Previous month:
May 2008
Next month:
July 2008

June 2008

Academia, this People Business

I have been thinking about the parallels (and differences) between industry and academia since I finished reading Execution by Larry Bossidy and Ram Charan. As much as the parents of some students would like to believe they are the customers because they are the ones footing the bills, I tend to view the companies that hire our graduates as the customers - diverse and numerous enough to avoid being "market-setters" - and the students as our products, especially at the undergraduate level. At the graduate level, we produce both students and research papers, and our customers would be the places that hire the former and the journals that publish the latter. In both cases, a strategy of moving up-market requires incentives so that potential customers give our product a try (get at least one graduate in the door); even if we already have excellent products, we might want to further increase quality by buying better raw materials (offering more merit-based scholarships to incoming freshmen).

As an example, Goldman Sachs only started recruiting systematically on Lehigh's campus after hiring a couple of students in our flagship honors program in Integrated Business and Engineering. Academia does have one advantage over industry, in the sense that its products have a brain, so universities do not have the same downside risk - just imagine, "Dear companies, Lecture 7 of IE 333 is all wrong! We're recalling all graduates from the Classes of 1986 to 2003 who've taken the class with Professor Absentminded for brain adjustment! Load them onto the Bieber bus and ship them back to Bethlehem! We'll provide you with a replacement at no cost!" As a matter of fact, a defective product is much harder to quantify in academia - what is the university's fault and what is the student's? A student struggling in one career path might become a star in another.

I was impressed by Bossidy's and Charan's emphasis on people processes - a mix of brutal honesty and coaching. People are even more central to universities' mission, and in that regard, I find it striking that professors receive so little formal training when it comes to teaching and advising. Lehigh was unique, among the places I interviewed with, in requiring a teaching seminar in addition to a research one - the vast majority of universities organize a research seminar only, and rely on teaching evaluations of courses the student was a TA (Teaching Assistant) for to judge her teaching abilities, although the TA is the focus of only one or two questions out of the fifteen or twenty on the form. It is a widely accepted fact in most research universities that, to get tenure, you need to be an excellent research but only an average teacher; the reverse is not true: an excellent teaching record won't help you if your research portfolio is only average. In other words, when it comes to teaching, the incentive is only to clear a threshold. Lehigh has shown more enthusiasm for teaching than most places, in part (I believe) because of its history as a teaching institution - the focus on research dates from the mid-1990s only. The Faculty Development program helps professors improve their teaching; it organizes the new professors' orientation every year and I still have the book on good college teaching we were given in 2004 (and yes, I read it).

But, since a key objective of research universities is to advance the state of knowledge through research, it makes sense that their core people process would focus on the advising of graduate students. New doctorate-holders could not be less prepared for that job. (I would like to say that France, which requires would-be professors to pass a second examination before they can supervise theses, shows more interest in teaching its faculty how to advise students; unfortunately, that examination - to the best of my knowledge - is a research examination.) You learn on the fly, observing which students are motivated by your advising style and which ones are not, discovering said advising style along the way. For instance, I have a carrot-driven rather than stick-driven advising style: research is fun, let's write a paper. I hate having to confront students when they don't perform. According to Bossidy and Charan, this is an extremely common problem among supervisors - who doesn't prefer to be nice? This leads many people to postpone confronting non-performing workers in the hope the situation will sort itself out without requiring them to intervene.

Which brings me to my last point. Cheating students and non-performing workers alike often gleefully think that they've tricked the system and fooled their bosses. It turns out they haven't - their bosses know the truth. In many cases, what delays the students' or workers' demise is that the boss is reluctant to deal with the problem, not that the person is a particularly good liar. This might not cheer much the employee's co-workers, who have to pick up the slack and correct the person's mistakes. But people inclined to do the bare minimum might feel a bit more motivated if they realize the boss is seeing through their game. Unfortunately, it is doubtful those people are the ones reading Bossidy's and Charan's book.


