Danske Bank

Stabilize risks of discontinuous payoffs

with Fuzzy Logic
Antoine Savine
 Danske Bank

Discontinuous payoffs with Monte-Carlo

• Discontinuous profiles produce unstable risks with Monte-Carlo

− Example: 110 1y Digital in Black-Scholes MC, vol = 20%
− 10,000 to 250,000 simulations
− Seed changed between pricings
− Decent convergence on price, risk all over the place

Fuzzy Logic 2
 Danske Bank

Malliavin’s calculus

• The industry developed 2 families of methods to deal with this problem:

• One is the Likelihood Ratio Method (LRM) and more general Malliavin’s calculus
− Apply
 
ai   
Ea  f X  
 ai  f  X    a , X  dX
X

a  a , X 
   
  f X ai a , X dX   f X   i
  a , X  dX
X X  a, X 
     
  f X log  a , X   a , X dX  E f X log  a , X 
X
  ai  
 ai     
− Compute log   a, X  a pathwise
i

− No need to differentiate a discontinuous payoff, differentiate log-likelihood instead

− Tractability, efficiency and stability depend on model

Fuzzy Logic 3
 Danske Bank

Payoff smoothing

• Payoff smoothing is considerably simpler (and surprinsingly effective):

Replace discontinuous payoffs by ”close” continuous ones
− Digital  Call spread
− Barrier  Smooth barrier

• Benefit: work on payoffs only, no work required on models

• Problem: to identify (and smooth) all discontinuities
we need to know exactly how the payoff is calculated
• Today, this is a manual process: smooth(x,p,n,e)
=
Traders rewrite payoffs one by one
p
using a ”call-spread” or ”smooth” function:
n

x
-e/2 0 e/2

Fuzzy Logic 4
 Danske Bank

Smoothing a digital

• Digital  Call Spread

C
• Dig =   K
 CS = finite difference continuous approximation
• Alt. interpretation: spread the notional
evenly across strikes between (K-e/2,K+e/2)

Fuzzy Logic 5
 Danske Bank

Smoothing a barrier

• Spread notional across barriers in (B-e/2,B+e/2)

total KO
• Example: 1y 110-120 RKO in Black-Scholes (20%)
no KO
• Weekly monitoring, barrier shifted by 0.60 std partial KO

• Smooth barrier: knock-out part of the notional

depending on how far we land in (B-e/2,B+e/2)

Fuzzy Logic 6
 Danske Bank

Smoothing everything

• How can we generalize digital/barrier smoothing to any payoff?

• We need to identify (and smooth) all discontinuities

• Hence, we need to know exactly how the payoff is calculated
• We need access to the ”source code” of the payoff

• ”Hard coded” payoffs only allow evaluation against scenarios: ”compiled code”
• But scripted payoffs provide full information: ”source code”

Fuzzy Logic 7
 Danske Bank

Scripting Languages

• Simple ”programming” language optimized for the description of cash-flows

• Describe the cash-flows in any transaction: vanillas, exotics, even CVA/KVA
• And then evaluate the script with some model
• Example: simplified Autocallable
− If spot declines 3 years in a row, lose 50% of notional
− Otherwise make 10% per annum until spot raises above initial level

Fuzzy Logic 8
 Danske Bank

The many benefits of scripting cash-flows

• Run-time pricing and risks of transactions

− Users create/edit a transaction by scripting its cash-flows
− Then produce value and risk by sending the script to a model

• Homogeneous representation of all cash flows in all transactions

− Software can manipulate, merge and compress cash-flows across transactions
− Crucial for ”derivatives on netting sets” like VAR/CVA/XVA/KVA/RWA

• Crucially for us, provide ”source code” cash-flow information to the software
− So the software can, among (many) others:
− Obviously, evaluate cash-flows in different scenarios
− Detect contingency, path-dependence or early exit and select the appropriate model
− Optimise the calculation of value and risk sensitivities
− Detect (and smooth) discontinuities

Fuzzy Logic 9
 Danske Bank

Identifying discontinuities

• In programming, incl. scripting, discontinuities come from control flow

− If this then that pattern
− Call to discontinuous functions (meaning themselves implement if this then that statements)

• Autocallable:
− 3 control flows
− Unstable risk

Fuzzy Logic 10
 Danske Bank

Manual smoothing

• Replace if this then that by smooth

• Does the job, but hard to write and read and error prone, even in simple cases
Fuzzy Logic 11
 Danske Bank

Automatic smoothing

• Automatically turn
− Nice, easy, natural scripts

− Into smoothed scripts that produce stable risks

• We have access to the script  we can identify control flows

• We only need to smooth them all

Fuzzy Logic 12
 Danske Bank

Fuzzy Logic to the rescue

• Turns out we don’t need to transform the script at all!

