Best Practices in Research for Quantitative Equity Strategies

Joseph A. Cerniglia, Frank J. Fabozzi and Petter N. Kolm

The Journal of Portfolio Management Special QES Issue 2016, 42 (5) 135-143; DOI:

https://doi.org/10.3905/jpm.2016.42.5.135

Abstract

The authors examine the research process and principles underlying successful models used in

quantitative equity strategies. They identify three key factors they see contributing to improved empirical

work: 1) making research design a top priority, 2) making new, more extensive datasets available, and 3)

making advances in computational areas such as econometrics, machine learning, and statistics. The authors

explain these key factors and also share insights on how to integrate market dynamics, data, research design,

advance modeling techniques, and economic/financial introspection into the research process.

In this article, we examine the research process and principles underlying successful models used in

quantitative equity strategies. The research process is at the heart of the development of successful

quantitative strategies. Key factors for their success are the availability of more and better data, advances

in computational and econometric methods, and better understanding of how to enhance judgment in

the research process. Our discussion does not provide rules to follow, but rather tenets that emerged from

our collective experience of researching and developing quantitative models.

By identifying and examining the characteristics of quantitative strategies, we attempt to highlight some

best practices in quantitative modeling and aspire to outline a broader paradigm for building successful

models. These characteristics may not be strictly statistical or mathematical in nature, but rather they emphasize

the integration of market dynamics, data, research design, modeling techniques, and economic and

financial judgment.

In Superforecasting, Tetlock and Gardner [20151 state that "foresight isn't a mysterious gift bestowed at

birth. It is the product of particular ways of thinking, of gathering information, of updating beliefs." In this

article, we share some insights on accomplishing this objective by providing a framework and some thoughts

on building quantitative forecasting models. Our discussion centers on developing quantitative models

regardless of asset class, but our examples draw heavily from equities. Although we focus on

quantitative research methodologies, we think that some of these ideas are valuable for a fundamental

research process.

1
WHAT ARE QUANTITATIVE MODELS?

In this article, we refer to quantitative modeling in a broader sense. A quantitative strategy is a systematic, data-

and model-based approach to making investment decisions. We can further qualify quantitative strategies by

their underlying core characteristics. By examining these core characteristics, we can attempt to identify

some best practices in quantitative modeling to develop a paradigm that will lead to building successful

models.

2
The most important characteristic of the quantitative modeling approach is the scientific approach. This

approach provides a paradigm that guides and informs empirical work. Similar to other fields that take a

scientific approach—including natural sciences, medicine, and social sciences—this approach in quantitative

modeling attempts to describe, inquire, and interpret with precision. The characteristics of a scientific

approach as it relates to quantitative equity strategy modeling include the following:

• Development of a thoughtful hypothesis or thesis to be evaluated

• Use of empirical work to attempt to put precision around investment decisions and economic

reasoning

• Reliance on high standards of analytical rigor

• Use of sensitivity analysis to challenge assumptions and context in which the strategy was developed

• Incorporation of adjustments to the strategy based on judgment

• Ability to explicitly measure results

• Incorporation of revisions or updates to the model as new information becomes available

HOW ARE QUANTITATIVE STRATEGIES DEVELOPED?

Dating back to the 17th century, the scientific method is an approach for examining and understanding

phenomena, developing new theories, or modifying or integrating existing theories based on the

presentation of empirical and measurable evidence subject to specific principles of reasoning. Research based

on the scientific method typically takes steps to 1) define a question, 2) collect information and resources, 3)

form an explanatory hypothesis, 4) test the hypothesis by performing experiments and collecting data in a

reproducible manner, 5) analyze the data, and 6) interpret the data and draw conclusions. After drawing

these conclusions, the researcher may then go back to reformulate the explanatory hypothesis and repeat

the (3)-(6) cycle again.

3
Within the context of investment management, empirical analysis uses data and tools to design, research, and

evaluate hypotheses/models. The primary function of empirical research is to create (some) evidence for

trading models. A large part of empirical analysis is research design. A well-thought-out research design

provides support and credibility for validating the investment insights underlying trading models.

