# Using Scala for Payoff Scripting

The advantages of payoff scripting based on a build-in interpreter or “on-the-fly compiler” instead of implementing the payoffs in C++ are obvious. Faster time-to-market because there is no need to recompile and deploy the C++ pricing library and people without deep C++ knowledge are able to develop and test new structured products. One disadvantage is often the execution speed of the chosen scripting language. Examples of languages I have seen/used for payoff scripting are Python  (C++ interface boost::python), Lua, tinycc and CINT. When it comes to execution speed none of these are suited to build high performance solutions, see. e.g. [1].  This is especially true if the Monte-Carlo scenario generator is running on a GPU. The payoff scripting on the CPU can then easily become the bottleneck of your pricing library.

Scala is a modern programming language that integrates object-oriented and functional language features. The Scala compiler generates byte code for the Java VM. Therefore the execution speed of a Scala script is comparable with Java and roughly a factor of two slower than C++ [1].

The Scala compiler itself is a Scala object and can be used at runtime to compile and link new scripts or classes. In addition using JNI it is fairly easy to attach a Java VM to a C++ process and to exchange data between C++ and Scala. Also Scala offers a lot of features to design an “internal”, user-friendly domain specific language (DSL) for payoff scripting.

The code for a small QuantLib/Scala Monte-Carlo simulation in action is available here. It depends on QuantLib 1.0 or higher, a Java 1.6 VM and Scala 2.8/9. Overwrite the PayoffImpl.scala class to implement different payoffs without recompiling the C++ code.

# VPP Pricing IV: Variance Reduction for Perfect Foresight

Even though perfect foresight provides only an upper bound to the real VPP value the differences are often neglectable and the implementation efforts  are small compared with “exact” pricing based on finite difference methods or least square Monte-Carlo. Perfect foresight is the method of choice in conjunction with a linear programming optimizer if the problem contains time-integral constraints. Therefore it is worth to test the efficiency of two standard variance reduction techniques, namely antithetic sampling and quasi Monte-Carlo (QMC) together with a Brownian Bridge. Both methods are explained in [1], antithetic sampling in chapter 4.2 and quasi Monte-Carlo in section 5. Randomized QMC is used to calculate the error estimates for QMC as it is outlined in chapter 5.4.

Using the parameterization of the previous section VPP Pricing III, QMC in conjunction with a Brownian Bridge clearly out-performance the other algorithms for a 6 month contract as can be seen in the diagrams below. The code is available here. It depends on the latest QuantLib version from the SVN trunk or the upcoming QuantLib 1.2 release. If you want to generate the plots you’ll also need R.

[1] P. Glasserman, Monte Carlo Methods in Financial Engineering.  ISBN-0387004513

# VPP Pricing II: Mixed Integer Linear Programming

The next two steps are defining a simple VPP contract (or a simplified gas-run power plant) and setting up a mixed integer linear programming optimization (MIP) to calculated the intrinsic value and an upper bound for the extrinsic value based on a Monte-Carlo simulation and assuming perfect foresight. The third step outlined in the next part will then be the “exact” pricing of the extrinsic value using dynamic programming and finite difference methods.

The set-up of the simplified gas-run power plant is similar to the one explained in chapter 4.2.3 of the text-book [1]. In general the power plant has three power output level:

• Plant is off, $P=0$
• Generation at minimum load $P=P_{min}$
• Generation at maximum load $P=P_{max}$

The power plant has a fixed efficiency rate

$\zeta=\frac{MWh Power Output}{MWh Heat}$.

Ramp rates will be neglected, but the power plant has a minimum uptime $t_{up}$ and a minimum downtime $t_{down}$. The start-up costs are given by a fixed start-up cost $\eta$ (in €) and the price of the gas needed to produce the start-up heat $\theta$ (in MWh).

The mixed integer linear optimization is running in one hour blocks and is using three decision variables per hour $i$. The binary decision variable $\beta_i$ is true if the power plant is running at minimum load $P_{min}$ or at maximum load $P_{max}$ and $\beta_i$ is false if the plant is off. The real decision variable $0\le s_i \le 1$ is equal to one if the plant is started in hour $i$, which is implied by the following constraint

$\beta_i - \beta_{i-1} \le s_i$.

