Risk and Decision Analysis Today

In the always high-pressure world of the equity markets, the pressure has been dialed up on one of the most intensely speculative forms of financial risk analysis, the pricing of equity options.  Earlier this year a software developer in Ireland teamed up with British hardware company to release a white paper describing what they are calling a new speed record in the pricing of options ––1 million options evaluated in 17 seconds.

When you consider what an option is and how it functions in the equity markets, this speed is mind-bending––and, not incidentally, worrisome.  An option is a contract to either buy or sell an asset at a given price.  The option itself is purchased, and this creates a market for options as well as for their underlying equities.  The seller offers an option called a put and the buyer's option is called a call.  Options have expiration dates after which they are worthless.  Between the time when an option is purchased and when it expires, it's value can fluctuate.  This makes the trading of options pretty risky, and an obvious application for any kind of Monte Carlo in Excel software.  Monte Carlo simulation is just one of a number of statistical analysis techniques that are used to value options, but it is the one used by the high-speed entrepreneurs.
 

Conveniently, for their blazing model  the British companies used what's called a European option, which must be exercised on its expiration date. This simplified their calculations by eliminating variability in time and the conditions that change over time–-and as anyone who creates financial analysis models will tell you, simplicity and speed are directly related. Nevertheless, when the acceleration software and the hardware were cranked up to full speed, the white paper reports, they were processing 30 billion option pricing iterations per second.
 

Writing about a hypothetical case in the reinsurance industry, Lina Chan and Domingo Joaquin sought to predict how a stop-loss underwriting opportunity would affect a reinsurer's bottom line. Chan, a managing partner in CP Risk Solutions, is a fellow of the Society of Actuaries, and Joaquin is an associate professor of finance at Illinois State University.
 

To create their predictions, they first established what level of loss in capital position would be unacceptable, and then, using Monte Carlo simulations in Excel, they analyzed three variations of the hypothetical underwriting arrangement.  For each version of the deal, they ran simulations using log-normal, inverse Gaussian, and log-logistic probability functions. 

 

I was surprised at sunshine-to-gloom differences in researchers' simulation results.  The gloomiest was obtained with by the model using the log-logistic function, this prompted Chan and Joaquin to endorse the reinsurance deal involving the most sharing of risk––and, of course, of profit.  But what was most striking about their study were the possible courses of action that could have resulted from the analysts' reliance on only one probability function.  By creating a multi-perspective set of risk analyses, they demonstrated how to effectively squeeze the riskiness of the hypothetical deal down to almost nothing.

» Full text of case study.  

Recently a great deal of public attention has focused on economic efficiency in the health care industry, but one rarely mentioned element of overall efficiency is the financing of hospitals. Although many hospitals are nonprofit, even these need operating capital, and an infusion of capital usually is accompanied by financial risk.  A recent article in Becker's Hospital Review highlighted the importance of financial risk analysis in the process of choosing sources of credit, and its observations should ring true for any business, for-profit or not.   

In the article, Pierre Bogacz of HFA Partners, a firm specializing in risk management for nonprofits, examines a typical decision hospitals face regularly, whether to renew an existing letter of credit (LOC) or turn to variable rate financing.  He recommends that each financing vehicle be stress-tested using a financial risk simulation that takes into account the hospital's entire balance sheet. Financial risk modeling--I assume using Monte Carlo software--is the way to calculate the risk-adjusted cost of debt, and the simulations that result can serve as a valid basis of comparison among various sources of credit. 

Monte Carlo simulation is often referred to as a computational space hog.  But how much space a simulation hogs, of course, is a matter of how much data, how many variables, and the complexity of the statistical analysis.  Climate prediction is a wonderful case in point. 

When it comes to predicting atmospheric events, the TV weather guys are pretty good at getting it right for the next few days.  That's because their forecasts are based on accurate models, and one of the factors in their accuracy is that they account for geographical space in small increments.  For short-term forecasting, the grid spacing is a few tens of kilometers.

