Why machine learning is not good in the investment field _ Asset Management

Why machine learning is not good in the investment field

Original 2017-04-05 Ishikawa Volume letter Investment

Http://mp.weixin.qq.com/s/RgkShbGBAaXoSDBpssf76A

“

The essence of data snooping is this focusing on interesting events are quite different from trying to figure out which Eve NTS are interesting.

Attention to interesting events and figuring out which events are interesting are two different things, which is the nature of data accommodation.
”

1

Signature

Recently, one piece of news blew up the investment circle: BlackRock, the world's largest investment management company, announced that it would use machines (BlackRock AI Artificial Intelligence or machine learning algorithms machine learning algorithm) to replace some fund managers for stock selection. In recent years, with its application and outstanding performance in face recognition, credit anti-fraud and even chess and Weiqi, artificial intelligence is more and more familiar to people. A lot of people are starting to see that in the near future machine learning algorithms will achieve better results than people in the two-tier market. And BlackRock's announcement undoubtedly put artificial intelligence on the cusp again. One of the most fundamental points of this is:

Machine learning through a variety of non-linear algorithms (such as neural networks, decision trees, genetic algorithms) can be used in a large number of historical transactions from the data to dig out the investment patterns that humans can not see. The choice of stocks on the basis of these models will yield substantial gains.

Although I am firmly optimistic about the field of quantitative investment, I am conservative and cautious about the idea that machine learning can replace human beings in stock selection. This is because financial analysis is a non experimental science (nonexperimental) and therefore cannot be controlled (scientific control or controlled experiments). This means that while there is a large amount of data on financial transactions, it is not possible to design experiments to control the variability of independent variables and to test the assumptions made by repetitive tests (for example, some type of stock-picking pattern found in machine learning). Such data analysis is mostly a seemingly significant but actually deceptive pattern (especially for sample data), a phenomenon called data snooping.

Data snooping: Mining the nonexistent schema from the data (finding patterns in the, do not exist).

The problem of data accommodation exists in all non experimental studies, especially when we use complex machine learning algorithms for stock selection. This is because the complex nonlinear algorithm contains a large number of parameters, through the combination of these parameters can always find some people do not understand the stock model to obtain excess income. If these patterns are not properly understood and interpreted from a business perspective, data accommodation will make complex machine learning algorithms an efficient tool for discovering ineffective coincidences from historical data, as the reference at the beginning of this article says.

2

Using pseudo prime numbers to select stocks

Look at a stock Bishi algorithm. The traditional fund manager may not have racked his brains to think of such a pattern, but machine learning algorithms can easily (but wrongly) find it. This algorithm utilizes a property of prime numbers (prime numbers) to pay for a variant of the Ma Xiao theorem: In addition to 2, any one prime x satisfies "2 of the x-1 of the second party by its own remainder is 1".

For example, 13 is a prime number, and 2 of 13-1 (or 12) of the second is equal to 4096. Divide it by 13 to get 315, and the remainder to 1. It can be proved that all the primes other than 2 satisfy this property. But the numbers that satisfy this property are not necessarily primes, they are called pseudo primes (also known as Carmichael numbers). There are seven pseudo primes within 10,000:561,1105,1729,2465,2821,6601, and 8911. We use these pseudo primes to stock stocks: Select the stocks in the stock number that contain the above pseudo primes to invest. According to this rule, Ametek (a manufacturing company, Stock number 03110510) stands out. More surprisingly, it has gained 95 times times its cumulative earnings over the past 40 years, far exceeding the Dow Jones industry or the S & P 500 index.

There is no doubt that this is an extraordinary stock, and our pseudo prime strategy has achieved great success. However, do not rush to get excited first. We need to take a good look at the relationship between pseudo prime numbers and stock picking. The answer is no matter. So does this strategy really find an effective stock-picking pattern? The answer is in the negative.

Some people will jump out and say, "As long as it works, it doesn't matter why it works." ”。 This kind of cognition is very dangerous. For the non experimental problem of stock selection, because it is impossible to test the hypothesis by the control experiment, it is very important to understand why the algorithm of machine learning is effective at least from the business. Therefore, "as long as it works" is a very irresponsible attitude.

This example represents the problem of a lot of machine learning algorithms: we can always use complex non-linear algorithms (such as neural networks) to find the invincible stock-picking patterns in the back-test by using the excessive optimization parameters. In the process, we have fallen into the trap of data accommodation.

3

Cognitive biases exacerbate data accommodation

Data accommodation issues can easily occur under these conditions, and they are clearly present in the level two market:

There is a lot of data.

Many people are using the same data for analysis.

Lack of business theory or the inability to control variables.

Cognitive biases are "as long as it works, why it doesn't matter."

The first three of these are the objective conditions of the market, and the last one is rooted in people's cognitive errors. Human cognition always tends to pursue unusual events. It is only when some "unusual" coincidences occur that we tend to be concerned. Swiss psychologist Carl Jung called the synchronicity of coincidence to be a synchronic one.

Synchronic: Refers to "meaningful coincidence", used to explain phenomena that cause and effect laws cannot explain, such as dreams come true, think of someone, and so on ("Speak of the Devil, The Devil"). Jung believes that these seemingly causal events have a causal and meaningful relationship, which often depends on the subjective experience of the person. When both occur simultaneously, it is called the phenomenon of synchronic.

