When I was on the faculty at Yale I knew people in admissions and it's not clear to me that they were the best able to spot potential in 18 year olds. In studies of expert performance admissions people are less good at predicting UG[undergrad] GPA than a simple algorithm. (The "algorithm" is simply a weighted sum of SAT and HS GPA!)
When I worked at a bank, I remember we had a bunch of 'underwriters', and their job was to evaluate the creditworthiness of loan applicants. Coming out of grad school, I had no experience with this, and they had this very complex, holistic methodology. Later, when I became head of capital allocations and we started quantifying their portfolios, I invariably found they could be replicated via some pretty simple rules. This got me excited about going to Moody's to help develop their new RiskCalc Private firm model, which today dominates private firm underwriting worldwide. Anyway, the simple trick was always the same:
1) identify the key indicators
2) transform them appropriately (eg, turn 'profits' into percentile for profit/assets, or some sigmoidal transformation like exp(x)/(1+exp(x))
3) weight or crosstab the indicators
4) add 'em up, and transform output into meaningful buckets used for pricing (200 bps) or risk classification (eg BBB)
Risk is invariably on a log scale, so usually the ordinal rankings correspond to exponential increases in expected default rates.
Anyway, with hindsight, I found most of the facade put forth by various departments (eg, auto lending, health care lending), was very misleading. Everyone made the simple complicated. I think deep down, no one likes to think a computer can do their job, and there are many instances where exceptions matter, so a great deal is made out of these special cases. Yet the false positives made them great anecdotes, but horrible for generalizations. Thus, simple rules dominate their much more costly, confusing, and non-quantitative product created by teams of analysts.
These jobs seem rather common. For example, In the nineties many 'traders' I knew would simply buy from a customer at the bid, and then sell to another customer at the ask, and go home 'making' $10k a day. They would then get a $500k annual bonus. A college admissions director probably interviews people all day, and writes pages of material supplementing their decisions, but all to trail a simple linear rule. I also worked in an econ department, where we created tons of forecasts with lots of commentary, all dominated by vector autoregressions. Our asset/liability committees would often have a lot of commentary about where various interest rates are going, as if these forecasts were any better than what could be lifted from forward prices.
People in these situations are rather quite pathetic. Many are really good people, and would take a big pay cut if they started over, so its not as simple as just choosing a new career.
27 comments:
The quote from Steven Hsu is assuming, but I think that it slightly misses the point of college admissions, at least at many top tier universities. It assumes that the goal is to select the incoming class in which each student is individually as smart as possible, with the assumption that they will go to class and cogitate and get good grades. However, college is a very social experience and that applies to learning as well. The people who contribute the most to their classmates and who add the most value to the school may or may not be the ones with the highest grades.
Many college admissions officers are quite explicit that their goal is not to select the class of individually smartest people. MIT, for instance, once claimed that they evaluate applicants on two dimensions, which we might oversimplify as "smartness" and "leadership/interestingness", and they pick the best students on Dimension 2 who pass some threshold on Dimension 1.
Of course, one could debate the value of such criteria, and wonder whether admissions officers choose them specifically so that their job (as they define it) couldn't be done by a computer.
Separately, I suspect that doctors are another example of people whose jobs could, in many cases, be done by computers. A doctor is basically a human expert system who has been taught various things about various diseases and applies their formula to every patient whom they treat. Granted, different patients respond different ways to treatment, but the initial treatment and fallback treatments that a doctor will prescribe for a given condition are almost always the same. It would be possible, given sufficient time and medical information, to program a computer with a complicated decision tree to treat patients as well as or better than a human doctor. (Part of the difficulty in doing this, of course, is that the medical community jealously guards their knowledge, requiring you to go to medical school to obtain it.) In fact, such a system, MYCIN was built in the 1970's for diagnosing infectious blood diseases and recommending antibiotic treatment. The system did better than most physicians but was never used. The barriers to such a medical A.I. system ever being extensively used, of course, include the desire for the "human touch", the popular opinion that doctors should be perfect (even if they aren't), and the enormous legal liability involved (not to mention the AMA being an anti-competitive trade guild trying to reduce the supply of medical practitioner so as to drive up the incomes of doctors).
The only thing a loan underwriter really needs to know is loan to value and foreclosure costs & timeframe. Keep it low enough and creditworthiness is basically irrelevant, you will get paid from the collateral no matter what the borrower does.
"I invariably found they could be replicated via some pretty simple rules."
Can't simple rules be gamed though -- and didn't that happen with FICO scores during the mortgage bubble? I recall reading that Moody's placed a lot of weight on FICO scores but didn't take into consideration that some of the high-scoring borrowers were indigents with virtually no credit history who had been coached into building a short credit history that would give them good FICO scores.
"Separately, I suspect that doctors are another example of people whose jobs could, in many cases, be done by computers."