But climate change models attempt look a hundred years ahead--as well as a hundred years back--and these truly long-term environmental risk predictions are not nearly so accurate, even with the widespread use of Monte Carlo software.  The inaccuracies result from the fact that because of the limitations on computing power at individual institutions, the present climate change models must use expanded grid spacing.  Otherwise, computation for the statistical analysis involved would overwhelm the computers trying to run the models That's the reason that Oxford University professor Tim Palmer has proposed a "global" facility to meet the computational needs of climate scientists.  "We do not," he says, have the computing power to solve the known partial differential equations of climate science with sufficient accuracy."  

Particle physicists also have the need for huge models, including those with mammoth Monte Carlo simulations, and they have met it with CERN, the European Organization for Nuclear Research. It is Palmer's idea that a parallel organization in which national climate change centers could collaborate on an international climate prediction would support significant advances in our understanding of climate change.  At this facility dedicated computer power would allow scientists to squeeze down the grid scale, still be able to run Monte Carlo software to control for approximations at this level of detail, and reveal in much higher resolution how the earth's climate is changing and how human activities affect this.  

A recent blog in Investment Week that mentioned the history of Monte Carlo simulation and its use in finance led me to take a harder look at what I thought I knew about how financial risk analysis was launched.

I had long believed that Monte Carlo simulation was developed by a team working at Los Alamos Scientific Laboratory during the 1940s.  The blog mentioned Stanislaw Ulam playing solitaire. Both turned out to be true.  Ulam was part of the team working on nuclear weapons at Los Alamos, and he prefaced his own account of his inspiration from solitaire by saying,"After spending a lot of time trying to estimate them by pure combinatorial calculations, I wondered whether a more practical method than "abstract thinking" might not be to lay it out say one hundred times and simply observe and count the number of successful plays. This was already possible to envisage with the beginning of the new era of fast computers. . . ."  He and John von Neumann began to work on the calculations that eventually became essential to the Manhattan Project. 

So far, so true.  But how did Monte Carlo simulation enter the finance arena?  The blog fast forwards thirty years to 1976 and Roger G. Ibbotson and Rex A. Sinquefield with their publication of "Stocks, Bonds, Bill, and Inflation: Simulations of the Future."

True––but not so fast.  In the intervening years and especially during the 1950s, there was considerable development and dissemination of Monte Carlo simulation technique by the U.S. Air Force and the Rand Corporation.  This brought the technique closer to the realm of finance, but we're not there yet.  

The earliest publication I can dig up on Monte Carlo and financial risk simulation is David B. Hertz's "Risk Analysis in Capital Investment," published in the Harvard Business Review in 1964.


From Harvard Business Review, circa 1964

Okay, the 1960s.  That still leaves unattended by history almost fifty years, the advent of desktop computing, the commercialization of Monte Carlo software, acceleration through parallel computing, and the wafting up on the horizon of cloud computing.

So, in the words of too many finance journals, "more research is necessary."

A recent blog for Discover magazine reports on robotics research in which the robots evolve, at least in the sense that they progressively change shape.  The research by Josh Bongard, University of Vermont,  tracks the metamorphosis of robot with a simple shape--in this case, the shape of a snake––to a more complex machine-animal shape––for Bongard's purposes, a dog.

The basic "animal" is controlled by a "nervous system", a neural network controller. This neural net  "evolves" through thousands of repetitions toward a controller that creates the most successful response––in terms of preprogrammed body shapes––to a stimulus.  The point of the project was to simulate animal adaptation using neural networks in a genetic algorithm optimization process.



Changes in shape are one way that living animals respond, over long periods of time, to changes in their habitat.  Needless to say, the neural networks used by Bongard to evolve in a much shorter time were extremely complex, and their computational demands heavy. The evolution of machine snake to machine dog required numerous computers and more than 100 CPU years.