Execution and Operations Research

I've just finished reading Execution, by Larry Bossidy and Ram Charan. It has been a good read, especially since many comments the authors make (about execution not receiving the attention it deserves) can be applied to operations research too - after all, OR is fundamentally about the best way to execute a strategy, no matter how many academics would like to think of themselves as high-level strategists. Here are a couple of relevant quotes:

  • "People think of execution as the tactical side of business. That's the first big mistake. [...] No worthwhile strategy can be planned without taking into account the organization's ability to execute it." (p.21, hardcover edition)
  • "It's a pleasant way to view leadership: you stand on the mountaintop, thinking strategically and attempting to inspire your people with visions, while managers do the grunt work. [...] Who wants to tell people at a cocktail party, 'My goal is to be a manager', in an era when the term has become almost pejorative?" (p.24)
  • "The real problem is that execution just doesn't sound very sexy." (p.31)

(As an aside, the book even ranks #1, as of this writing, in the Amazon.com Bestsellers list for books in operations research. Thrilled that the concept of operations research is reaching a mass audience, I could not help but check the list. Sadly, quite a few books listed don't have much to do with OR, and many others are textbooks rather than books intended for the general public. I suspect the label is slapped over any operations-related book that is hard to categorize. "I don't really know what this book on operations is about; that's what research must be like. Hey, we do have a label for stuff in operations and research!" Wikipedia explains that OR "uses methods like mathematical modeling, statistics, and algorithms to arrive at optimal or near optimal solutions to complex problems. [... OR] helps management achieve its goals using the scientific process.")

Most relevant to operations researchers is the last chapter of the book - "The Operations Process: Making the Link with Strategy and People." Bossidy and Charan explain that "many companies translate their strategic plans into operations [... by spelling out] the results you're supposed to achieve, such as revenues, cash flow, and earnings, and the resources you're allotted to achieve them, but the process doesn't deal with how - or even whether - you can get the results." OR would play a critical role in addressing that issue, if only more managers gave it a try.

Some executives sneer at the idea of using optimization models in their decision-making process - how will they find the precise value of this cost parameter? What if that assumption doesn't hold? But the truth is, any business is run using numbers: the performance targets decided by top management. Wall Street analysts produce their own numbers too when they study companies. All of these are driven by specific assumptions, which might or might not prove well-founded. 

Nobody is saying Wall Street analysts should find themselves another job because picking the right assumption is too difficult, so why aren't more businesspeople relying on OR models to determine the best way to reach their targets? "The financial targets are often no more than the increases from the previous year's results that top management thinks security analysts expect. [...] No one necessarily knows how and why those numbers are reached, but they nonetheless become marching orders for the coming fiscal year." (p.229) Linking strategy and operations using OR would ground the numbers in a more realistic framework and provide a plan to reach them.

Bossidy and Charan insist on having several contingency plans. OR even has a name for this: adaptable optimization, where the decision-maker decides on an operating plan early on in the process but also devises a set of prespecified policies he can implement once he knows more about the market conditions. This is due to my former doctoral adviser at MIT, Dimitris Bertsimas, and his then-student, Constantine Caramanis, now at UT Austin. Nobody ever said OR should consist in one model, and produce one answer. Instead, you want OR models to give you one good plan that can be quickly modified into another - also good - plan if conditions change, rather than one great plan whose performance decreases drastically if one tiny assumption proves to be unfounded. The issue isn't with OR per se, but the way it has been implemented so far. Who wouldn't be suspicious of a single computer model that pretends to know what's best for you when you aren't even sure of the conditions you'll face in the next quarter?

Maybe another reason why OR hasn't achieved quite the recognition it deserves in practice is that the models are often limited to one aspect of the decision-making process, such as inventory control, while top managers use a different set of performance metrics. In other words: who cares that inventory costs are minimized? what impact does OR have on what investors and analysts care about? If OR solves models based on the preoccupations of low-level management, it will remain a discipline for low-level managers. Integrating operations with finance might become a necessary step to increase OR's relevance and visibility.


First Universal Product Code...

