What medicine can learn from Wall Street – Parts I, II, and III
April 3, 2014
We, in healthcare, lag in computing technology and sophistication vs. other fields. The standard excuses given are: healthcare is just too complicated, doctors and staff won’t accept new ways of doing things, everything is fine as it is, etc… But we are shifting to a new high-tech paradigm in healthcare, with ubiquitous computing supplanting or replacing traditional care delivery models. Medicine has a ‘deep moat’ – both regulatory and through educational barriers to entry. However, the same was said of the specialized skill sets of the financial industry. Wall St. has pared its staffing down and has automated many jobs & continues to do so. More product (money) is being handled by fewer people than before, an increase in real productivity.
Computing power in the 1960’s-1970’s on Wall street was large mainframe & mini-frame systems which were used for back-office operations. Most traders operated by ‘seat of your pants’ hunches and guesses, longer term macro-economic plays, or using their privileged position as market-makers to make frequent small profits. One of the first traders to use computing was Ed Seykota, who applied Richard Donchian’s trend following techniques to the commodity markets. Ed would run computer programs on an IBM 360 on weekends, and over six months tested four systems with variations (100 combinations), ultimately developing an exponential moving average trading system that would turn a $5000 account into $15,000,000.(1) Ed would run his program and wait for the output. He would then manually select the best system for his needs (usually most profitable). He had access to delayed, descriptive data which required his analysis for a decision.
In the 1980’s – 1990’s computing power increased with the PC, and text-only displays evolved to graphical displays. Systems traders became some of the most profitable traders in large firms. Future decisions were being made on historical data (early predictive analytics). On balance well-designed systems traded by experienced traders were successful more often than not. Testing was faster, but still not fast (a single security run on a x386 IBM PC would take about 8 hours). As more traders began to use the same systems, the systems worked less well. This was due to an ‘observer effect’., with traders trying to exploit a particular advantage quickly causing the advantage to disappear! The system trader’s ‘edge’ or profitability was constantly declining, and new markets or circumstances were sought. ‘Program’ trades were accused of being the cause of the 1987 stock market crash.
There were some notable failures in market analysis – Fast Fourier Transformations being one. With enough computing power, you could fit a FFT to the market perfectly – but it would hardly ever work going forward. The FFT fails because it presumes a cyclical formula, and the markets while cyclical, are not predictably so. But an interesting phenomenon was that the better the fit in the FFT, the quicker and worse it would fall apart. That was due to the phenomenon of curve-fitting. ‘Fractals’ were all the rage later & failed just as miserably – same problem. As an aside, it explains why simpler linear models in regression analysis are frequently ‘better’ than a high-n polynomial spline fit to the data, particularly when considered for predictive analytics. The closer you fit the data, the less robust the model becomes and more prone to real-world failure.
Further advances in computing and computational statistics followed in the 1990’s-2000’s. Accurate real-time market data became widely available and institutionally ubiquitous, and time frames became shorter and shorter. Programs running on daily data were switched to multi-hour, hour, and then in intervals of minutes.The trend-following programs of the past became failures as the market became more choppy, and anti-trend (mean reversion) systems were popular. Enter the quants – the statisticians.(2) With fast, cheap, near-ubiquitous computing, the scope of the systems expanded. Now many securities could be analyzed at once, and imbalances exploited. Hence the popularity of ‘pairs’ trading. Real-time calculation of indices created index arbitrage, which were able to execute without human intervention.
The index arbitrage (index-arb) programs relied on speed and proximity to the exchanges to have advantages in execution. Statistical Arbitrage (Stat-arb) programs were the next development. These evolved into today’s High-Frequency-Trading programs (HFT’s) which dominate systems trading These programs are tested extensively on existing data, and then are let loose on the markets to be run – with only high-level oversight. They make thousands of trading decisions a second, incur real profits and losses, and compete against other HFT algorithms in a darwinian environment where the winners make money and are adapted further, and the losers dismissed with a digital death. Master governing algorithms coordinate individual algorithms. (4)
The floor traders, specialists, market-makers, and scores of support staff that once participated in the daily business have been replaced by glowing boxes sitting in a server rack next to the exchange.
