After people have had time to digest the raw, theoretical value of full second order risk, the immediate question is "Sure you can find examples on small trades, but what does it look like on a large portfolio?".
There is now a simple tool online to analyze various types of portfolios for different days. We of course would like to direct everyone with midcurve risk to look at the that portfolio over the historical dates. Midcurves are complicated products, as they are vanilla in nature but modeled using a correlated multi-asset model and the relationship between the assets (the best hedge portfolio) depends on curve shape, vol surface shape, and level of both. We use a correlated Gaussian distribution assumption since the vol surface in the tool is a flat normal surface.
But here at PLATSON we never stop asking questions, because it is our goal to highlight the shortcomings in current risk management now that we have a tractable solution. So we built a full portfolio model. We created a yield curve set, with one index curve and one discount curve, each with 14 points (a mixture of deposit rates and swap rates). And we created a vol surface with 126 (14 expiries, 9 tenors) points, defined as a flat normal volatility surface. There are no complications moving to SABR, but we choose to keep models as straight forward as possible so clients can try to replicate our results.
We then generated a randomized portfolio of swaps, swaptions, and midcurve options. The only restrictions were that the total tenor of the swap must be less than 30 years (the maximum of our yield curve), swaptions must be inside the 15y15y point (the bottom right of our grid), and midcurves have forward starts of 6m to 5 years and final tenors of 1 to 5 years. Our midcurve model used a standard correlated Gaussian distribution and the correlations were set at reasonable levels.
After generating this random portfolio, we used traditional risk metrics to hedge the delta (at the bucket level) and vega (using macro hedging on a reduced grid of "liquid" points, to simulate a modern market. In total we had 226 swaps, 228 swaptions, and 80 midcurves.
The initial Delta, as expected, is near 0. The Gamma is a complicated set of numbers, but net, is very small. In fact, hidden from traditional risk is how much cross gamma goes unquantified.
The vega, while net is small, has significant bucket risk. Our experience says this is quite standard, as it is near impossible to hedge at the individual bucket level. Oddly, we see something that regularly occurs in our experience: a vega butterfly occurring between a low liquidity area (12y tails) and the surrounding liquid points, 10y and 15y. This was not planned in advance, though the hedging algorithm we use may favor such an outcome quite strongly.
We then randomized a market move. We attempted to make the moves meaningful (how boring would a 0 move day be for risk assessment?) without being excessive. We bounded the moves in ways we believe would achieve this, and have:
We ended with a modest bear steepening, with some volatility in the libor/ois basis, with vol dropping a relatively parallel 2-3 bp vols. In our opinion, there is nothing surprising about this move, and it is a good test of modern risk management metrics. We have seen many days this kind of move is indicative of, and most risk managers would expect good understanding of book dynamics on these days.
How did risk actually change over the day?
The Delta change is not trivial. The 15y bucket has significant shift of risk from the forward to the discount curve, implying significant new basis swap exposure (and this is offset by the longer buckets) exposing a book to significant shape risk in the LOIS curve, something we personally have experience struggling with in the JPY and AUD markets.
The vega shifts are even more severe in the back end, replicating the risks that many banks now run with structured products.
How do traditional metrics perform when analyzing the delta? Let's look at the prediction error you would find the next morning to get a feel.
Pretty terribly when it comes to the back end risk. In fact, almost all the features of the risk we were most worried about are errors when looking at traditional metrics. I know as a trader, waking up to an extra 40-50k of long end risk was always cause for concern, especially when trading currencies like AUD where liquidity in the long dated swaps was hard to come by.
But how about PLATSON? Is it any better? Because we not only quantify the cross gamma (important for midcurves) but also the vanna and volga(important for long dated options), it is significantly better.
There is no perfect risk system, of course. But I can say with certainty I never worried too much about 2-3k of delta, especially since that is well below the minimum traded increment in most markets.
PLATSON passes the delta test as well as can be hoped. The worst section is about 2.5k of 12y/15y/25y butterfly, significantly better than 50k of 25y outright risk.
But what about Vega? Most traders have indicated it is the difficulty of quantifying vega changes on day that most trouble them. The cost of hedging delta is low, but vega is both expensive and difficult to manage and having early warning of significant Vega changes would better allow them to move risk to customers (via sharing axes with sales) or be part of suitable trades in the interbank.
Reducing errors in Vega prediction of hundreds of thousands of dollars per bp vol is significant. If it even allows you the chance to save just 1/10th of a bp vega on the risk you are moving, the risk could easily be worth, to a single book, hundreds of thousands of dollars a year.
Of course, our experience says you only know you are caught the wrong way around when the PnL posts late in the evening, well after most chances to get back onside are long gone. What would you see? Your spreadsheets would show:
Finance would surprise you with a 250k hole in your PnL:
And those people on the PLATSON method would see:
Want to get your risk and PnL right on the big days?
Contact us today to learn more about the PLATSON solution and how we can generate these risks with almost no increase in computational power required. Our optimizations are unique, our runtimes even on small portfolios are comparable, and the savings could begin to be realized from day 1.
Kommentare