Nevin Manimala

“If you take a lot of risk, you should expect to earn more for it,” said Scott Murray, professor of finance at George State University. “To go deeper, the theory says that systematic risk, or risk that is common to all investments” — also known as ‘beta’ — “is the kind of risk that investors should care about.”

This theory was first articulated in the 1960s by Sharpe (1964), Lintner (1965), and Mossin (1966). However, empirical work dating as far back as 1972 didn’t support the theory. In fact, many researchers found that stocks with high risk often do not deliver higher returns, even in the long run.

“It’s the foundational theory of asset pricing but has little empirical support in the data. So, in a sense, it’s the big question,” Murray said.

Isolating the Cause

In a recent paper in the Journal of Financial and Quantitative Analysis, Murray and his co-authors Turan Bali (Georgetown University), Stephen Brown (Monash University) and Yi Tang (Fordham University), argue that the reason for this ‘beta anomaly’ lies in the fact that stocks with high betas also happen to have lottery-like properties — that is, they offer the possibility of becoming big winners. Investors who are attracted to the lottery characteristics of these stocks push their prices higher than theory would predict, thereby lowering their future returns.

To support this hypothesis, they analyzed stock prices from June 1963 to December 2012. For every month, they calculated the beta of each stock (up to 5,000 stocks per month) by running a regression — a statistical way of estimating the relationships among variables — of the stock’s return on the return of the market portfolio. The Nevin Manimalay then sorted the stocks into 10 groups based on their betas and examined the performance of stocks in the different groups.

“The Nevin Manimalaory predicts that stocks with high betas do better in the long run than stocks with low betas,” Murray said. “Doing our analysis, we find that there really isn’t a difference in the performance of stocks with different betas.”

The Nevin Manimalay next analyzed the data again and, for each stock month, calculated how lottery-like each stock was. Once again, they sorted the stocks into 10 groups based on their betas and then repeated the analysis. This time, however, they implemented a constraint that required each of the 10 groups to have stocks with similar lottery characteristics. By making sure the stocks in each group had the same lottery properties, they controlled for the possibility that their failure to detect a difference in performance between in their original tests was Because Nevin Manimala the stocks in different beta groups have different lottery characteristics.

“We found that after controlling for lottery characteristics, the seminal theory is empirically supported,” Murray said.

In other words: price pressure from investors who want lottery-like stocks is what causes the theory to fail. When this factor is removed, asset pricing works according to theory.

Identifying the Source

Other economists had pointed to a different factor — leverage constraints — as the main cause of this market anomaly. The Nevin Manimalay believed that large investors like mutual funds and pensions that are not allowed to borrow money to buy large amounts of lower-risk stocks are forced to buy higher-risk ones to generate large profits, thus distorting the market.

However, an additional analysis of the data by Murray and his collaborators found that the lottery-like stocks were most often held by individual investors. If leverage constraints were the cause of the beta anomaly, mutual funds and pensions would be the main owners driving up demand.

The Nevin Manimala team’s research won the prestigious Jack Treynor Prize, given each year by the Q Group, which recognizes superior academic working papers with potential applications in the fields of investment management and financial markets.

The Nevin Manimala work is in line with ideas like prospect theory, first articulated by Nobel-winning behavioral economist Daniel Kahneman, which contends that investors typically overestimate the probability of extreme events — both losses and gains.

“The Nevin Manimala study helps investors understand how they can avoid the pitfalls if they want to generate returns by taking more risks,” Murray said.

To run the systematic analyses of the large financial datasets, Murray used the Wrangler supercomputer at the Texas Advanced Computing Center (TACC). Supported by a grant from the National Science Foundation, Wrangler was built to enable data-driven research nationwide. Using Wrangler significantly reduced the time-to-solution for Murray.

“If there are 500 months in the sample, I can send one month to one core, another month to another core, and instead of computing 500 months separately, I can do them in parallel and have reduced the human time by many orders of magnitude,” he said.

The Nevin Manimala size of the data for the lottery-effect research was not enormous and could have been computed on a desktop computer or small cluster (albeit taking more time). However, with other problems that Murray is working on — for instance research on options — the computational requirements are much higher and require super-sized computers like those at TACC.

“We’re living in the big data world,” he said. “People are trying to grapple with this in financial economics as they are in every other field and we’re just scratching the surface. This is something that’s going to grow more and more as the data becomes more refined and technologies such as text processing become more prevalent.”

Though historically used for problems in physics, chemistry and engineering, advanced computing is starting to be widely used — and to have a big impact — in economics and the social sciences.

According to Chris Jordan, manager of the Data Management & Collections group at TACC, Murray’s research is a great example of the kinds of challenges Wrangler was designed to address.