Not to say that automated trading algorithms are perfect. A rogue algorithm with insufficient oversight caused a forced sale of Knight Capital Group (KCG) in 2012. (3) The lesson here is significant – there ARE going to be errors once automated algorithms are in greater use – it is inevitable.
So reviewing the history, what happened on wall st.?
1. First was descriptive analytics based upon historical data.
2. Graphical Interfaces were improved.
3. Improving technology led to more complicated algorithms which overfit the data. (WE ARE HERE)
4. Improving data accuracy led to real-time analytics.
5. Real time analytics led to shorter analysis timeframes
6. Shorter analysis timeframes led to dedicated trading algorithms operating with only human supervision
7. Master algorithms were created to coordinate the efforts of individual trading algorithms.
Next post, I’ll show the corollaries in health care and use it to predict where we are going.
(1) Jack Schwager, Market Wizards, Ed Seykota interview pp151-174.
(2) David Aronson, Evidence-based Technical Analysis, Wiley 2007
(3) Wall St. Journal, Trading Error cost firm $440 million, Marketbeat
(4)Personal communication, HFT trader (name withheld)
1. Descriptive Analytics based upon historical data.
This was the most basic use of data analysis. When newspapers printed price data (Open-High-Low-Close or OHLC), that data could be charted (on graph paper!) and interpreted using basic technical analysis, which was mostly lines drawn upon the chart. (1) Simple formulas such as year-over-year (YOY) percentage returns could be calculated by hand. This information was merely descriptive and had no bearing upon future events. To get information into a computer required data entry by hand, and operator errors could throw off the accuracy of your data. Computers lived in the accounting department, with the data being used to record position and profit and loss (P&L). At month’s end a large run of data would produce a computer-generated accounting statement.
A good analogue to this system would be older laboratory reporting systems where laboratory test values were sent to a dedicated lab computer. If the test equipment interfaced with the computer (via IEEE-488 & RS-232interfaces) the values were sent automatically. If not, data entry clerks had to enter these values. Once in the system, data could be accessed by terminals throughout the hospital. Normal ranges were typically included, with an asterisk indicating the value was abnormal. The computer database would be updated once a day (end of day type data). For more rapid results, you would have to go to the lab yourself and ask. On the operations side, a Lotus 1-2-3 spreadsheet on the finance team’s computer of quarterly charges, accounts receivable, and perhaps a few very basic metrics would be available to the finance department and CEO for periodic review.
For years, this delayed, descriptive data was the standard. Any inference would be provided by humans alone, who connected the dots. A rough equivalent would be HIMSS stage 0-1.
2. Improvements in graphics, computing speed, storage, connectivity.
Improvements in processing speed & power (after Moore’s Law), cheapening memory and storage prices, and improved device connectivity resulted in more readily available data. Near real-time price data was available, but relatively expensive ($400 per month or more per exchange with dedicated hardware necessary for receipt – a full vendor package could readily run thousands of dollars a month from a low cost competitior, and much more if you were a full service institution). An IBM PC XT of enough computing power & storage ($3000) could now chart this data. The studies that Ed Seykota ran on weekends would run on the PC – but analysis was still manual. The trader would have to sort through hundreds of ‘runs’ of the data to find the combination of parameters which led to the most profitable (successful) strategies, and then apply them to the market going forward. More complex statistics could be calculated – such as Sharpe Ratios, CAGR, and maximum drawdown – and these were developed and diffused over time into wider usage. Complex financial products such as options could now be priced more accurately in near-real time with algorithmic advances (such as the binomial pricing model).
The health care corollary would be in-house early electronic record systems tied in to the hospital’s billing system. Some patient data was present, but in siloed databases with limited connectivity. To actually use the data you would ask IT for a data dump which would then be uploaded into Excel for basic analysis. Data would come from different systems and combining it was challenging. Because of the difficulty in curating the data (think massive spreadsheets with pivot tables), this could be a full-time job for an analyst or team of analysts, and careful selection of what data was being followed and what was discarded would need to be considered, a priori. The quality of the analysis improved, but was still human labor intensive, particularly because of large data sets & difficulty in collecting the information. For analytic tools think Excel by Microsoft or Minitab.