...was scanned on a June 26. I wrote a post about it on the Engineering Pathway website. Check it out! While you're browsing the "Today in History" posts, make sure you read Alice Agogino's post about Marie Curie defending her thesis and receiving the Nobel five months later (yes, I really mean months) and her account about the first shopping cart. I particularly enjoyed the part where the inventor hired people to push his new shopping carts and demonstrate how wonderful they were because customers were adamant they preferred the old baskets. While more recent carts designed by IDEO (pictured in the post) have not been widely implemented, I have seen similar - although less futuristic-looking - ones at my local supermarket, with one top and one bottom tray. I don't particularly like them. They have less capacity than the regular carts and it is easy to put in single items but much harder to fit the paper or plastic bags you get at the cashier's; since the trays are rather shallow, you cannot just pile bags on top of each other in the trays because they will fall out. Anyway, it is good to see innovative concepts, no matter how flawed, find their way to the market - once in a while, customers will adopt something that suits their needs, and the inventor might not even need to hire actors to convince them.


Someone Dislikes Math

I came across an article entitled "Charter school - how not to unwind after work" in The Economist dated June 7th. The subtitle ("how not to...") puzzled me, since the article is devoted to the CFA (Chartered Financial Analyst) exams and the growing number of students taking them. There really didn't seem to be any reason to pass judgment on people studying for an exam in order to improve their career opportunities. Then I reached the last paragraph, which describes the high dropout rates ("only a fifth of the candidates who start complete all three stages; a quarter do not even turn up at the exam"), and the last sentence made the subtitle clear: "Mind you, faced with questions such as 'What are the desirable statistical properties of an estimator?', they can hardly be blamed for that." You can sense the author's absolute distaste for anything math-related right there - he or she pities the poor souls forced to stare at math in their free time. Any introductory statistics course will cover estimators; we're not talking about graduate-level statistics here. I would find the reporter's reaction funny if that didn't reinforce the general public's conception that it is okay not to want anything to do with math - that it is okay not to want anything to let other people deal with math for you and hope they understand the topic more than you do, it is so okay you can even brag about it in a leading international publication, although the Economist's writers make sure they know all kinds of fancy economics and intricate rules of state elections and deadlines for cease-fires in all troubled countries you can think of.

So here are the two desirable properties of statistical estimators.

1. You want your estimator to actually estimate the quantity you're trying to figure out. That makes sense, doesn't it? This is called having an unbiased estimator. A textbook example of biased estimator is to think about the employee at the deli who puts the meat he's just cut for you on the scale, but the scale is not tuned properly and shows there's 0.1 pound of food on the scale when it's in fact empty. All the measurements are going to be off by 0.1 pound. If you use the scale reading to estimate the true weight of the meat, you're using a biased estimator. When you can make observations without equipment, an unbiased estimator of the expected value of a random variable (like the expected number of broken eggs in any given plastic box) is the average of all the observations you have collected (the average number of broken eggs in the specific 43 plastic boxes you have analyzed when you were at the supermarket). 

2. If you use your estimator several times, you don't want the numbers you get to vary widely from each other. That's because you don't want to make a lot of trials before getting a good idea of what the unknown parameter is. So you say you want your estimator to have low (ideally, minimal) mean-squared error, which is equal to the variance of the observations when the estimator is unbiased.

The single best-known biased estimator (read: the one that causes most mistakes) is found in variance estimation problems. When you consider the variance of a random variable X (in math: E[(X-E(X))^2] ), you need to compute expected value twice, and one expected value is inside the other. Because of this last point, the natural strategy of replacing expected values by sample averages throughout creates a bias in the resulting estimator. Specifically, it underestimates the variance. (You can read the formulas in the "sample variance" section of this Wikipedia article. After trying to put them in plain English for this blog, I decided that simply was not doable. I wouldn't want anyone to end up more traumatized by math because of this post.) That also means the resulting confidence intervals are too narrow. If you have a lot of observations, then the difference between "right" and "wrong" variance is small, but it exists nonetheless - and underestimating risk is not a good idea these days.

Luckily, most calculators and spreadsheet software compute sample variance the right way. Unluckily, they also offer an option to compute it the wrong way, just in case you enjoy showing wrong results to your boss. In Excel, VAR(...) computes the unbiased estimator of the data between parenthesis, while VARP(...) computes the biased one. Why anyone would ever want to use that one eludes me. But since no one in their right mind would type VARP rather than VAR when they calculate the variance of their data, I guess VARP will remain one of these Excel functions no one cares about.