4
• Mullifactor strategies: Models that invest in equities based on multiple characteristics that replicate how
investors make decisions. These strategies have an edge in their superior processing of information and
identification of differentiated insights.
• Allocation decisions: Strategies that make decisions based on allocations to different countries, sectors,
factor timing, regime switching, etc.
• Stock—specific strategies: Strategies that focus on the information specific to an individual equity security.
• Factor strategies: Rule-based strategies that trade well-known equity risk premiums to earn returns.
• Event studies: Strategies that generate alpha from specific events.
• Market microstructure strategies: Strategies that exploit profitable opportunities arising from the trading
flows and dynamics of equity markets.
• Statistical arbitrage: Strategies that exploit systematic relationships among equity securities with
similar characteristics. In contrast to pure riskless arbitrage, statistical arbitrage is a risky form of
strategy.
• Textual strategies: Quantitative strategies that trade based on qualitative textual signals, such as news
reports, company documents, or Internet searches.
• Thematic/macro strategies: Strategies that trade baskets of equity securities based on broad themes in the
economic environment. technology, demographics, or other fields.

Exhibit 1 lists different types of quantitative strategies. The typical steps in developing these strategies are as

follows: 1) formulate trading ideas and strategies, 2) develop signals, 3) acquire and process data, 4) analyze

the signals, 5) build the strategy, 6) evaluate the strategy, 7) backtest the strategy, and 8) implement the

strategy. We next examine each of these steps.

• Formulating Investment Ideas and Strategies

A successful quantitative strategy often starts as an idea based on economic intuition, a market insight, or an

anomaly. Background research can be helpful for understanding what others have tried or implemented in

the past.

To distinguish between a trading idea and a quantitative strategy, we look at the economic motivation for

each. A trading idea has a shorter-term horizon, often associated with a specific event or mispricing. A

quantitative strategy has a longer time span and exploits opportunities to process information better,
receive premiums associated with anomalies, or identify mispricings.

• Developing Signals

After having established the idea of the strategy, we move from the economic concepts to the construction of

signals that may be able to capture our intuition. Signals provide building blocks for the model used to

5
create an investment strategy.

Built from data, signals are quantitative measures that represent an investment idea. How signals are built

varies depending on the investment thesis and the data representing the thesis. For example, a

quantitative signal could be based on a stock's underlying characteristics such as its return on equity

or valuation ratio. A sentiment signal could be developed from unstructured text from various company-

issued reports or news about the company.

• Acquiring and Processing Data

Data are critical to a strategy's success. A strategy relies on accurate and clean data to build signals. Data need

to be carefully stored in an infrastructure that is scalable and flexible. Upon acquiring new data sources,

researchers expand the information set necessary to create new insights.

6
 • Ana lyzing the Signa ls

Researchers perform a variety of statistical tests and econometric techniques on the data to evaluate the

empirical properties of signals. This empirical research is used to understand the risk-and-return potential

of a signal. For example, a researcher might be interested in statistically testing whether a signal's Sharpe

ratio is larger than 1. This analysis may form the basis for building a more complete trading strategy.

• B uilding the S tr a tegy

A model represents a mathematical or systematic specification of the trading strategy. There are two

important considerations in this specification: 1) the selection of the specific signals and 2) how these

signals are combined. Both considerations may be motivated by the economic intuition driving the

trading strategy.

• Evaluating, Testing, and Implementing the Strategy

The final steps involve assessing the estimation, specification, and forecast quality of the model. This analysis

includes examining the goodness of fit (often done in sample), forecasting ability (often done out of sample),

and sensitivity and risk characteristics of the model.

Empirical validation and testing are key drivers in the development of quantitative trading strategies. They

bridge the gap between stylized financial models and the real world represented by the markets.

Financial models are often crude approximations of reality—with regimes in which they work acceptably, and

regimes in which they do not work at all or work very poorly at best. Careful systematic empirical research can

help identify these regimes. The researcher's judgment and experience become a critical factor in this step.