Years ago, the WSJ's Holman Jenkins didn't quite go that far, but he argued for much greater use of computers in diagnosis. His point was that physicians should be more like airline pilots keeping an eye on the autopilot and stepping in for the complex stuff rather than micromanaging every decision, and that this would result in fewer medical errors.
Yeah - what Dave said - the over-complex heuristics are a kind of security through obscurity measure against gaming. It's not perfect - but it is slowing down the Batesian Mimicry mechanism you write about.
I ran a chain of retail stores that sold expensive merchandise. I told our managers 98% of the time, the system can handle the business... I need you to handle the 2%. It works like this. Out of 100 people that come in the store, 2 will come in barefooted and disheveled. One will be a homeless person and the other will be a rich famous author that lives in your town and is trying to get some clothes for a tv interview. Your staff, the security system and the company policies won't be able to tell the difference between these two people because their behavior is identical. We've already been replaced by computers for the norm, it's the exception that we need humans for. That one barefooted millionaire with oatmeal in his hair would spend $10,000 without even thinking about it.
If you think algorithms can defeat individual creativity, then how do you explain Demand Media's successful IPO last week and how they're using people to game Google's algorithm down on the content farm.
Dave Harrison
tradewithdave.com
I would mirror your earlier commenters in that the human process is important in identifying situations that game the simple system.
I think you are correct that simple algorithms can account for most of the information regarding many decisions, but I think relying on them in an automated fashion will get you killed particularly in a trading environment.
If you just blindly follow these algorithms you will select risk that the market knows something about that your algorithm does not.
The best pairing to me is an algorithm in the hands of someone who both understands how it works (including its blind spots) and understands the fundamentals of the real world. The algorithm then does the first pass of identifying potential winners/losers and then a person then checks the model's blind spots.
While I won't argue against the point that analytical methods appear to bloat, or at least attempt to be more comprehensive over time, I must ask: why does this occur?
Is it not the natural tendency to try and glean more insight and information? After all "...there are many instances where exceptions matter..." and the unexpected tends to be precisely where one is not looking.
It just seems that a lot of tail risk is blithely ignored when sticking to a set of rules, as opposed to adapting your methods with whatever it is you're looking at.
Accounting is all about applying very simple algorithms. They are hundreds of excellent computer programs. Yet we pay a lot for a good human accountant. Are they redundant?
I think sometimes the "clustering illusion" (http://en.wikipedia.org/wiki/Clustering_illusion) comes into play - humans can tend to over think things. Also, humans easily confuse correlation with causation, leading to incorrect assumptions and ultimately poor choices.
Don't forget the effects of performativity. In many instances (particularly related to finance), following a particular rule/algorithm will improve its predictive power.
Donald MacKenzie basically wrote 'An Engine, Not a Camera' to explain how this happened to the Black-Scholes-Merton pricing model. Highly recommended reading.
I now believe that even at first seemingly "poor" automation pays off very well because of the performativity effect - people will change other assumptions and processes to comply with the way the automation works, basically "smoothing" the variations in a whole chain of processes, improving the results of the initial automation, and paving the way to further automation at larger scales of the system.
Pharmacists - besides the occassional need for a compounding pharmicist why couldn't a computer do as good a job dispensing medication?
If what you say is correct, some competitor should be able to start a {brokerage, insurer, etc} using simple linear regressions instead of employees. They would have much lower costs and a product of equal quality. They would drive the firms who hire these people out of business.
How do you explain the fact that this has not happened?
This is why we'll never have anything like a flat tax or greatly simplified tax code. Because...a computer could do your taxes quickly, accurately and fairly.
Just think how many accountants and lawyers that would put out of work overnight.
That is what got me so excited when I first started programming, I though "If I am ever in a shit, repetitious job again, I will just automate myself".
No because I'm a programmer. I tell the computer what to do.
Programming I think will still be outside the frame of what a computer can do by itself. If a computer can put a cap on a bottle it did so because it was programmed to do so...... so I still have my job.
Automating jobs away is easy in the cases where understanding was already replaced by statistics. Why do people do that? Because they have to, their performance is measured statistically. Even in the cases where performance can’t be correlated with monetary measurements.
Here is another job to automate: the chief of police. Her job is to keep the crime down”, which in turn is measured with crime trends in percentages. A simple learning system cold be devised that based on various statistics of officers performance can devise hiring and promoting policies that would improve measured statistics. Will you fill safer living in that neighborhood? You bet, at least statistically speaking. Another thing you can count on is that the amount of reported crime will decrease too. Only losers would be manufacturers of the garden fertilizers, the demand for their product would suffer a few percentages.
We can automate now using simple rules. For example Insurance companies have complex underwriting system but the most heavily weighted factor is the credit score. If you are slow to pay your bills, you are more likely not going to be proactive to fix a link in the roof or fix your car which will lead to a greater expense paid by the insurance company.