I was fascinated by this seemingly whimsical research not only because I love animals but because I can see its potential for practical applications in business and industry––say, in solving industrial operations problems or Six Sigma and other process improvement efforts. In fact, Bongard's research on shape optimization raises the question of whether neural networks and genetic algorithm could be used to optimize the structure of an organization. 

Media reports on the  global search for alternative and sustainable energy sources often dwell in the happy realms of possibility and leave me happily clinging to a cheerful bits of information they offer up––"if everyone over the age of 21 replaced on incandescent lightbulb with a fluorescent," and blah, blah––when was the last time you read such an account that came to an unhappy conclusion or the prospect of failure? And who knows whether these bits are facts or factoids, the unreliable cousins of fact.  More important, where do the calculations in them come from?

I never stopped to think about the sources or pertinence of the peppy facts and factoids I like so much until I came across a brief blog mention of scientist Seth Darling at the U.S. Department of Energy's Argonne National Laboratory.  Darling is a photovoltaics expert who is trying to separate fact from factoid and frame a realistic picture of the costs of solar electrical generation.  He is using his Monte Carlo software to "lift up the rug" under which many assumptions about solar energy have been swept.  

Darling points out that the photovoltaics industry is expanding rapidly, with the number of its stakeholders growing in parallel: investors and funding agencies, technology developers, regulators, and policymakers.  None of these stakeholders can rely on cheerful factoids. They have to make too many decisions under uncertainty, and they need reliable information on which to base statistical analysis, risk assessment, and production predictions. Darling is trying to provide an analytical framework for testing assumptions behind solar electrical production, calculating its lifetime costs, and comparing these with conventional generation methods.  He calls this a "levelized cost of energy."  This goes beyond immediate financial risk analysis to incorporate over the lifetime of the production resource such usually hidden variables as the cost of financing, insurance, maintenance, and depreciation. 

There's a new game out there for all of you in the game of risk management. "Take Charge: A Risk Simulation Game," developed by the Indian risk management services provider Aujas, debuted at a recent summit for CSOs––chief security officers, as opposed to chief strategic officers––in Bangalore.

"Take Charge" is a team sport that is played within hypothetical business environments and marketplaces, and as the teams interact, their prowess as decision makers is evaluated. The goal for each team is to balance risk analysis with business growth.  This requires not only standard fare such as financial risk analysis but project risk management strategies incorporating such issues as information security and global accessibility.  The team that best keeps its eyes on the prize, that best understands the goals of their imaginary business and creates the best frameworks for making decisions under high levels of uncertainty is the winner. 

The point of this high-intensity play is to reveal that calculated risk is essential to growth and profit and to highlight the role of the risk officer as a key player in corporate strategy.  But many of you, dear readers, are already very much aware of your importance in strategy making. You're already out there on the real board playing the game of taking charge. 

Financial advisers took a hit from the 2008 meltdown of the markets.  Many investors, finding fault with their advisers' lack of prescience or actual handling of their investments during the crisis, decided they could do just as well managing their own investments––and they ditched their advisory firms.

 

So far their results probably haven't been bad.  For the past two years stocks have been making steady gains, so these new independents have no reason to second-guess their decisions. But a recent blog on CNBC.com put out the strong opinion that it's probably time for the investors who cut their advisers loose to swallow their pride and kiss and make up. 

 

The basic rationale behind this opinion goes something like this: in an environment with increasingly complex markets and rapid trading automated by neural networks, the everyday investor does not have the necessary skills in financial risk analysis or access to the essential risk analysis solutions to survive.  In addition, new increases in market volatility make it difficult for the amateur, without benefit of Monte Carlo software, to keep pace.  And furthermore, this free spirit most likely will not have the time or discipline to absorb and process the deluge of information the markets pour out. 

 

Overall, it's a pretty good argument, but I find this last bit––the requirement of time and discipline–-the most convincing.  If most readers are as lazy as I am, financial advisers should see a big uptick in their stock.