MODELS AND JUDGMENT

The well-known statistician George Box stated, "All models are wrong; some models are useful" (Box [19761).

Models simplify the world around us through idealization. Naturally, this idealization describes the most

salient features of markets, and it is important to note that not every market dynamic is included in a model.

The construction of idealized representations of the financial markets is a vital part of academic and

practitioner research.

Although models are quantitative in nature, the research process is subject to data and design decisions that

are more qualitative in nature. Judgment calls include deciding how to cleanse the data, how to select a

7
specific model, how to aggregate signals, and which risk measures to rely on. Researchers make these

decisions, of course, based on their experience and preferences. Besides the research process itself,

judgment is also prevalent in the feedback mechanism of backtesting and running the strategy.

Generally, most quantitative models are based on two approaches of thinking—hypothesis based

(deductive) and pattern based (inductive). Each approach requires a different model-building research

process. For the hypothesis-based approach, the starting point is some insight about why a trading

opportunity exists. It is dependent on an economic thesis or hypothesis on how the market works or

why the opportunity exists. Frequently, the "story" precedes the empirical work.

The second approach is inductive or pattern based. This approach is exploratory in nature, and the discovery

of insights emerges from the empirical work. A key feature is that learning occurs throughout the process. In

this approach, it is critical to be able to distinguish between correlation and causation. Are measured

statistical correlations spurious or causal? Understanding underlying economic mechanisms and theory

may provide insights to this question.

Best practices involve understanding how to make better decisions in the research design process. It is

useful to draw on sciences from other disciplines that study decision making, often in experimental settings;

these include psychology, philosophy, and organizational behavior.

A good question to ask is, "How do we make better decisions in the development of quantitative

strategies?" We compiled the following list of attributes from the research of various experts in the areas

of decision sciences, including Learner [1978, 1983], Tetlock and Gardner [2015], and Tversky and

Kahneman [1974], as well as from our own personal experience:

• Start a research project with scientific caution.

• Understand the assumptions underlying the research methodology decisions and why those

assumptions were chosen.

• Formulate an alternative hypothesis.

• Make (reasonable) changes to key assumptions to evaluate their robustness.

• Look for contrary evidence of the findings.

8
 • Tweak the research question being asked and try answering this revised question as a way of

gaining additional perspective.

• Focus on causation and probabilities, rather than statistical correlations.

• Break down the investment thesis into its underlying assumptions and scrutinize each

assumption.

• Examine what you know about the investment thesis from what you do not know or cannot know.

The benefits of using quantitative models extend beyond pure quantitative trading. These models provide

valuable analytical tools in a traditional, fundamental investment decision-making process. It is important to

differentiate between how purely quantitative investors use model forecasts and how fundamental investors

use them. For quantitative investors, model forecasts produce an expected return forecast on a security or

a set of securities. For fundamental investors, model forecasts create new insights to synthesize

with other qualitative information (e.g., management meetings and industry strategy) being acquired to

make investment decisions. Working with fundamental investors, we apply quantitative models to

understand complex relationships, to verify investment theses, and to discover new opportunities.

Collaboration of quantitative researchers with fundamental investors is a social experience that can create an

"investment edge." The process of building a quantitative model jointly produces unique investment insights.

Numerous studies and anecdotes provide evidence that combining computer-based forecasting and

human judgment results in better outcomes. We can draw on the literature from other disciplines to assist in

providing insight into how to better integrate the two. For example, in "freestyle chess," a chess tournament in

which players are open to consult any resource available to assist them, the winners of tournaments are

humans paired with machines—beating machines only, or human experts alone, or machine alone (Cowen

[2013]). The key to the winners' success is being able to synthesize information from multiple sources, while

recognizing the strengths and weaknesses of each approach.

