Source: Minority Investors Author: Wang Yudong
Original title: 5 Thoughts on Quantitative Research

In recent years, on the basis of active research, we have carried out some quantitative explorations, and our understanding of relevant methods has gradually deepened. In particular, the basic principles behind specific strategies have directly determined the trade-offs of subsequent research.

This article will share 5 personal thoughts on quantitative research in the form of questions and answers:

1. What can the statistics show?

“From rolling the dice to the alpha dog” has a story:

Jagger is an excellent mechanical engineer. He is attracted to the casino turntable: there are 38 numbers on the turntable. Ideally, the probability of each number appearing is 1/38, but under the process at the time, the machine was unable to achieve perfect symmetry, and the presence of defects would cause the turntable to skew certain numbers.

He hired 6 assistants for several days to record every number that was transferred out of each roulette machine and analyze the pattern. He found that on the sixth roulette wheel, the nine numbers appeared significantly more frequently. As a result, he played this biased turntable machine, bet heavily on these 9 numbers, and earned 70,000 yuan on the same day.

Quantitative backtesting can be compared to turning a turntable:Count the historical turnaround results of different strategies to find “bias”, so you can bet on investment directions that are likely to be profitable.

Jagger's story isn't over yet. The casino's management discovered an anomaly and changed the mechanical settings of the sixth wheel. He was unable to continue making money; in the end, he had to leave the casino.

In the stock market, price formation is the result of a game where participants each pursue their own interests to maximize their own interests.It is clearly different from an objective turntable with specific mechanical structural defects and no changes:

The “bias” of the stock market refers to the interactive characteristics of participants' behavior (not objective physical), and people from all parties will adjust their behavior based on history and current changes, and even consider one more step based on expectations about the behavior of counterparties; furthermore, many “game points” are actually “adjustable” for specific groups.

As a result, it is impossible to simply assume that the regularity in the backtest is the result of deviations in the participants' behavior.Static bets based on historical indicators are likely to be hunted down by dynamic players who think more slowly. So, how do you determine whether the “specific turntable” in the stock market has been changed by anyone?

There are two types of methods:

The first is data-driven high-frequency trading: faster feedback, faster adjustments, as long as you leave before losses eat away profits, don't worry too much about what you did right before and the reasons behind it (even if objectively, there must be a reason);

Second, it is logically-oriented low-frequency portrayal: it is necessary to understand the microstructure behind the data, especially the specific rules and the expression of interests of relevant parties, pay attention to rule changes, follow the process of interpretation, and adjust one's own participation posture.

This article only focuses on the second type of method. The underlying logic is:

Statistical results can only tell us what happened in the past, but the surface relationships of historical data are not naturally applicable to the future; what really works is the “microstructure” behind the data.

In other words, what can be played over and over again is the specific microstructure and its causal relationships under specific rules, not a mysterious number game. If the rules and microstructure of the game change, even the most prominent rules in history are no longer applicable today.

2. Are the explanations we've given really reasonable?

As a result,A historical statistical result with abnormal benefits can only be the starting point for research.

We need to explain the observed phenomena, and make further inferences, tests, and even find clear microstructures and corresponding rules and environments. The 4-step method of phenomenon - explanation - reasoning - verification is the only way to go for all research.

In reality, many people have this misunderstanding: for a market vision, as long as a “reasonable” economic explanation is given, then this rule is repeatable.

Just like, Aristotle discovered that stones land faster than feathers, and gave the explanation that “heavy objects land faster than light objects”. The crux of the question is, if this explanation holds true, then its inference must also stand up to data verification: would a 10-kilogram feather land faster than a 5-kilogram iron ball? Obviously, Aristotle's “reasonable” explanation was wrong, but the theory dominated Europe for almost 2000 years.

Don't underestimate people's ability to make up for themselvesDon't be satisfied with the reasons most people have already mentioned or the explanations given in some literature. If these explanations hold true, what are the inferences? Is this contrary to objective data?Are there any other explanations that are more reasonable and that the inference has been verified by the data?

Be especially wary of “panacea” explanations:For example, as long as a strategy has excess profits within a certain period of time, is it “excessive or insufficient market reaction”? So, what should a reasonable market be like? As another example, when a strategy or factor fails, is it because “trading is crowded”?

This kind of explanation can only fill an understanding gap where there is an explanation but no clear explanation has actually been found; it lacks the necessary microstructure, and there is no clear reasoning.It's just revolving around existing data, and it's not possible to push the research furtherThey may even think that they have a reasonable explanation and are blindly confident and take excessive risks.