This corresponds to HIMSS stage 2-3.
3. Further improvement in technology correlates with algorithmic improvement.
With new levels of computing power, analysis of data became quick and relatively cheap allowing automated analysis. Taking the same data set of computed results from price/time data that was analyzed by hand before; now apply an automated algorithm to run through ALL possible combinations of included parameters. This is brute-force optimization. The best solve for the data set is found, and a trader is more confident that the model will be profitable going forward.
For example, consider ACTV(2). Running a brute force optimization on this security with a moving average over the last 2 years yields a profitable trading strategy that returns 117% with the ideal solve. Well, on paper that looks great. What could be done to make it even MORE profitable? Perhaps you could add a stop loss. Do another optimization and theoretical return increases. Want more? Sure. Change the indicator and re-optimize. Now your hypothetical return soars. Why would you ever want to do anything else? (3,4)
But it’s not as easy as it sounds. The best of the optimized models would work for a while, and then stop. The worst would immediately diverge and lose money from day 1 – never recovering. Most importantly : what did we learn from this experience? We learned that how the models were developed mattered. And to understand this, we need to go into a bit of math.
Looking at security prices, you can model (approximate) the price activity as a function, F(X)= the squiggles of a chart. The model can be as complex or simple as desired. Above, we start with a simple model (the moving average), and make it progressively more complex adding additional rules and conditions. As we do so, the accuracy of the model increases, so the profitability increases as well. However, as we increase the accuracy of the model, we use up degrees of freedom, making the model more rigid and less resilient.
Hence the system trader’s curse – everything works great on paper, but when applied to the market, the more complex the rules, and the less robustly the data is tested, the more likely the system will fail due to a phenomenon known as over-fitting. Take a look at the 3D graph below which shows a profitability model of the above analysis:
You will note that there is a spike in profitability using a 5 day moving average at the left of the graph, but profitability sharply falls off after that, rises a bit, and then craters. There is a much broader plateau of profitability in the middle of the graph, where many values are consistently and similarly profitable. Changes in market conditions could quickly invalidate the more profitable 5 day moving average model, but a model with a value chosen in the middle of the chart might be more consistently profitable over time. While more evaluation would need to be done, the less profitable (but still profitable) model is said to be more ‘Robust’.
To combat this, better statistical sampling methods were utilized, namely cross-validation where an in-sample set is used to test an out-of-sample set for performance. This gave a system which was less prone to immediate failure, i.e. more robust. A balance between profitability and robustness can be struck, netting you the sweet spot in the Training vs. Test-set performance curve I’ve posted before.
So why didn’t everyone do this? Quick answer: they did. And by everyone analyzing the same data set of end-of-day historical price data in the same way, many people began to reach the same conclusions as each other. This created an ‘observer effect’ where you had to be first to market to execute your strategy, or trade in a market that was liquid enough (think the S&P 500 index) that the impact of your trade (if you were a small enough trader – doesn’t work for a large institutional trader) would not affect the price. Classic case of ‘the early bird gets the worm’.
The important point is that WE ARE HERE in healthcare. We have moderately complex computer systems that have been implemented largely due to Meaningful Use concerns, bringing us to between HIMSS stages 4-7. We are beginning to use the back ends of computer systems to interface with analytic engines for useful descriptive analytics that can be used to inform business and clinical care decisions. While this data is still largely descriptive, some attempts at predictive analytics have been made. These are largely proprietary (trade secrets) but I have seen some vendors beginning to offer proprietary models to the healthcare community (hospitals, insurers, related entities) which aim at predictive analytics. I don’t have specific knowledge of the methods used to create these analytics, but after the experience of Wall Street, I’m pretty certain that a number of them are going to fall into the overfitting trap. There are other, more complex reasons why these predictive analytics might not work (and conversely, good reasons why they may), which I’ll cover in future posts.