TAXONOMY OF QUANTITATIVE EQUITY STRATEGIES

Quantitative strategies exist across different markets. The characteristics underlying these strategies vary

substantially. We categorize strategies along a number of dimensions, such as the asset class, type of

securities, horizon, trading style, and investment philosophy. Each of these categories influences the

quantitative modeling process, often starting with the research design, data, modeling techniques, and

9
evaluation methods.

Commonality in these traits allows us to classify strategies into groups. In Exhibit 1, we attempt to create a

simple taxonomy of quantitative strategies. There is some overlap in this classification because strategies

share common traits. We also considered how investors implement the strategies.

Quantitative investment strategies differ in their motivation to trade, the frequency of trading, information used

to trade, and the markets traded. The strategies employ different holding periods and trading frequencies—

the latter of which can occur in milliseconds, or extend to months or years. Separately, the holding period

of each trade varies along similar horizons. Both trading frequencies and holding periods are functions

of the investment theory underlying the strategy and the empirical results uncovered in research.

TECHNIQUES FOR EMPIRICAL WORK

In addition to insights about the market, quantitative computational methods are critical for success. Many of

the traditional financial econometric techniques continue to be widely used. Their success results from their

tractability—being well understood and fairly straightforward to implement. These include regression-

based techniques, such as Fama-MacBeth and generalized least squares, and nonparametric techniques,

such as portfolio sorts. Our understanding of how to effectively apply and interpret these techniques has

matured.

Researchers continue to extend and innovate upon traditional computational approaches. For example,

Patton and Timmermann [2010] propose new ways to test for monotonicity (in portfolio sorts) in the

expected returns of securities sorted by characteristics that theory predicts should earn a systematic

premium. They provide a summary statistic for monotonicity, allowing researchers to decompose the results to

better diagnose the source of a rejection of (or failure to reject) the theory being tested.

New computational methods continue to emerge and flourish. An increasingly popular computational

field among quantitative researchers is statistical learning (sometimes referred to as machine learning).

These analytical tools—which can be classified into supervised and unsupervised methods—are valuable for

building models because they reveal the structure of data, incorporate nonlinearities into the model, and

provide robust predictions. In our view, these approaches should not be viewed or used as "black boxes"

but rather as analytical tools.

Researchers in finance are applying these newer methodologies to create insights into the dynamics of

10
equity markets. Moritz and Zimmermann [2014] address the research question of which variables provide

independent information about the cross-section of stock returns. Their computational approach, called

deep conditional portfolio sorts, is designed to deal with a large number of variables and potential
nonlinearities and interactions. When estimating the model, the authors incorporate concepts from the

statistical learning literature, mirroring methods used to estimate decision tree and ensemble methods.

Ogneva, Piotroski, and Zakolyukina [2015] use the lasso (least absolute shrinkage and selection

operator) model by Tibshirani [1996] to select a parsimonious set of fundamental variables for a
probability of recession given a failure model. Lasso estimates a sparse solution of a regression problem by

setting some of the regression coefficients to zero. They are primarily interested in the out-of-sample

classification ability of a proposed model, using cross-validation as a performance measure.

Unstructured data have become more valuable in developing quantitative signals for equities. Textual analysis

of corporate disclosures such as financial statements, earnings releases, and conference call transcripts are

sources of unstructured, qualitative data. Li [2010] provides a survey of various techniques to extract

signals from textual data, showing that the communication patterns of management could reveal

certain management characteristics that have an impact on understanding corporate decisions and

forecasting stock returns. Focusing on research related to earnings quality, stock market efficiency, and

corporate financial policies, he highlights two general approaches for conducting content analysis using a

rule-based dictionary approach and statistical approach, such as the naïve Bayesian machine-learning

algorithm.

Equity markets react to news flow. Although it is potentially rich in information, this source of data also

contains substantial noise. Heston and Sinha [2015] compare different methods of textual analysis using

news reports from Dow Jones to predict cross-sectional stock returns. They analyze the horizons over which

returns are realized for sentiment signals created using different computational methods. Some signals provide a

forecasting horizon of up to a quarter, whereas others forecast returns over shorter horizons, such as a day.