The minimum up-time $t_{up}$ and the minimum down-time $t_{down}$ is a consequence of the constraints

$\begin{array}{rcl} \beta_i &\ge& \sum_{t=i-t_{up}+1}^{t=i} s_t \\[7pt] \beta_i &\le& 1-\sum_{t=i+1}^{t=i+t_{down}} s_t \end{array}$

The real decision variable $0\le \gamma \le 1$ is equal to one if the power plant is running at maximum load $P_{max}$ and zero if the power plant is either running at minimum load $P_{min}$ or if the plant is off, that means

$\beta_i \ge \gamma_i$

Let $P_i$ be the power price, $G_i$ be the gas price and $CO_2^i$ be the carbon dioxide price at hour $i$. The objective function is then given by

$P\& L = \sum_{t=1}^N\left[\left(\gamma_iP_{max} + P_{min}(\beta_i-\gamma_i)\right) \left(P_i - \frac{G_i+CO_2^i}{\zeta}\right) - s_i\left(\eta + \theta (G_i+CO_2^i)\right)\right]$

For a one year span the problem consists of $3\cdot 365 \cdot 24 = 26280$ decision variables $\{\beta_i, \gamma_i, s_i\}$ and $4 \cdot 365\cdot 24 = 35040$ constraints. This comparable small problem can be solved using e.g. the Gnu Linear Programming Kit (GLPK). For an overview on open source linear/mixed integer programming solver see [2].

The model parameters and the example forward curves are outlined in the previous entry VPP Pricing I. The diagram below shows the intrinsic value and the upper bound for the total value (intrinsic plus extrinsic value) based on Monte-Carlo, perfect foresight and MIP for different  power plant efficiencies $\zeta$. The parameters of the VPP contract are given by

$t_{up}=t_{down}=2h, P_{min}=8MW, P_{max}=40MW, \eta=300 EUR, \theta=20MWh$,

the (fixed) carbon dioxide price is 3.0€ per MWh heat.

The source code is available here. It depends on GLPK and the latest QuantLib version from the SVN trunk or the next QuantLib 1.2 release.

It is now quite easy to add and price time-integral constraints, e.g. the following constraint restricts the number of starts within a year to be less than or equal to a given number

$\sum_{t=1}^N s_i \le \#Starts$.

The following diagram shows the results for $\#Starts \le 25$ and a minimum load $P_{min}=25MW$.

The source code is available here. It depends on QuantLib 1.1 and if  you want to generate the plot directly from the C++ program you’ll also need R, RCPP and RInside.

[1] M. Burger, B. Graeber, G. Schindlmayr, Managing Energy Risk, ISDN 978-0-470-ß2962-6

# Swing Option: Linear vs. Dynamic Programming II

In order to get an improved impression on the differences between Monte-Carlo/linear programming and  finite difference methods/dynamic programming  a more realistic swing option with hourly payoff profile is evaluated based on a German hourly forward curve. The forward curve is taken from the Kyos example download page. The parameterization of the Kluge model is outlined in [1].

With maturity of twelve weeks the example swing option provides $12*7*24=2016$ exercise opportunities. The number of overall exercised swing opportunities is constraint by

$100 \le \sum_{i=1}^{2016} \beta_i \le 500$.

The size of two dimensions of the finite difference method (dynamic programming) is therefore already been given. 2016 steps are needed in time direction and 500 steps are needed in the “consumed exercises” direction. Together with the two other directions – one for the power price and one dimension for the jump process – and without further simplidsfications this forms a pretty large finite difference problem. The Monte-Carlo based linear programming approach reduces the computational burden but will lead to an upper bound of the swing option price. The diagram below shows the corresponding results.

The code is available here. It depends on the GNU Linear Programming Kit, the Boost Thread library for parallelization and at the time of writing on the latest QuantLib version from the SVN trunk. If you want to generate the plot directly out of the C++ program you also need R, RCPP and RInside. To utilize all CPU cores please use the -DNTHREADS=(number of CPU cores)  compiler switch. In addition to run the program you have to download the forward curve German power from the Kyos web page and convert it into a text file EEX_2010-2013.txt of the folliwing format

4-Oct-2010 36.73 32.09 27.32 23.22 25.71 35.60 47.32 …

5-Oct-2010 38.12 31.12 22.76 25.65 27.87 34.60 50.01 …