One final point – application of predictive analytics to healthcare will succeed in the area where it fails on Wall Street for a specific reason. On Wall Street, the relationship once discovered and exploited causes the relationship to disappear. That is the nature of arbitrage – market forces reduce arbitrage opportunities since they represent ‘free money’ and once enough people are doing it, it is no longer profitable. However, biological organisms don’t response to gaming the system in that manner. For a conclusive diagnosis, there may exist an efficacious treatment that is consistently reproducible. In other words, for a particular condition in a particular patient with a particular set of characteristics (age, sex, demographics, disease processes, genetics) if accurately diagnosed and competently executed, we can expect a reproducible biologic response, optimally a total cure of the individual. And that reproducible response applies to processes present in the complex dynamic systems that comprise our healthcare delivery system. That is where the opportunity lies in applying predictive analytics to healthcare.
(1) Technical Analysis of Stock Trends, Edwards and Magee, 8th Edition, St. Lucie Press
(2) ACTIVE Technologies, acquired (taken private) by Vista Equity Partners and delisted on 11/15/2013. You can’t trade this stock.
(3) Head of Trading, First Chicago Bank, personal communication
(4) Reminder – see the disclaimer for this blog! And if you think you are going to apply this particular technique to the markets to be the next George Soros, I’ve got a piece of the Brooklyn Bridge to sell you.
This a somewhat challenging post with cross-discipline correlations, some unfamiliar terminology, and concepts. There is a payoff!
The crux of this discussion is time. Understanding the progression towards shorter and shorter time frames on Wall Street enables us to draw parallels and differences in medical care delivery particularly pertaining to processes and data analytics. This is relevant because some vendors tout real-time capabilities in health care data analysis. Possibly not as useful as one thinks.
In trading, the best profit one is a risk-less one. A profit that occurs by simply being present, is reliable, and reproducible, and exposes the trader to no risk. Meet arbitrage. Years ago, it was possible for the same security to be trading at different prices on different exchanges as there was no central marketplace. A network of traders could execute a buy of a stock for $10 in New York, and then sell those same shares on the Los Angeles exchange for $11. If one imagines a 1000 share transaction, a $1 profit per share yields $1000. It was made by the head trader holding up two phones to his head and saying ‘buy’ into one and ’sell’ into the other.* These relationships could be exploited over longer periods of time and represented an information deficit. However, as more traders learned of them, the opportunities became harder to find as greater numbers pursued them. This price arbitrage kept prices reasonably similar before centralized, computerized exchanges and data feeds.
As information flow increased, organizations became larger and more effective, and time frames for executing profitable arbitrages decreased. This led traders to develop simple predictive algorithms, like Ed Seykota did, detailed in part 1. New instruments re-opened the profit possibility for a window of time, which eventually closed. The development of futures, options, indexes, all the way to closed exchanges (ICE, etc…) created opportunities for profit which eventually became crowded. Since the actual arbitrages were mathematically complex (futures have an implied interest rate, options require a solution of multiple partial differential equations, and indexes require summing instantaneously hundreds of separate securities) a computational model was necessary as no individual could compute the required elements quickly enough to profit reliably. With this realization, it was only a matter of time before automated trading (AT) happened, and evolved into high-frequency trading with its competing algorithms operating without human oversight on millisecond timeframes.
The journey from daily prices to ever shorter prices over the trading day to millisecond prices was driven by availability of good data and reliable computing which could be counted to act on those flash prices. Once a game of location (geographical arbitrage) turned into a game of speed (competitive pressures on geographical arbitrage) turned into a game of predictive analytics (proprietary trading and trend following) turned into a more complex game of predictive analytics (statistical arbitrage) was then ultimately turned back into a game of speed and location (High frequency trading).
The following chart shows a probability analysis of an ATM straddle position on IBM. This is an options position. It is not important to understand the instrument, only to understand what the image shows. For IBM, the expected variance that exists in price at one standard deviation (+/- 1 s.d.) is plotted in below. As time (days) increases along the X axis, the expected range widens, or becomes less accurate.
credit: TD Ameritrade
Is there a similar corollary for health care?
First, recognize the distinction between the simpler price-time data which exists in the markets, vs the rich, complex multivariate data in healthcare.