A growing amount of information comes from Twitter, Internet searches, and other sources of text-based

social media. These data sources can also be useful for building quantitative measures of investor

sentiment. For example, Da, Engelberg, and Gao [2015] use daily Internet search volume to construct a

measure of market-level sentiment. They show that this measure is useful in predicting short-term reversal,

increased volatility, and fund flows.

11
It is important to understand the intuition, assumptions, and strengths and weaknesses of computational

approaches. The choice of a method involves trade-offs. The computational approach should align with the

data structure, research design approach, and underlying investment strategy being research and traded. For

example, research that uses a hypothesis-based (deductive) approach to modeling typically relies on more

traditional computational approaches, such as regression and information coefficients. Pattern-based

approaches are applied to unstructured data sources and/or data that contain nonlinearities or other unusual

features.

EXAMINING SOME CRITICAL ASPECTS OF QUANTITATIVE STRATEGIES

There are many advantages to using a quantitative model. The scientific approach to developing

models bring advantages such as rigor, creativity, avoidance of biases, and process.

Rigorous analysis is an underlying principle of the scientific approach. A rigorous approach allows one to

validate ideas through a framework incorporating statistical rigor by employing backtests, in-sample/out-

sample comparisons, and Monte Carlo analysis to study the robustness and sensitivity of a strategy to a

given choice of parameters. This verification should also incorporate new market and theoretical

developments. New assumptions (paradigms/theories) require the reconstruction of prior assumptions and

the reevaluation of prior facts.

The numerous computational tools—from statistical learning to traditional statistical approaches—and

the expanding set of data sources provide tools and raw materials to explore new trading ideas. Similar to an

artist who has access to paints and canvases, a researcher benefits from creativity—which results from hard

work, introspection, and inspiration. This creativity is the driver for new investment ideas. With the right

tools, researchers have the ability to develop ideas about investing strategies and create models.

All decisions are subject to biases. In investing, the behavioral biases are well documented—confirmation bias,

optimism bias, and overconfidence, to name a few. Quantitative models give us more objective

benchmarks to measure our decisions and to (partially) eliminate these biases.

The decisions we make to build quantitative strategies are also affected by biases. We sometimes see

data and construct a story to explain what happened. Taleb [2007] calls this "narrative fallacy"— looking

backward and creating a story to fit events. Being aware of potential biases and understanding how our

assumptions drive our choices in the modeling process are key to building successful strategies.

12
Quantitative strategies are systematic; that is, the underlying strategy is consistently applied to identify and

implement trading opportunities in a structured framework. A framework brings structure and logic to a

disorderly and complex activity of identifying opportunities in the markets. This framework provides a

process—a common plan of direction and action to use in developing, evaluating, and implementing

investment ideas through quantitative models.

13
In markets filled with near-constant information flow, not all information influences asset prices. The benefit of

having a systematic model is having a consistent process to focus efforts on information that influences

prices—and to avoid reacting to noise (useless information).

Empirical research is often based on historical data, and there is a limit to how much information about

the future we can infer from the past. Sometimes, quantitative investors are at risk of being too systematic in

their approach. There is always the risk of low-probability events that will challenge the underlying

assumptions. Because markets change and the current environment looks different from the past, we need

to evaluate whether those changes are structural or transitory. It is important to understand how both types of

changes will impact the performance of quantitative models. We need to continually evaluate our models and

the markets they operate in, revising a model when our judgment and experience indicate it is no longer

effective or a different opportunity set for alpha arises.

• At the Core of a Quantitative Model: The Data

At the heart of a quantitative model is data. Quantitative analysis relies on nonexperimental inference. How

the data are used and the source of the data are of critical importance. "Garbage in and garbage out" is a

commonly used phrase referring to how the data inputs of a model can affect its output. For researchers of

quantitative strategies, this means that a quantitative process is only as good as its data.

Data impact the outcome of a research project. In any dataset, there are some data features we understand, and

some we do not. For researchers, it is critical to explore data features and expose unexpected features of

the data.

