A Parsimonious Framework for Pricing Future Generation Structured Notes The objective is to build a versatile framework that can be quickly readjusted to price all kinds of complex structured notes, including interest rate-, FX-, equity-, and commodity-linked notes. Possibly with path-dependent features. Possibly with callable features.This sounds like Holy Grail. However, here are some mitigating circumstances:a) it is not a front-office tool, it is a middle-office tool to run infrequently (say, once a week) in order to tune a very rough approximations used for future risk exposure computations, to improve collateral management, and to check on front-office pricing systemb) thus, 3% relative error in pricing is, perhaps, acceptablec) thus, precision on sensitivities is not requiredd) this tool will only be run for “problem” structures, as a matter of reconciliation between different systems.The framework should be sufficiently versatile and economical, as resources to build and thencustomize, calibrate and run the tool are very limited. Bottom line: when new type of structure, say FX-based targeted redemption note, comes along customization of pricing function should be straightforward and quick.Here are several model design decisions and issues I want to discuss but I welcome any comment from the Quant community.Model 1. For short-term LIBOR rate and/or commodity price: Several forward price processes on the underlying asset are run simultaneously. As forward matures, its price is fixed thereafter. SABR is adopted for each forward process until maturity with parameters calibrated to volatility smile at maturity point (where available). All correlations between forwards, their volatilities, and cross are historical and fixed For FX: spot lognormal process: Run with 2 short-term rates for 2 currencies, see the process above, with correlated diffusion terms. FX drift at each step of simulation is determined by the simulated short rate differential. FX instantaneous volatility is fixed and along with all correlations is determined from history. For equity: spot lognormal process: with r-q drift and historical volatility uncorrelated to fixed income risk factors or spot SABR process: if “rich” dataset of implied volatility surfaces is available for calibrationMy impression is that the alternatives are either not versatile enough or very expensive to calibrate properly (like local volatility model, if developed from scratch). What do you think?Model 2. Sub-optimal exercise decision in the space of Monte-Carlo scenarios generated by the defined above simulation process. Longstaff-Schwartz technique used with quadratic polynomial functions over the total set of simulated risk factors.Issues in question:1) Should BGM drift formula change when volatility becomes SABR-like ?2) Is there any consensus on how to extrapolate volatility of volatility outside the time framewhere the smile is available to calibrate SABR? Any particular parametric form?3) Are there better sub-optimal exercise techniques than Longstaff-Schwartz, and in particular for SABR-like simulations of risk factors? Any pointers here?4) Any papers on how well this technique works for path-dependent callables, also what about the degree of polynomials used here, do quadratic polynomials work ok ? Thank you

QuoteThe objective is to build a versatile framework that can be quickly readjusted to price all kinds of complex structured notes, including interest rate-, FX-, equity-, and commodity-linked notes. Possibly with path-dependent features. Possibly with callable features.My first question would be: do you have a budget? Euphemistically put, it seems to be awful ambitious.Is this a s/w framework? How many people? Do you - for example - have a senior C++ /C#/Java designer/developer for this?Can your quantify words like 'quickly', 'versatile'. I think these are at odds with 'parsimonious'QuoteThis sounds like Holy Grail. However, here are some mitigating circumstances:a) it is not a front-office tool, it is a middle-office tool to run infrequently (say, once a week)in order to tune a very rough approximations used for future risk exposure computations,to improve collateral management, and to check on front-office pricing systemSmall software and innocent-looking software systems have a habit of 'blossoming' into huge systems. Be careful.The mythical man month

Ambitious indeed! You might get some ideas by looking at a couple of commercial products (I have no affiliation of any kind with the vendors):FEA's Derivatool: http://www.barra.com/products/derivatoo ... =1Summit's MUST: http://www.misysbanking.com/Misys_Banki ... ex.htmlI'd hazard a guess that such a project would be a black hole from which it would take years to emerge. Best of luck!

Yes, thank you. I am familiar with MUST, Numerix, and some other stuff.I appreciate the concern. Not to worry! I am talking about core computational engine in MatLab,no interface, no overhead. I anticipate, simulation framework and Longstaff won't take much to implement in MatLab. Calibration routines would take the most time, I think.Right now, I am more concerned with the overall project design and some theoretical issues.I have little but sufficient resources, and it is a long-term project with several stages.

QuoteOriginally posted by: BoryaRight now, I am more concerned with the overall project design and some theoretical issues.And so you should be; get the design wrong now and you are in the black hole described by DavidJN.QuoteI have little but sufficient resources, and it is a long-term project with several stages.Does this mean you do or do not have sufficient resources? If you have, I can humbly offer may services to do an OO design for you, then no more headaches.

Dear Cuchulain,Thank you for the offer. I prefer to do my own design but I leave the door open to show me that I am not capable.Hence, the post. Is there anything wrong with the scheme I proposed in my initial post?Devil is in details. For example, here are some details for IR modeling.1) Volatility smiles for caplets are bootstrapped from caps in the manner described by Carol Alexander in the March 2003 issue of Wilmott magazine2) SABR model parameters are calibrated to each smile separately using the formulas given by Pat hagan and the others in September 2002 issue of Wilmott. Forward Libor measure is used for each forward short rate (different measures for different rates).3) After this is done, BMG drifts are added to the equation for forward short rates to bring them to the same measure, but volatilities, calibrated above, are not changed.4) After this is done, Monte-Carlo simulation is performed, predictor-corrector is used5) After this is done, Longstaff procedure is performed with the given exercise payoff function for a simulation path with quadratic polynomials.It looks straightforward to me. I want to anticipate the problems. Thanks again.

BoryaQuoteIt looks straightforward to me. I want to anticipate the problems. Thanks again. I thnk the devil will get into the details when you get to the data level. For example, creating a generic MC engine (is that the goal BTW) will be a bit tricky because each problem has its own peculiarities. Having gotten one model finished, then go the next one but things need to be modified. That's life! Edit: Borya, do you plan to write MC engine?LIBOR?

Yes, I plan to write generic MC for multidimensional SABR, every time I have volatility fixed (for equity or FX) I'll just makevol of vol equal to 0. No local volatility.I will use it for LIBOR, commodity, FX, and equity.What do you think?