Last May 6, the Dow Jones Industrial Average made a rapid series of inexplicable drops, and, in fact, in one five-minute period fell more than 500 points.  Then, just as inexplicably, the market recovered.  The causes of the so-called Flash Crash remained mysterious until September, when the SEC issued a report on the rapid fluctuation of the market.  It found that a single "large fundamental trader" had used an algorithm to aggressively hedge its market position quickly.

 

Since then the role of neural networks and algorithms in automated transactions has received a good deal of attention from the media.  The online edition of this month's Wired offers a fascinating perspective on algorithms as investors.  It reveals how neural networks and other automated types of statistical analysis  can chew through news of the financial markets--essential a big pile of data-- to instantaneously produce a financial risk analysis, make a larger determination of the results of a prospective trade portfolio risk management terms, and make the trade.  The speed with which a computer can function as an investor is part of the problem. It produces a kind of feedback loop in which each instantaneous trade produces instantaneous responses from other computers trolling the markets. 

 

The trend toward computer control of financial markets, however, does not continue unfettered. The month after the Flash Crash, the SEC instituted some "circuit breakers," rules to stop trading when the feedback loops begun too intense and the markets fluctuate too rapidly.

All of this presents an interesting and larger question: How much control can we delegate to computers--not just in the financial realm but in our social and creative lives--before we have to scramble to catch up with them and regain control?   

Baby Boomers are coming face to face with the realities of retirement, and their financial advisers are having to dig deep to come up with strategies that will calm their fears of a recurrence of the financial meltdown of 2008.  In this climate, one term that comes up repeatedly is fixed income, which usually means bonds.  Here, it is interesting to note that even fixed is not as certain as it sounds.  Prices and rates of return for bonds vary over time and in opposition to those of equities.  Even given this dynamic, the challenge for bond fund managers is essentially the same as for equity fund managers--how to diversify a portfolio's holdings to minimize risk and optimize return.  

In 2008 and early 2009 the the credit risk of corporate bonds was painfully in evidence, and since then financial planners have been sharpening their credit risk management tools to stabilize returns on bond portfolios.  It has been generally accepted by investment professionals that the greater the number of financial instruments in a portfolio, the broader the spread of credit risk.  A recent credit risk analysis by the BondDesk Group found, however, that spreading risk over an increasing number of investments is effective only up to a point, after which further investments offer no further protection against loss.

 

The BondDesk Group used Monte Carlo simulation software to determine two values, tail risk (loss of 20%) and black swan risk (loss at a catastrophic level of 50% or more), in portfolios that progressively increased in size from 2 to 50 bonds. Taken together the two measures of risk predicted which bonds would default.  Interestingly, the simulations revealed that both kinds of risk were reduced by increases in portfolio sizes up to 10 bonds, and in both cases, these benefits began to diminish with bond number 11.  

 

In elementary school arithmetic, most of us who want to do well struggle to come up with the correct answer to that problem posted on the blackboard. Unfortunately, that's the way many grown-up decision makers approach risk management.  At a recent Palisade Users Conference, v.p. Randy Heffernan offered up some fun and insightful comments about risk analysis and the need many managers seem feel to boil risk assessment down to a single "correct" answer--the Number.  

 

The Number, he points out, harbors a number of evils that bedevil rational decision making. The reason for this is that risk is, by definition, uncertainty, and uncertainty is often a compound of a number of unknowns.  Uncertainty embodies many possible outcomes or answers. Trying to identify a single resolution to uncertainty leads to simplistic and often dead-wrong answers. Randy points out that the way to get the best of something as vaporous as uncertainty is to use probability or--in the case of multiple unknowns--probabilities.

 

A probability expresses a range of outcomes or numbers, and this, Randy proposes, allows the risk manager a fuller understanding of any particular course of action.  "It means thinking in two dimensions," he says, "not just 'what if' but 'how likely.'''

 

Almost every well brought up manager thinks risk analysis, the process of quantifying risk, is important.  But there is more than one approach to quantification, Randy counsels.  If you want to do really useful risk analysis, forget about coming up with The Number and concentrate concentrate on seeing the range of possibilities. 