We can classify data in a number of different ways. More recently, it has become common to characterize data

as structured or unstructured. Structured data are organized into tables with clearly identified and

organized information. Unstructured data, such as text containing natural language, do not have a

formal structure. It requires specialized processes to extract the important attributes that can be used in

various computational techniques—and thereby, it introduces new opportunities and challenges for

researchers. The infrastructure to store and access this information is still evolving; thus, it requires

substantial effort to utilize this information in modeling.

Data containing errors, missing values, and other flaws affect the validity of the analysis. For example, Kothari,

Sabino, and Zach [2005] find that non-surviving firms tend to be either extremely bad or extremely good

14
performers. Survivor bias implies truncation of such extreme observations. The authors show that even a small

degree of such nonrandom truncation can have a strong impact on sample moments of stock returns.

Data are often available from multiple sources, and the number of available data sources is increasing. It is well

known that different data sources maintain a different level of detail. These differences can have a large impact on

strategy development. It is important to understand the comparative characteristics of databases. In particular,

the researcher needs to explain how discrepancies between databases affect the research output. The issue is

to determine whether the use of a particular data source might have influenced the results.

15
Characteristics of good data include the following:

• Clear and consistent definitions of what the data represent

• Reasonable detail underlying the data

• Appropriate data availability: length, frequency, timeliness

• Consistent view of history

• Free from survivorship bias

• Free from look-ahead bias

Successfully working with data means understanding the nuances of data sources. The following are a

few best practices for data.

• Understand how the database evolves over time. Most databases change over time, and those
changes include what data were collected, how the data were collected and its coverage.

• Understand how the database's standard procedures work and how they differ among different
data sources. Most databases have standardized procedures for reporting certain items in their
system in order to ensure comparability.

• Beware of potential biases in the data.

• Choose one data source to build the model and a second data source to confirm the model.

• Include statistics that describe and compare the usability of data items with regard to standard empirical
applications in finance.

• Look for economic explanations of any outliers. For example, Brown, Lajbcygier, and Li [2008]

examine the economic significance of outliers in their dataset. In their work, they show that the

outliers result from firms with materially different financial situations. In contrast to outliers caused

by bad data, this set of outliers had material implications for the conclusions of their results.

There may be opportunities in using less clean data sources or data with shorter history, and so on. This

type of data might provide a source of alpha that others overlook because of the work and patience

16
required to make the data usable for a research process.

How does poor data affect strategies? High-quality data are critical to success. The validity and power of the

results rely on well-prepared datasets. For example, Ljungavist, Malloy, and Marston [2009] document

changes in the collection and recording of historical I/B/E/S analyst stock recommendations. They show

that these changes are nonrandom, and the consequences of these changes affect returns generated on

trading signals using these data.

17
WHAT DO WE MEAN BY "GOOD" MODELS AND STRATEGIES?

In this section, we describe five key properties of "good" quantitative models. We leverage the work of

Gabaix and Laibson [2008], who describe critical properties for building economic models. These

properties are intended to be broad guidelines.

Parsimony: Parsimony means models with few assumptions. All models are only approximations
of reality, and some features will always be omitted. Our assumptions are based on the results of empirical

research (parameter estimates), economic intuition, and judgment and theory (for example, priors,

structural assumptions). Having too many assumptions tends to lead to overfitting. When overfitting

occurs, our models are unable to provide accurate out-of-sample predictions.

Tractability: Tractable models are easy to analyze, providing transparency to the user. We should have
explicit descriptions of our research choices for data, research design, and computational choices. Having

tractability enables us to question our model assumptions and make changes when necessary.

Conceptual insightfulness: Our model should align with market dynamics, investor behavior, and
investment theory. The empirical analysis and our hypothesis about why the strategy works should be

mutually reinforcing.

Predictability: We desire models that give us robust forecasts. As practitioners, we are primarily concerned with
the profitability and risk of a strategy and how well the model's motivation fits economic theory and market

behavior.