MIT Student Evaluations

I recently came across MIT student evaluations dating back from 1991. Several versions exist on the web (for instance here and here). I just thought I'd give my Top Ten for today's post.
10. "The class is worthwhile because I need it for the degree."
9. "He teaches like Speedy Gonzales on a caffeine high."
8. "Text is useless. I use it to kill roaches in my room."
7. "Text makes a satisfying 'thud' when dropped on the floor."
6. "His blackboard technique puts Rembrandt to shame."
5. "The recitation instructor would make a good parking lot attendant. Tries to tell you where to go, but you can never understand him."
4. "The absolute value of the TA [Teaching Assistant] was less than epsilon." [For those of you non-mathy type, epsilon is a very small positive number. Like 0.00000000001. Absolute value is always positive. Draw your own conclusions.]
3. 'What's the quality of the text?' "Text is printed on high-quality paper."
2. "Problem sets are a decoy to lure you away from potential exam material."
1. "The course was very thorough. What wasn't covered in class was covered in the final exam."


Doing Good, or Making Things Worse

We've all gotten used to hotels' little cards in rooms explaining they will change the bed sheets every three days to limit the use of detergent in their laundry. I recently stayed at a hotel that proudly advertised the fact that patrons would not receive a paper invoice; instead, the invoice would be emailed to them to save paper. Isn't the sight of big hotel chains fighting against pollution and deforestation just heart-warming?... In its May 22nd issue, The Economist published an article about the ugly truth behind the rise of Internet computing: companies like Yahoo!, Google and Microsoft need more and more servers to handle the increase in web traffic and usage - all the gigabytes of free storage space on Google Mail, for instance, need to be stored somewhere. The article gives a few impressive statistics: "America alone has more than 7,000 data centres," "In America the number of servers is expected to grow to 15.8m by 2010—three times as many as a decade earlier," and "Google is said to operate a global network of about three dozen data centres with, according to some estimates, more than 1m servers." The issue is that the servers need to be kept at room temperature (without cooling, they tend to overheat), and thus use massive amounts of energy. This has become such a problem that "Microsoft is looking for a site in Siberia"; Iceland has also become a popular destination.

I do very much enjoy scanning papers and keeping them on my Google Mail account where storage is practically unlimited; I find it much more convenient that keeping a loose piece of paper. It also seems more environmental-friendly because it limits the use of paper, and makes the end users (me or the hotel patron) feel good about themselves by doing their part to fight deforestation. But how much good does "paperlessness" bring to the environment, if it just increases the energy required to cool down the servers? If I write something on paper, I just use one sheet once. The server needs to be cooled down day after day. Of course, in practice, I might well print it, so you need to take into account the cost of powering my computer and my printer, but I doubt a desktop uses an amount of energy of the same order as a server. All this to say - what exactly are the cost benefits of going paperless? I know having lots of trees reduces global warming, but making Iceland and Siberia melt instead to reduce paper use isn't the right way to tackle the problem. It turns out that AMD recently sponsored a study on that exact question, i.e., what is an estimate of total power consumption by servers in the US? The numbers are staggering: AMD puts the price tag at $2.7 billion (in utility bills for the companies owning the server farms) in 2005. The study's author also states that, under reasonable assumptions, "total electricity used by servers by 2010 [only 18 months away!] would be 76% higher than it was in 2005." Thankfully, computer companies have recognized this challenge and quietly begun to address it. One approach they are considering is called virtualization, to make more efficient use of idle servers. Nanotechnology also appears to hold much promise (see "Low-power electronics with nanoscale devices" in this page maintained by HP). While hotels and the like look forward to decreasing their costs by going paperless, Internet companies are scrambling for a way to prevent their own costs from ballooning. Whether they will find one before their utility bills become unmanageable remains to be seen.


Internet Revenue Management

Internet providers are implementing new schemes targeting their most active users - those who share a lot of video and music files. The Washington Post, in an article dated June 4, describes Comcast's and Time Warner Cable's practices; Comcast will delay traffic of the heaviest users, while TimeWarner plans to adopt an approach similar to cell phone billing, where customers are charged more "for larger volumes of data and faster Internet access".