In layman's parlance, when two things are not connected in time and space, people think that there is a supernatural mystical force linking them together and that the coincidence has some meaning.

For example, in the above example, the stock code contains pseudo prime numbers and stocks to obtain a huge excess returns is a pure coincidence, such a coincidence by machine learning algorithm found and presented to the user. If the user does not attempt to understand whether the two really have a relationship, it will be because of the synchronic and the wrong coincidence to give some meaning, that machine learning has found a good way to choose the stock model.

4

Luck or strength.

So much has been said before, and the aim is certainly not to negate the application of AI and machine learning in the two-tier market. But I would like to say that for any pattern of artificial intelligence discovery, it works only if we can understand the meaning of it clearly and unambiguously. It is impossible to tell whether a good result comes from luck or strength.

Before, I wrote an article, "Better luck than good." 》。 The use of sequential statistics (order statistic) In this article explains this truth:

Among the many stocks, the best will always have a very good yield; Among the many strategies, the most powerful one always brings an amazing rate of return. However, by calculating the distribution of extreme values (sequential statistics) of independent samples, this result is inevitable.

Let's review the examples in that article. Suppose a stock investment strategy's annual yield X conforms to the mean value of 10%, the standard deviation is normal distribution of 20%. Assuming that there are m different policies in the market, the yield y of the best of them is the function of X, y = max (X1, X2, ..., Xm). The following figure is the comparison between the yield distribution of the best one and the distribution of a single policy yield when m = 3000: The yield distribution of the optimal strategy moves to the right and becomes narrower on the horizontal axis.

The following figure shows the results of the Prob (y≥0.7) variance with the number m of the policy. The mean value of Y and the variation of the standard deviation with M are also given. As M grows, we are more and more determined that there will always be strategies to stand out, with an annual yield of more than 70%. This kind of judgment can also be proved by the mean and variance of Y: With the increase of the number of strategies, the mean value of the annual yield of optimal strategy is increasing, and the standard deviation is decreasing.

The result shows that when there are a lot of different strategies, the best one is always exceptional. But the question we really care about is whether this strategy has found a false pattern in the vast historical data, or a true scientific investment model. We have to figure out how it works from a business level.

“

As with no black box, if you don ' t know why it works, your won ' t realize when it ' s stopped working. Even a broken watch is right twice a day.

Machine learning algorithms are like a black box, and if you don't know why it works, you don't know when it's going to expire. Even a lockout watch can be correctly two times a day.
”

5

Artificial intelligence is a long way off

In fact, people use algorithms to select stocks is nothing new. The risk multiple factor model can be regarded as an algorithm for selecting stock. Of course, it works because of its use of factors, such as growth factors, scale factors, momentum factors, have a clear business base. In recent years, many people use machine learning complex algorithms, such as support vector machines, to improve the multiple-factor stock selection. These nonlinear algorithms construct a lot of non-linear factors. For example, if the algorithm tells us "the concept plate of the male, and the logarithmic market value of the three-month momentum of the e-time is greater than PI" is a good model, then we have to ponder over.

As for the application of artificial intelligence in the level two market investment, a quantitative investment predecessor who has rich actual combat experience has expounded the following views, which I strongly endorse:

“

We can believe that it can capture subtle patterns that humans simply cannot perceive. But can these patterns last? These patterns will be just random noises that don't repeat. Experts in the field of artificial intelligence have assured us that they have a lot of precautions to filter for instant noises. And these tools do have a significant effect on consumer marketing and credit card fraud detection. The pattern of consumer behavior and fraud obviously has a long duration, which makes these artificial intelligence algorithms work even if they contain a large number of parameters. In my experience, however, such precautions are far from sufficient to predict financial markets, and the excessive fitting of historical data noise can have serious consequences. ...... Compared to the large number of independent consumer behavior and credit transaction data available, we have access to statistically independent financial data that is very limited. You might say that we have a lot of time-sharing financial data to use. In practice, however, these data are sequence-related and not independent of each other.
”

The predecessor gave his opinion about when artificial intelligence was effective:

Based on the correct econometric or theoretical basis, rather than the randomly discovered pattern.

The required parameters are less than the historical data.

Only linear regression is used, and no complex nonlinear function is used.

The concept is simple.

All optimizations must be implemented in mobile windows that do not contain future unknown data, and the effect of this optimization must be continually confirmed by future unknown data.

The more rules the policy has, the more parameters the model has, the more likely it is that data will be indulged. It is often a simple model that can stand the test of time.

6

And look at BlackRock's decision.

As the world's largest asset management company, BlackRock announces the use of artificial intelligence instead of fund managers is not to be ignored, and will inevitably be a stone to stir the waves. The agency predicts that by 2025, 10% of the world's financial institutions will be replaced by machines. This is probably not a relationship with the increasingly high alpha. After all, in the long run, the vast majority of fund managers do not win the index, so what is the use of these fund managers?

By quoting my partner, boss Gao, it might be better to understand the BlackRock decision:

“

Excess income more and more expensive, open source is not good, try to throttle. The equilibrium state of the final investment market is that the marginal cost of excess income is exactly equal to the excess income. Such costly investment funds will eventually be squeezed out of the market by lower-cost funds.
”