A recent chat with Palisade customer Vertex Pharmaceuticals reinforced something I learned a few years ago when I was working with a biotechnology start-up: the development of a new drug begins with a bright idea and then enters a long, dark tunnel of uncertainty and risk.  The odds that the idea will ever emerge in the marketplace are very long, 10 to 1, and the costs of development are gi-normous--from $60 to $100 million to get a new drug even as far as phase 2 clinical trials.  But then. . . .the payout can also be gi-normous.

 

At every step in the development process, pharmaceutical risk assessment is crucial to a development company's viability.  The company has a pool of drug "candidates" in its so-called "pipeline," the pathway that leads a candidate from preclinical development through phases 1,2, and 3 of clinical trials and, with much luck and funding, into the market.  At each stage, the pharmaceutical risk management process must weigh the probabilities and potential benefits of a drug reaching the market, factor into that calculation the optimal timing of investment in development, and decide whether and when to invest in further development.  

 

 

At the heart of this study is a multivariate statistical analysis that evaluates 16 measures of personal and public "harm" caused by 20 legal and illegal drugs--from heroin and Ecstasy to alcohol and khat.  Multivariate analysis is related to Monte Carlo simulation, and like that statistical technique incorporates sensitivity analysis.  

 

You probably won't be surprised to learn that the study finds the most "harm" comes from alcohol and the least from mushrooms, or that the authors conclude that "aggressively targeting alcohol harms is a a valid and necessary public health strategy."  

 

On the way to these final judgments, there are some very interesting findings, and if only for the glitzy graphs representing the statistical analyses--and the current push to legalize marijuana in California--I recommend taking a look at Easton's blog.

 

For some amusement and bemusement scroll down to the comments to the blog--e.g., "I especially like the 'pink' bits in the last diagram, 'drug-specific imparement of mental functions', which is presumably what you are actually paying for."  These opinions demonstrate aptly that--statistical analysis, sensitivity regression analysis, Monte Carlo simulation--you can lay the facts on the line but no two people will look at them the same way.

Sooner or later, it had to happen. . . .  Tweets have been linked to stock market behavior.  This was not a case of inside information.  Researchers from Indiana University have demonstrated that public mood, as expressed in millions of Tweets, can predict stock market behavior with fair reliability .   

 

Analyzing a collection of 9.8 million Tweets from 2.7 million users in 2008, the team used a "subjectivity analysis" tool called OpinionFinder and a Profile of Mood States (a psychological measure) to create a time series that tracked daily variation in public mood as exhibited by the language of the Tweets.  It then compared the fluctuations in mood with those of the closing values of the Dow Jones index.

 

To make these comparisons, the team trained a neural network on the data.  Of course, this was not just any neural network.  It was a Self-Organizing Fuzzy Neural Network, one in which organizes its own"neurons" during the training process.  

 

In an article in last week's Oil & Gas Journal, Tom Smith focused on the use of neural networks in oil and gas exploration.  Because of the technology's usefulness in classifying data and identifying patterns, it has become widely used to reduce the risk and time in the siting of oil and gas wells. All well and good, but not good enough apparently to satisfy the growing intensity of exploration.

 

Oil and gas apparently aren't the only things spurting out of the oil fields. These areas are gushing data, so much data that conventional neural networks can't process all of the information.  Author Smith believes that the next step in reducing risk and wasted time in exploration will be the "unsupervised" neural network.  It pushes the Known off the computer screen and replaces it with an Automated Unknown.  

 

While the "supervised" neural network processes classified data, that is, known information, the "unsupervised" neural net can classify unclassified data and then process the patterns that result.  This makes it invaluable for seismic interpretation, that is, for detecting and analyzing subtle geological variations that may be related the potential to extract usable oil or gas.

 

Smith predicts that unsupervised neural networks will be a "disruptive" technology in seismic interpretation.  A disruptive technology is an unexpected innovation that changes the direction of progress in an industry, like digital downloads in the music industry.  If he's right, it just got a whole lot easier to strike oil. 