No explanation is absolutely reasonable, but the proposal of a new reasonable explanation, while being able to understand more visions, from a practical point of view, should be able to open up new dimensions, incorporate incremental information, and promote the deepening of research. In theory, every explanation is “subject to falsification”. Valuable research is a process of continuously finding new explanations and continuing to test and think.The key to explanation is to point to clear practices.

In stock research, what kind of explanation would be relatively reasonable and more logical? Judging from existing experience, it directly involves the interests of relevant parties, especially the regularity brought about by maximizing the interests of the superior party, and the inevitable choices brought about by maximizing the interests of the dominant party, which has more value that can be grasped.

3. Has the concept been stolen?

Quantitative methods have made many strategies testable to a certain extent. However, in the process of quantification, it is easy to have the problem of “stealing concepts”.

In other words,Quantification for the sake of quantification ignores the problem to be solved; even the object of the test is actually another stock selection method that has little to do with it.

For example, value factors, is buying undervalued (such as PE) a value investment? Coupled with the quality factor of high ROE (high profitability), is it a value investment?

Build a backtest along this line of thought,What you test is only the indicator itself being tested, not the problem to be solved in the first place. Excessive simplification makes quantitative verification meaningless.

When it comes to value investing, according to Graham's definition, it's bought at a price lower than the intrinsic value, and “intrinsic value” means “value proven to be reasonable.”

For example, if the market value of a company falls below net working capital, it is profitable to buy a company to pay off its debts or go bankrupt and liquidate it. This isValue investing from an “arbitrage” perspective; As another example, if you can clearly see that a high-growth company can reliably reach a certain level of profit in the future, then even if PE is currently high, the intrinsic value may be higher than the market price. This isValue investing from a “growth” perspective...

The quantitative strategy based on indicators such as PE and PB only emphasizes the level of this indicator itself, but it does not portray the essence of the “value investment” expressed above:First, we need to determine what is reasonable valueThe market price should be “relatively” reasonably low, not “absolutely” low. (Bypassing the key question and stealing the concept)

Taking bank stocks as an example, the few with the lowest PB in the past few years should instead be avoided. The reason is the deeper “microstructure”: their history is very burdensome, and regulation will not allow them to clear risks all at once. Their performance will be weaker than their peers for a long time, causing stock prices to continue to be under pressure. Buying based on low PB is likely to be at odds with value investing.

These kinds of questions frequently arise in a number of quantitative studies:The “gimmick” is huge, but in essence, no problem has been solved. It just emphasizes that a combination of several unrelated variables in history would have a good backtest effect.

A clear portrayal of the object is the starting point for quantitative research. What are we portraying? Has there been a change of concept? Did we deviate from the original proposition to be tested? What did we test in the end? Which “points can be grasped”? This is worth reflecting on.

4. Have you been kidnapped by specific metrics?

Many people may have this view: historical backtesting has tested the effectiveness of specific indicators.

In fact, specific indicators are only proxy variables corresponding to the prefactors of principle.On the surface, historical data verifies the indicators; in fact, it tests the underlying principle (microstructure).

The underlying logic is that under the same principle, if one indicator is portrayed by a different indicator, similar results should be obtained; if, under the same principle, the indicators are only slightly adjusted or depicted from a different angle, then the previous backtest results are meaningless. (The inference of the principle has been falsified)

The so-called “kidnapped by indicators” means“Mysterize” some indicators that have historically had excess earningsIt's as if it had some kind of mysterious “stock selection ability,” and even a slight deviation would no longer have “magic power.” Generating this sense of mystery is a weakness of human nature.

Instead of getting too bogged down in the details of an indicator, we should go back and see the principles it points to, and even the common points or clear falsification of different portrayals of the same factors.

Historical stock prices are highly serendipitous. Theoretically, the results of different proxy indicators based on the same principle are the same. Even if you arrange the combinations over and over again to come up with a better backtest result,It has not brought any benefit to the principle itself, let alone any reason to obtain better actual benefits in the future.

What's more, many indicators, in essence, are not necessarily “factors,” but can only be considered a classification standard.

Imagine this scenario: At some point in the future, we will go back and reflect on our current stock selection gains and losses, and sum up our lessons. Are some stocks performing well or not because their market capitalization is small, the turnover rate is low, or the rise and fall have been large in recent days?Is there a “reliable point” in advance to explain the origin of a phenomenon using a phenomenon in this way?