A New York Times article on the same theme, though, points out that very few users would actually reach their maximum amount allotted. For instance,  "[in a trial run by Time Warner], new customers can buy plans with a 5-gigabyte cap, a 20-gigabyte cap or a 40-gigabyte cap. Prices for those plans range from $30 to $50. Above the cap, customers pay $1 a gigabyte. Plans with higher caps come with faster service." The article goes on explaining, with welcome specificity (lacking in the Post's article) that "Streaming an hour of video on Hulu.com, which shows programs like "Saturday Night Live", “Family Guy” and “The Daily Show With Jon Stewart,” consumes about 200 megabytes, or one-fifth of a gigabyte. A higher-quality hour of the same content bought through Apple's iTunes store can use about 500 megabytes, or half a gigabyte."

Predictably, users' groups and representatives of Internet companies are reacting as if making heavy users pay for clogging the bandwidth would endanger the survival of the Internet. From a revenue management perspective, I'd be curious to see how users react to those prices. An issue with the flat-fee structure is that people have not been educated on how many mega- or gigabytes they use each month and cannot judge for themselves whether they will be affected by the new pricing scheme. Maybe, before scaring average customers into believing they'll have to pay more to access the Internet, or before pretending that average customers have nothing to worry about, Internet providers should try harder to let them know how much they are currently using.


Innovation in America

The Economist's online edition of June 3rd has a column on whether America can keep its innovative edge. The column's author opposes the creation of a National Innovation Foundation (first suggested by the Brookings Foundation), which would have a budget of $1 billion to $2 billion to "promote partnerships between universities and industry, and support regional industry clusters with federal grants". I agree; taxpayers' money would be better used strengthening cash-starved institutions such as the National Science Foundation, which already runs a Small Business Innovation Research (SBIR) program. We do not need more federal agencies; we need better funding for the ones we already have. There is also ample evidence that governmental interference to create innovation clusters doesn't work, and a system that produced Dell and HP cannot be a bad system.

I was disappointed that the column, although insisting that "America has long been a global powerhouse of innovation, breeding thousands of firms [...] from ideas born in garages", does not give any thought to what will make tomorrow's innovation, and whether tomorrow's type of innovation will be any different from yesterday's. In other words, has all the low-hanging fruit been plucked? Companies that revolutionize their industry rely on disruptive ideas - for instance, Michael Dell realized that he could sell computers without using middlemen. But as Americans' environment becomes more technology-intensive, it seems reasonable to assume any breakthrough made in the States will rely more and more on advanced technological concepts. That doesn't mean college students will stop creating companies in their dorm room, simply that these companies are less likely to have a major impact on their industry. Teams participating in the MIT's "100k competition", for instance, always include a researcher or PhD student. Even Mena Trott, the founder of the blogging behemoth Six Apart - which commercializes Typepad - and former English major, would not have been able to start her business without the help of her husband, who wrote the computer code for her (see Face Value in The Economist, November 23, 2006). While everyone likes a good "self-made man/woman" story, the barrier to entry in the field of significant innovation (worth making a company for) might well have been raised out of the reach of the average college kid.

This should be good news for graduate students in engineering, who have the technical knowledge their younger counterparts lack. Many graduate students in engineering, however, come from countries that do not particularly value taking risks or encountering failure; they are also reluctant to spend time on side projects instead of trying to graduate (in contrast with undergraduates, who are predominantly American, they also have almost no support system in the States). The increased reliance on doctoral students in engineering comes at a time of heightened concern about studying in the States by those foreign students - I won't discuss this heavily-publicized point here, except to mention that patent applications often include graduate students, and students of lesser quality produce patentable inventions of lesser quality, even if the patent is ultimately granted. (Don't count on the employees of the Patent Office to recognize that the same idea could have been implemented in a much more cost-effective way with this or that radical new chip design, for instance. Who can tell of what could have been?)