But these types of simple rules are good ways to evaluate 80% of high volume transactions. It is the remaining 20% of high volume transactions that need more complex rules or human intervention. (Store manager handling homeless person and eccentric millionaire is a great example.)
This 20% evaluation is where 80% of the cost and effort occurs. I have seen examples where you can't program but you need a domain expert to evaluate it.
The market place knows this. A previous comment gave an example of a Doctor being replaced by a system like this. For the simple 80% high volume transactions, we already have medRx.com sites, physician assistants, RNs, emergency room staff doing these types of transactions. It is the remaining 20% high volume transactions that bring in 80% of the revenue that is handled by specialists such as cancer treatment, surgery, and costly exams like MRI, CAT scans.
We can only hope to automate the first 80%, you are welcome to diagnose cancer, I just will not trust my life to it.
"you will get paid from the collateral no matter what the borrower does"
Yes, but there's a lot of friction involved in extracting the value of the collateral.
Yes, but there's a lot of friction involved in extracting the value of the collateral.
That's why you adjust your required loan-to-value (LTV) according to the cost of such friction. You would require lower LTV in New Jersey than you would in Texas.
The exceptions are important so the question is can any formula detect them? When I worked in University admissions we normally just added grade points to get a student score. This was fine until I noticed the Malayan student who had an A grade in Malayan.
The exceptions are important so the question is can any formula detect them? When I worked in University admissions we normally just added grade points to get a student score. This was fine until I noticed the Malayan student who had an A grade in Malayan.
If all the rules became written and simple enough to be programmed, it might be much easier to exploit the system. Like that person who sued McDonalds for the coffee being too hot, however stupid that sounds.
There are some jobs that can be automated, but, on the other hand, automation gives potential competitive advantage to those who approach the situation creatively.
Since for a lot of people their job is just a job, and not the way to express themselves, creativity is not always there. Well, that's probably the way out of this, at least until computers become as intelligent as humans.. Routine things will be left to computers, and people will have to start working creatively. Also, automation itself is a creative activity - isn't that why there are so many consultants in the IT industry?
Computers may make our lives easier and can do many "pointless" and mundane jobs. How many of us are working less hours as a result? The ones who are have been made redundant. If you keep taking people out of the workforce and replacing them with machines and cheap labour, society suffers as the economy eventually shrinks due to peoples decreasing earnings. What do people who cant find a job because machines of cheap labour can do a cheaper job? Eventually the wealth will be in the hands of a small minority while the rest of us live in extreme poverty. In short, if everyone is a wee bit inefficient and overpaid, it is better for everyone because jobs create wealth and wide spread wealth creates more wide spread wealth.
Anonymous, if it were true that automation shrinks the economy, then the people who smashed machinery during the nineteenth-century industrial revolution were right, and we should all return to the textile factories.
On the contrary, automation lets people produce more per hour, which leads to increased wealth. (There are only two ways to increase production - work more efficiently or work more hours.) Compared to the start of the industrial revolution, we are extremely efficient now, having automated thousands of jobs. As a result, we as a society create more wealth and have a much higher standard of living.
Bendodge, anonymous has a point. The ultimate idea of automation is to replace inefficient people with more efficient machinery/computers. Whether people are left with some jobs or not is only a question of how advanced the automation is. From this standpoint, there is a huge difference between 19th century and our days. Read that post about the system being able to handle 98% of the retail store business for example.
Anonymous said, "In short, if everyone is a wee bit inefficient and overpaid, it is better for everyone because jobs create wealth and wide spread wealth creates more wide spread wealth." That is not correct. Inefficiency sucks wealth out of society. When everything costs more, everyone has less. When the cost of living is less, people have more resources to improve themselves.
Anonymous seems to think that people cannot be trained for productive work if they are replaced by automation, so once out of work, permanently out of work. Our history proves that incorrect. In my own case, I worked at a picture framing gallery that went out of business, so I got a job as an optician. When I saved enough money, I bought optical books and became a certified optician. The higher pay allowed me to buy a home computer, and after I studied programming, became a software engineer. Automation made the optical books cheaper, automation made the home computer cheaper, and that gave me oportunities that I would not have had otherwise!
Beautiful ideation.
But let me chip in my 2 cents too.
Rules are not always boolean in nature; it does not always fit into a sort of an if-else structure. How would a fully automated autopilot (without human pilots) handle a situation of hijack? We could perhaps program it for such a situation too, but the question is, that how many different situations could you program it for? How would you think of all such situations in the first place?
Your article reminded me of the Swarm intelligence theory, which states something like that small and unintelligent units can produce an overall intelligence behavior based on simple rules. Perhaps we could program computers using such rules which would handle much more complex situations than their own complexity: like sum of the parts is greater than all the parts would normally make combined into one.
Post a Comment