When we invest logically based on specific principles,There must be a clear “game point” change that will occur in the futureIt may be better performance, it may have caught on a certain hot topic or topic, it may even be a substantial restructuring, etc.

On the other hand, looking at these indicators (many people call them “factors”), could this be the reason for the changes in “game points” described above? Can it help us make better judgments about changes in “game points” beforehand? Or is it just because of this kind of stacking that would yield better backtesting results?

When we're faced with an indicator and think about whether it should be incorporated into an existing model, let's first ask: what does it portray? Does it have anything to do with the principles of the model itself? Rejecting indicator kidnapping and adhering to the proven principle itself is a realistic attitude.

5. What incremental dimensions can improve “comparative advantage”?

Whether it's active research or quantitative models, the ultimate goal is to find a point that has a relative advantage and consolidate and expand it through continuous accumulation.This is an inevitable requirement of a market game.

The so-called “comparative advantage” includes familiarity with investment instruments and market rules, awareness of game pricing characteristics, understanding of the interests and behavior of all parties involved in specific events, and even insight into industry-level game processes, etc., and is even more direct — a wider network of contacts, more and more timely information, and the ability to process massive amounts of data.

Einstein said,“You can't solve this problem at the same level of thought that created it.”

In stock investment, the focus of the market game is the rise and fall of stock prices. The most direct information is price data generated by transactions. Price changes have no momentum or reversal, but the reason for the continuation or change of trend is not necessarily at the level of volume and price.

For example, in the turntable machine mentioned at the beginning of this article, the nine numbers on the sixth turntable appear more frequently because of a specific microstructure: the mechanical settings are imperfect and biased, and momentum itself does not cause momentum. At the same time, the current machine has not been readjusted, which is an important prerequisite for betting profit.

In addition to stock prices, we can introduce data on the company's performance: if net profit growth continues to accelerate in the future, it will generally lead to “momentum” in stock prices; however, if performance changes or an inflection point occurs, it may lead to a “reversal” in stock prices. In this way, the quantified portrayal extends to the fundamental dimension of momentum.

If there is still no certainty about the company's future business situation, can we break out of the strange circle of linear extrapolation and find points that influence the pace of profit release from a higher dimension, such as the company's financing behavior? As a result, incremental dimensions of information are constantly being introduced...

In addition to this, we can also narrow down the scope of the study: for example, looking at the entire market, it is difficult to make a clear judgment, but will there be a clear boom or reversal in a specific industry? Also, are there current opportunities for specific companies at the intersection of a few leads?You can't demand a certain dimension to have a conclusion, but you should try to find a foothold in each dimension.

Stock research does not use various models to repeatedly torture known data, but rather revolves around the problem itself to be solved, disassembling it layer by layer: from stock prices to performance, to the behavior of listed companies... and even narrowing the encirclement circle, focusing on intersections, etc.In the “iteration”, investigation is carried out one by one based on the list, in order to find “grasiveness” and obtain a comparative advantage in a certain dimension and at certain points.Even in exploration, I discovered an unprecedented way of thinking and further promoted the improvement of the information dimension.

Computer programs can optimize known dimensions and given information to the greatest extent possible under a given data set. For example, the momentum at the performance level was found from stock price momentum, and the best participation posture in history was given. But what if there are no “points to seize” at these two levels? What if the most critical causal relationships aren't in the data currently available? What if the relevant “manageable points” only cover a small category of stocks, but the criteria for the relevant classification are not in the given information? ...

No matter how amazing the computing power is, it is difficult to exceed a given range of data.

At this point, further research is not about repeatedly torturing the data itself, but rather introducing what incremental dimensions, incremental information, and even more clear pertinence, more detailed classification, and related standards.

At this point, human intervention is necessary. Subjective exploration will bring richer perspectives. Subjective dimension reduction and dimension enhancement is an important driving force for further quantitative research.Machines can't do things outside of established boundaries; this can only be made up for by humans.

At the end of the article, let's borrow a famous quote from Mr. Hu Shi:

“I'm afraid that the truth will be endless; every inch of joy comes in.”

There is no research method that can be done once and for all. Any quantitative model is in the process of continuous optimization and improvement, and perceptions of market visions are also twistling in the midst of denial and denial.

We must not be limited to the arrangement and combination of established factors and repeated torture; we must break through known boundaries, face up to the problem to be solved, find “points that can be grasped” in the complementarity of quantification and initiative, and thus obtain a “comparative advantage” on specific issues.

edit/lambor