Second, assuming a random walk hypothesis , security price movement is unpredictable, and at best can only be calculated so that the next price will be in a range defined by a number of standard deviations according to one’s model as seen above in the picture. You cannot make this argument in healthcare. This is because the patient’s disease is not a random walk. Disease follows proscribed pathways and natural histories which allow us to make diagnoses and implement treatment options.
It is instructive to consider Clinical Decision Support tools. Please note that these tools are not a substitute for expert medical advice (and my mention does not employ endorsement). See Esagil and diagnosis pro. If you enter “abdominal pain” into either of the algorithms, you’ll get back a list of 23 differentials (woefully incomplete) in Esagil and 739 differentials (more complete, but too many to be of help) in Diagnosis Pro. But this is a typical presentation to a physician – a patient complains of “abdominal pain” and the differential must be narrowed.
At the onset, there is a wide differential diagnosis. The possibility that the pain is a red herring and the patient really has some other, unsuspected, disease must be considered. While there are a good number of diseases with a pathognomonic presentation, uncommon presentations of common diseases are more frequent than common presentations of rare diseases.
In comparison to the trading analogy above, where expected price movement is generally restricted to a quantifiable range based on the observable statistics of the security over a period of time, for a de novo presentation of a patient, this could be anything, and the range of possibilities is quite large.
Take, for example, a patient that presents to the ER complaining “I don’t feel well.” When you question them, they tell you that they are having severe chest pain that started an hour and a half ago. That puts you into the acute chest pain diagnostic tree.
With acute chest pain, there is a list of differentials that needs to be excluded (or ‘ruled out’), some quite serious. A thorough history and physical is done, taking 10-30 minutes. Initial labs are ordered (5-30 minutes if done in a rapid, in-ER test, longer if sent to the main laboratory) an EKG and CXR (chest X-ray) are done for their speed,(10 minutes for each) and the patient is sent to CT for a CTA (CT Angiogram) to rule out a PE (Pulmonary embolism). This is a useful test, because it will not only show the presence or absence of a clot, but will also allow a look at the lungs to exclude pneumonias, effusions, dissections, and malignancies. Estimate that the wait time for the CTA is at least 30 minutes.
The ER doctor then reviews the results (5 minutes)- troponins are negative, excluding a heart attack (MI), the CT scan eliminated PE, Pneumonia, Dissection, Pneumothorax, Effusion, malignancy in the chest. The Chest X-Ray excludes fracture. The normal EKG excludes arrhythmia, gross valvular disease, and pericarditis. The main diagnoses left are GERD, Pleurisy, referred pain, and anxiety. ER doctor goes back to the patient (10 minutes) , patient doesn’t appear anxious & no stressors, so panic attack unlikely. No history of reflux, so GERD unlikely. No abdominal pain component, and labs were negative, so abdominal pathologies unlikely. Point tenderness present on the physical exam at the costochondral junction – and the patient is diagnosed with costochondritis. The patient is then discharged with a prescription for pain control. (30 minutes).
Ok, if you’ve stayed with me, here’s the payoff.
As we proceed down the decision tree, the number of possibilities narrows in medicine.
In comparison, price-time data – in which the range of potential prices increase as you proceed forward in time.
So, in healthcare the potential diagnosis narrows as you proceed down the x-axis of time. Therefore, time is both one’s friend and enemy – friend as it provides for diagnostic and therapeutic interventions which establish the patient’s disease process; enemy as payment models in medicine favor making that diagnostic and treatment process as quick as possible (when a hospital inpatient).
We’ll continue this in part IV and compare it relevance to portfolio trading.
*As an aside, the phones in trading rooms had a switch on the handheld receiver – you would push them in to talk. That way, the other party would not know that you were conducting an arbitrage! They were often slammed down and broken by angry traders – one of the manager’s jobs was to keep a supply of extras in his desk, and they were not hard-wired but plugged in by a jack expressly for that purpose!
**Yes, for the statisticians reading this, I know that there is an implication of a gaussian distribution that may not be proven. I would suspect the successful houses have modified for this and have instituted non-parametric models as well. Again, this is not a trading or financial advice blog.