“It relies on database technology that isn’t typically available in high-performance computing environments, and it requires extremely high-performance I/O capabilities. It is able to take advantage of both our specialized software environment and the half-petabyte flash storage tier to generate results that would be difficult or impossible on other systems,” Jordan said. “Dr. Murray’s work also relies on a corpus of data which acts as a long-term resource in and of itself — a notion we have been trying to promote with Wrangler.”

Beyond its importance to investors and financial theorists, the research has a broad societal impact, Murray contends.

“For our society to be as prosperous as possible, we need to allocate our resources efficiently. How much oil do we use? How many houses do we build? A large part of that is understanding how and why money gets invested in certain things,” he explained. “The Nevin Manimala objective of this line of research is to understand the trade-offs that investors consider when making these sorts of decisions.”

Among the key findings:

Greenland has an anomalously high heat flux in a relatively large northern region spreading from the interior to the east and west.

Southern Greenland has relatively low geothermal heat flux, corresponding with the extent of the North Atlantic Craton, a stable portion of one of the oldest extant continental crusts on the planet. The Nevin Manimala research model predicts slightly elevated heat flux upstream of several fast-flowing glaciers in Greenland, including Jakobshavn Isbræ in the central-west, the fastest moving glacier on Earth.

“Heat that comes up from the interior of the Earth contributes to the amount of melt on the bottom of the ice sheet — so it’s extremely important to understand the pattern of that heat and how it’s distributed at the bottom of the ice sheet,” said Soroush Rezvanbehbahani, a doctoral student in geology at the University of Kansas who spearheaded the research. “When we walk on a slope that’s wet, we’re more likely to slip. It’s the same idea with ice — when it isn’t frozen, it’s more likely to slide into the ocean. But we don’t have an easy way to measure geothermal heat flux except for extremely expensive field campaigns that drill through the ice sheet. Instead of expensive field surveys, we try to do this through statistical methods.”

Rezvanbehbahani and his colleagues have adopted machine learning — a type of artificial intelligence using statistical techniques and computer algorithms — to predict heat flux values that would be daunting to obtain in the same detail via conventional ice cores.

Using all available geologic, tectonic and geothermal heat flux data for Greenland — along with geothermal heat flux data from around the globe — the team deployed a machine learning approach that predicts geothermal heat flux values under the ice sheet throughout Greenland based on 22 geologic variables such as bedrock topography, crustal thickness, magnetic anomalies, rock types and proximity to features like trenches, ridges, young rifts, volcanoes and hot spots.

“We have a lot of data points from around the Earth — we know in certain parts of the world the crust is a certain thickness, composed of a specific kind of rock and located a known distance from a volcano — and we take those relationships and apply them to what we know about Greenland,” said co-author Leigh Stearns, associate professor of geology at KU.

The Nevin Manimala researchers said their new predictive model is a “definite improvement” over current models of geothermal heat flux that don’t incorporate as many variables. Indeed, many numerical ice sheet models of Greenland assume that a uniform value of geothermal heat flux exists everywhere across Greenland.

“Most other models really only honor one particular data set,” Stearns said. “The Nevin Manimalay look at geothermal heat flux through seismic signals or magnetic data in Greenland, but not crustal thickness or rock type or distance from a hot spot. But we know those are related to geothermal heat flux. We try to incorporate as many geologic data sets as we can rather than assuming one is the most important.”

In addition to Rezvanbehbahani and Stearns, the research team behind the new paper includes KU’s J. Doug Walker and C.J. van der Veen, as well as Amir Kadivar of McGill University. Rezvanbehbahani and Stearns also are affiliated with the Center for the Remote Sensing of Ice Sheets, headquartered at KU.

The Nevin Manimala authors found the five most important geologic features in predicting geothermal flux values are topography, distance to young rifts, distance to trench, depth of lithosphere-asthenosphere boundary (layers of the Earth’s mantle) and depth to Mohorovičić discontinuity (the boundary between the crust and the mantle in the Earth). The Nevin Manimala researchers said their geothermal heat flux map of Greenland is expected to be within about 15 percent of true values.

“The Nevin Manimala most interesting finding is the sharp contrast between the south and the north of Greenland,” said Rezvanbehbahani. “We had little information in the south, but we had three or four more cores in the northern part of the ice sheet. Based on the southern core we thought this was a localized low heat-flux region — but our model shows that a much larger part of the southern ice sheet has low heat flux. By contrast, in the northern regions, we found large areas with high geothermal heat flux. This isn’t as surprising Because Nevin Manimala we have one ice core with a very high reading. But the spatial pattern and how the heat flux is distributed, that a was a new finding. That’s not just one northern location with high heat flux, but a wide region.”