• We just need to evaluate differently
the conditions and the conditional statements
• In classical (sharp) logic, a condition is true or false
− Schrodinger’s cat is dead or alive

• In fuzzy logic, a condition is true to a degree = degree of truth or DT

− Schrodinger’s cat is dead and alive
− For instance, it can be 65% alive and 35% dead

• Note: DT is not a probability

 Financial interpretation: condition applies to DT% of the notional
− A barrier is 65% alive means 35% of the notional knocked out, the rest is still going
− It does not mean that it is hit with proba 35%, which makes no sense on a given scenario

Fuzzy Logic 13
 Danske Bank

Fuzzy Logic

• Invented in the 1960s

• Not to smooth financial risks
• But to build the first expert systems

• Automate expert opinions

− Expert: it is dangerous for the engine to go fast and be hot
− Programmer: what do you mean by ”fast” and ”hot”?
− Expert: say fast = ”>100mph” and hot = ”>50C”
Code: if speed > 100 and heat > 50 then danger = true else danger = false endIf
Makes no sense: (99.9,90)  safe, (100.1,50.1)  dangerous !!

Fuzzy Logic 14
 Danske Bank

More Fuzzy Logic

• With fuzzy logic:

− isFast and isHot are DTs = smooth functions resp. of speed and heat valued in [0,1]

DT C1 and C2   DT1  DT2 DT C1 and C2   min  DT1 , DT2 

− DTs combine like booleans alt.
DT C1 or C2   DT1  DT2  DT1  DT2 DT C1 or C2   max  DT1 , DT2 

Code: danger = fuzzyAnd( isFast( speed), isHot( heat))

Results make sense: (99.9,90)  50% dangerous, (100.1,50.1)  25% dangerous

Fuzzy Logic 15
 Danske Bank

Back to payoff smoothing

• Digital smoothing = apply fuzzy logic to if spot>strike then pay 1

• Barrier smoothing = apply fuzzy logic to if spot>barrier then alive = 0

• Classical smoothing exactly corresponds to

the replacement of sharp logic with fuzzy logic in the evaluation of conditions

• Reversely, we can systematically apply fuzzy logic to evaluate all conditions

• Fuzzy logic for control flow effectively removes discontinuities
• And stabilises risk sensitivities

Fuzzy Logic 16
 Danske Bank

Evaluation of ”if this then that” with fuzzy logic

• Evaluation with sharp logic

1. Evaluate condition C
2. If C is true, evaluate statements between then and else or endIf
3. If C is false, evaluate statements between else and endIf if any

• Evaluation with fuzzy logic

1. Record state S0 (all variables and products) just before the ”if” statement
2. Evaluate the degree of truth (DT) of the condition
3. Evaluate ”if true” statements
4. Record resulting state S1
5. Restore state to S0
6. Evaluate ”else” statements if any
7. Record resulting state S2
8. Set final state S = DT * S1 + (1-DT) * S2

Fuzzy Logic 17
 Danske Bank

Fuzzy Logic implementation

• No change in scripts
• Change the evaluation of if this then that statements only
• Effectively stabilises risks sensitivities with minimal PV impact
• (With proper optimization) just as fast as normal
• Easily flick between sharp (pricing) and fuzzy (risk) evaluation

Fuzzy Logic 18
 Danske Bank

Advanced Fuzzy Logic: condition domains

• Consider the script

• The variable digPays is valued in {0,1}

• To apply fuzzy logic / call spread to digPays > 0 makes no sense
• Evidently the DT of digPays > 0 is digPays itself

• This is easy to correct but we must first compute the domains of all conditions
• Fortunately, we have access to the script
• We can write code that traverses the script and calculates condition domains
• This is not as hard as it sounds

Fuzzy Logic 19
 Danske Bank

Advanced Fuzzy Logic: beware the cat

• We are used to conditions being true or false

− Schrodinger’s cat is dead or alive
− The cat cannot be dead and alive
− If the cat is alive then it can’t be dead, etc.

• We may write scripts based on such logic, like for instance:

X has to be overwritten by either 0 or 1 right?

• With fuzzy logic, where spot > 100 to a degree and <= 100 to a degree
− The logic that underlies the script is not applicable
− And the result is a bias in the value by orders of magnitude!
− With fuzzy logic, best stick to the decription of cash-flows and refrain from injecting additional ”logic”

Fuzzy Logic 20
 Danske Bank

Advanced Fuzzy Logic: what epsilon?

• Fuzzy logic applies a ”call spread” to conditions:

− Too wide  risk of valuation bias
− Too narrow  unstable risks
− What is a reasonable size?

• Typical e for a condition expr > 0 = fraction of the std dev of expr
• Alternative = set e such that Proba ( - e/2 < expr < e/2) = target

• Solution: use pre-simulations to estimate the first 2 moments of all conditions

• And set their epsilon accordingly

Fuzzy Logic 21