Adaptability: Financial markets are constantly evolving, subject to sudden, unpredictable changes. Market
changes challenge the assumptions of models. It is important to understand and forecast the potential impact

of those changes. For example, change may be the result of innovation of financial products, change in the

preferences of market participants, and/or responses to exogenous financial and economic shocks.

Flexibility in our thinking and our models is necessary to adapt to changing market conditions.

CONCLUSIONS

In this article, we discussed a number of principles, insights, and experiences for building successful

quantitative equity strategies. To build these successful strategies, we need to combine judgment with the

18
scientific approach to identify and validate new opportunities in ever-changing markets. Understanding

the data and computational techniques is necessary for obtaining empirical evidence to support our

investment ideas. Without doubt, the research and development of successful quantitative models are a

blend of science and art. Our endeavors in model building require constant innovation and rigorous

analytics to be successful.

© 2016 Pageant Media Ltd

19
[References]

1. Box G.E.P. “Science and Statistics,” Journal of the American Statistical Association,
71(1976), pp.791-799

2. Brown S., Lajbcygier P., Li B. “Going Negative: What to Do with Negative Book Equity

Stocks.” The Journal of Portfolio Management, Vol.35, No.1(2008),pp.95 -102

3. Cowen T. “Average is Over.” New York: Penguin Group, 2013

4. Da Z., Engelberg J., Gao P., “The Sum of All FEARS Investor Sentiment and Asset Prices.”
Review of Financial Studies, Vol.28, No.1(2015), pp.1-32

5. Caplin A., Schotter A., Gabaix X., Laibson D. “The Seven Properties of Good Models.” In
The Foundations of Positive and Normat ive Economics: A Handbook Caplin A.,Schotter

A., pp.292-299. New York, NY: Oxford University Press, 2008

6. Heston S, Sinha N., “News versus Sentiment: Predicting Stock Returns from News

Stories.” Robert H. Smith School Research Paper, 2015.

http://ssrn.com/abstract=2311310 .
7. Kothari S., Sabino J., Zach T. “Implications of Survival and Data Trimming for Tests of

Market Efficiency.” Journal of Accounting and Economics, Vol. 39, No.1(2005) pp.129 -
161

8. Leamer E. Specification Searchers: Ad Hoc Inference with Nonexperimental Data. New

York: John Wiley & Sons, 1978

9. Leamer E. “Let’s Take the Con Out of Econometrics.” Amer ican Economic Review, Vol.
73, No.1 (1983), pp.31 -43

10. Li F. “Textual Analysis of Corporate Disclosures: A Survey of the Literature.” Journal of

Accounting Literature, 29 (2010), pp. 143-165
11. Ljungqvist A., Malloy C., Marston F. “Rewriting History.” Journal of Finance, Vol 64,

No.1(2009), pp.1935-1960
12. Moritz B., Zimmermann T. “Deep Conditional Portfolio Sorts: The Relation between Past

and Future Stock Returns.” Harvard University Working Paper, 2014

13. Ogneva M., Piotroski J., Zakolyukina A.”When is Distress Risk Priced? Corporate Failure

and the Business Cycle” USC, Marshall School of Business Working Paper, 2015
14. Patton A., Timmermann A. “Monotonicity in Asset Returns: New Tests with Applications

20
 to the Term Structure, The CAPM and Portfolio Sorts.” Journal of Financial Economics,
98 (2010), pp. 605-625

15. Taleb N. “The Black Swan: The Impact of the Highly Improbable.” New York: Random
House, 2007.

16. Tetlock P., Gardner D. “Superforecasting: The Art and Science of Prediction.” New York:

Crown Publishers, 2015.

17. Tibshirani R. “Regression Shrinkage and Selection via the Lasso.” Journal of the Royal

Statistical Society, Series B58, No. 1 (1996), pp. 267 -288

18. Tversky A., Kahneman D. “Judgment under Uncertainty: Heurist ics and Biases.” Science,

185 (1974), pp. 1124-1131

21