 

The increasing number of mentions of Monte Carlo simulation in the popular press usually refer to its use in the realm of finance--for such applications as determining value-at-risk, reserve estimation, and credit risk management--because this is where quantitative analysis hits us directly in the pocketbook and where the technique is relatively easy to explain.  But there is a parallel upturn of coverage in the realm of medicine, particularly in pharmaceutical risk management, that is mostly taking place out of the public eye. 

 

This coverage appears in specialized periodicals--such as Genetic Engineering -- their online counterparts, and in the offerings of online aggregators targeting in audiences in medicine, public health, and the pharmaceutical industry.  These articles deal with statistical analyses that are not so easy to explain-- pharmaceutical risk assessment in drug trials, diagnostic probabilities in new treatment regimes, risk analysis of public health hazards--and only a limited number of readers can understand them.

I mention this parallel stream of publishing because of the sheer number of medical, pharmaceutical and biotechnology studies that rely on Monte Carlo simulation. The steady rise in the number of Google alerts I receive is pretty clear evidence that the technique has escaped corporate headquarters and is deeply entrenched in the biosciences, going to work on life-and-death issues. 

The so-called Great Moderation, the economic period that began in the late 1980s and ended with the financial crisis of 2007, was characterized by less volatility and more stability and predictability in the financial markets. I recently saw commentary that identified one of the causes of the economy's' plunge from moderation to recession as "Monte Carlo simulation" and the projections it produced.

 

Over time and a few blog entries it became apparent that what should have come under fire was not Monte Carlo as a statistical technique but the probability distributions on which the financial risk analysis models were based.  During the Great Moderation, apparently, many financial analysts involved in portfolio risk management relied on "normal" probability distributions and continued to rely on them even after tail risk events--events in which risk exceeded 3 standard deviations away from an asset's current price--began to occur much more frequently than "normal" accounted for.  This, of course, meant that actual investment returns were far from those predicted by the models.

 

Commenting on recession-proofing and tail risk, PIMCO's Richard H. Clarida observes that as we begin to leave the recession behind we are entering a period in which the New Normal is the paradigm distribution function for asset management.  The New Normal "is flatter and the tails are fatter."  As he sees it the long reach of risk has just gotten longer.  I recommend his essay, which stresses how important it is to "get the tails right."

Distributions aren't cast in concrete, and many Monte Carlo software packages allow you to fit the distribution to the demands of the questions you're addressing with your risk analysis model.  If you find  yourself in agreement with the New Normal, it shouldn't be that hard to get the tails right.    

 

Big Data, of course, are the billions of bytes of information that come to a business through their use of Internet--clicks, click-throughs, patterns of browser use.  The more successful your e-commerce operation, the heavier the flow of Big Data.  But even the smallest e-enterprise can slurp an astonishing amount of it.  What's big about Big Data are the patterns in which they arrive, and these patterns are where the Experimentation comes in.  They tempt the enterprise to employ new stratagems to improve the effectiveness of the e-operations, and because of the fluidity of the Internet medium, a great deal of tinkering results.  If one ploy doesn't prove beneficial, it's pretty easy to try another one.

 

In effect, Big Data, provide a focus group, or a series of focus groups.  This, it seems to me, is where risk analysis, especially financial risk analysis, has a Big Role to play.  With all the historical data in your lap, any standard operations risk simulation can tell you a lot about the potential results of Ploy A versus the results of Ploy B.  Train a neural network on the data.  Then bring up the Excel, the Monte Carlo software, and let it chomp through the known patterns.  This should yield fair predictions of potential results. The ideal implementation, of course, is one in which the analytics play in real time as Big Data roll in.

While the McKinsey authors don't trouble themselves about specific statistical analysis techniques, they make an important point: the fluidity and speed of both data acquisition and quantitative analysis require a new managerial mindset toward experimentation and decision evaluation.