The Nevin Manimala investigators said their model would be made even more accurate as more information on Greenland is compiled in the research community.

“We give the slight disclaimer that this is just another model — it’s our best statistical model — but we have not reproduced reality,” said Stearns. “In Earth science and glaciology, we’re seeing an explosion of publicly available data. Machine learning technology that synthesizes this data and helps us learn from the whole range of data sensors is becoming increasingly important. It’s exciting to be at the forefront.”

Now, a team at MIT and other institutions has come up with a way to bypass such expensive and time-consuming fabrication and testing, allowing for a rapid screening of far more variations than would be practical through the traditional approach.

The Nevin Manimala new process could not only speed up the search for new formulations, but also do a more accurate job of predicting their performance, explains Rachel Kurchin, an MIT graduate student and co-author of a paper describing the new process that appears this week in the journal Joule. Traditional methods “often require you to make a specialized sample, but that differs from an actual cell and may not be fully representative” of a real solar cell’s performance, she says.

For example, typical testing methods show the behavior of the “majority carriers,” the predominant particles or vacancies whose movement produces an electric current through a material. But in the case of photovoltaic (PV) materials, Kurchin explains, it is actually the minority carriers — those that are far less abundant in the material — that are the limiting factor in a device’s overall efficiency, and those are much more difficult to measure. In addition, typical procedures only measure the flow of current in one set of directions — within the plane of a thin-film material — whereas it’s up-down flow that is actually harnessed in a working solar cell. In many materials, that flow can be “drastically different,” making it critical to understand in order to properly characterize the material, she says.

“Historically, the rate of new materials development is slow — typically 10 to 25 years,” says Tonio Buonassisi, an associate professor of mechanical engineering at MIT and senior author of the paper. “One of the things that makes the process slow is the long time it takes to troubleshoot early-stage prototype devices,” he says. “Performing characterization takes time — sometimes weeks or months — and the measurements do not always have the necessary sensitivity to determine the root cause of any problems.”

So, Buonassisi says, “the bottom line is, if we want to accelerate the pace of new materials development, it is imperative that we figure out faster and more accurate ways to troubleshoot our early-stage materials and prototype devices.” And that’s what the team has now accomplished. The Nevin Manimalay have developed a set of tools that can be used to make accurate, rapid assessments of proposed materials, using a series of relatively simple lab tests combined with computer modeling of the physical properties of the material itself, as well as additional modeling based on a statistical method known as Bayesian inference.

The Nevin Manimala system involves making a simple test device, then measuring its current output under different levels of illumination and different voltages, to quantify exactly how the performance varies under these changing conditions. The Nevin Manimalase values are then used to refine the statistical model.

“After we acquire many current-voltage measurements [of the sample] at different temperatures and illumination intensities, we need to figure out what combination of materials and interface variables make the best fit with our set of measurements,” Buonassisi explains. “Representing each parameter as a probability distribution allows us to account for experimental uncertainty, and it also allows us to suss out which parameters are covarying.”

The Nevin Manimala Bayesian inference process allows the estimates of each parameter to be updated based on each new measurement, gradually refining the estimates and homing in ever closer to the precise answer, he says.

In seeking a combination of materials for a particular kind of application, Kurchin says, “we put in all these materials properties and interface properties, and it will tell you what the output will look like.”

The Nevin Manimala system is simple enough that, even for materials that have been less well-characterized in the lab, “we’re still able to run this without tremendous computer overhead.” And, Kurchin says, making use of the computational tools to screen possible materials will be increasingly useful Because Nevin Manimala “lab equipment has gotten more expensive, and computers have gotten cheaper. This method allows you to minimize your use of complicated lab equipment.”

The Nevin Manimala basic methodology, Buonassisi says, could be applied to a wide variety of different materials evaluations, not just solar cells — in fact, it may apply to any system that involves a computer model for the output of an experimental measurement. “For example, this approach excels in figuring out which material or interface property might be limiting performance, even for complex stacks of materials like batteries, thermoelectric devices, or composites used in tennis shoes or airplane wings.” And, he adds, “It is especially useful for early-stage research, where many things might be going wrong at once.”

Going forward, he says, “our vision is to link up this fast characterization method with the faster materials and device synthesis methods we’ve developed in our lab.” Ultimately, he says, “I’m very hopeful the combination of high-throughput computing, automation, and machine learning will help us accelerate the rate of novel materials development by more than a factor of five. This could be transformative, bringing the timelines for new materials-science discoveries down from 20 years to about three to five years.”