calculations 2011-06-22 13:45:17.952000 ----------------------------------------------------

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

29 / 34

DEXISIONOur Analytics Suite

Capabilities

DEXISION can handle a number of nancial derivatives ranging from plain vanilla to complex and exotic

Example products:
plain vanilla options American options Asian options options on baskets certicates (bonus, express, etc.) swaps, swaptions real options portfolios of options life insurance contracts

Example underlyings:
indices stocks bonds interest rates currencies commodities

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

30 / 34

DEXISIONOur Analytics Suite

Capabilities

DEXISION can be benecially applied in a number of areas

nancial research: researchers, lecturers and students in (mathematical) nance nd in DEXISION an easy-to-learn tool to model, value and analyze nancial derivatives nancial engineering: nancial engineers and risk managers in investment banks, hedge funds, etc. can use DEXISION to quickly model and value diverse nancial products, to cross-check valuations and to assess risks of complex derivatives portfolios actuarial calculations: those responsible for the design, valuation and risk management of market-oriented insurance products can engineer, value and test new and existing products easily nancial reporting: IFRS and other reporting standards require the use of formal (option) pricing models when there are no market prices; DEXISION considerably simplies the modelling, valuation and risk assessment for illiquid, complex, non-traded nancial instruments and embedded options real options valuation: DEXISION oers unique capabilities to account for the specics of real options (as compared to nancial options)

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

31 / 34

DEXISIONOur Analytics Suite

Technology

DEXISION is based on a Python-LAMP environment and makes heavy use of Numpy

Suse Linux 11.1 as 64 bit operating system Apache 2 as Web server MySQL 5.0.67 as relational database Python 2.6 as core language (integrated via mod_python in Apache) Numpy 1.3.0 as fast linear algebra library Dojo 1.0 as JavaScript framework for the GUI SVG for all custom graphics MoinMoin Wiki (Python powered) for Web documentation

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

32 / 34

DEXISIONOur Analytics Suite

Technology

Our aim is to make DEXISION the Google of Derivatives Analytics

recently, Visixion added Web services to DEXISIONs functionalities which allow to integrate it into any environment once a structure is modeled in DEXISION, updates of valuations can be received in real-time via these Web services (with data delivered e.g. in XML format) during the Web service call, data/variables can also be provided a call to value an American put option on the DAX index could look like: https://company.dexision.com/DEXISIONeval.py?company=X&user=Y&pwd= Z&paths=50000&steps=150&portfolio=DAX/DAX_Am_Put_Dec_2011&DAX_ current=7200&DAX_vola=0.175&rate=0.03&strike=6800

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

33 / 34

DEXISIONOur Analytics Suite

Technology

Contact

Dr. Yves J. Hilpisch Visixion GmbH Rathausstrasse 75-79 66333 Voelklingen Germany www.visixion.com Derivatives Analytics and Python Programming www.dexision.com Derivatives Analytics On Demand E [email protected] T/F +49 3212 1129194

Y. Hilpisch (Visixion GmbH)

Derivatives Analytics

EuroPython 2011

34 / 34