Another issue is that, in the same way that the days of "ideas born in college dorm rooms" might be behind us, tomorrow's innovation might be a lot more team-driven due to the sheer complexity of the problems at hand. While America's culture emphasizes individualism, it is hard to think of another country currently benefiting from a similar quality of graduates and superior sense of teamwork. As a historical example, the Japan of the late 1980s comes to mind; the last paragraph of page one in this Time article ("What killed the boom") is eerily reminiscent of the present situation in the U.S., but competitors to the States remain a bit far behind to "close the gap" that easily, at least in the immediate future. That much is clear: the era of the 20-year-old innovator is over.


The Real Value-at-Risk

The Economist had a special report on international banking a few weeks back (May 17th, 2008). I particularly enjoyed the article on risk management ("Professionally gloomy"), despite the fact that the author has the wrong date for the concept of Value-at-Risk - it was developed by JPMorgan in the 1990s, not the 1980s, as the magazine itself noted correctly in a January 22nd, 2004 article, "Too Clever by Half". The definition of "the maximum amount of money a bank can expect to lose" isn't the clearest available out there, but it's better than nothing, and the rest of the article goes to some length to clarify the concept. VaR receives quite a bit of attention in the mainstream media because it has now become "a staple of the risk-management toolkit and is embedded in the new Basel 2 regime on capital adequacy."

The article points out that the VaR calculations, which use historical data from the previous three to four years, let banks take more aggressive positions "the longer things go smoothly", for the same total amount of money set aside. Then, when there is a crash, prices that showed little volatility and correlation become extremely volatile and correlated: suddenly, the bank loses a lot more money than the VaR it had calculated. (See also Buttonwood's column, February 17, 2004, and the article titled "The Coming Storm", February 19, 2004 and "Uphill work", September 6, 2007) Common sense, however, suggests that "the risk of a blow-up will increase, not diminish, the farther away one gets from the last one." VaR also "acts as an amplifier" by triggering sell orders in times of crisis, which depress prices further and trigger additional sell orders from other companies. The last issue mentioned in the article is that "VaR captures how bad things can get 99% of the time, but the real trouble is caused by the outlying 1%."

It is interesting to note that "VaR breeds complacency" is not a new observation - the same theme can be found in the January 2004 article mentioned above, for instance, and yet banks don't seem to have taken advantage of the previous four and a half years to address this flaw. A difficulty is to determine how many years of data banks should use - how much of what was true five or eight years ago remains true today? This is for instance the case for computer-based trading: if you used different computer systems back then, or relied more on traders' judgment, the sell-offs might exhibit different characteristics that limit their present usefulness in predicting asset price behavior. Using historical data to compute VaR, quite simply, might not be the right way to approach the problem; someone at PriceWaterHouseCoopers suggested to start with given amounts of money (say, $1 billion) and "work backwards to think about what events might lead to that kind of hit."

To compensate for the limitations of historical data, banks run stress tests, which are discussed in "Eggheads and long tails," May 17th, 2007. (The most hilarious sentences of the day: "[UBS's approach to risk is described as:] zero tolerance for fiefs; beware of tail risks, risk concentrations, illiquid risks and legacies; avoid risks that cannot be properly assessed or limited; and never be hostage to a single transaction or client. [...] Indeed, it is sometimes criticised for not taking large enough punts. Earlier this year [a prominent investment banker] left the firm, reportedly because he felt it was too conservative in using its own capital in private-equity deals." To understand why it's hilarious, recall the recent UBS meltdown, which I wrote about here.) The "eggheads" article also mentions some interesting assumptions behind Goldman Sachs's scenario planning, such as the need to protect against liquidity risk and the hypothesis for stress-testing purposes that "the firm is unable to hedge or sell positions and must hold them from peak to trough. One of its fixed-income tests is a replay of the 1998 LTCM and bond-market meltdown. In its equity division it uses a “supercrash test” that assumes an instantaneous fall in the price of equities of 50%." Details of the simulations, however, are scarce. A footnote in the "Trading places" graph in the "Coming Storm" article, showing how VaR has increased in every bank from 2002 to 2003, puts the discussion in perspective: "Figures are calculated differently by different banks." That doesn't exactly inspire confidence in Wall Street's risk management abilities.