Model thinking and strategic thinking: Why are the models that are always useless?

Have you encountered such a situation: the model you made is always useless, and it is estimated that there is no result? This article takes the learning ability of primary school students as an example to analyze how to use model thinking and strategic thinking, and hope to help you.

Imagine that you are a primary school principal, you must score the learning ability of all elementary school students, or to give everyone a label for good students and poor students. How would you fight?

Take out the sixth grade students because they have a high degree of reliability in their exams. All the good grades in all exams or recent exams are particularly high. They are defined as good students and are particularly low as poor students. Then use all the information you can get, such as family background, hobbies, etc., to train a classification model, and finally predict all students of one to sixth grade.

You are sure and you will definitely not do this. If you leave this task to the class teachers of each class, you will find that they must have adopted another approach.

Sixth grade people have tested a lot of tests, of course, it is very good. According to the historical results, especially the recent results, just play directly. The first grade only tested several times, and it was also necessary to judge according to the scores such as Chinese mathematics, art, music and sports morality. This means that the strongest service is the strongest. The confidence of the low grade is definitely not so high, but it is better than using what you like to eat, playing without playing, and not watching TV factors such as TV factors. All students' marking rely on the scores and some characteristics of the characteristics to judge the most effective.

The first method is model thinking. In this scene, you will not rely on model thinking at all, and you even think that the model of the model is ridiculous. But, changing the scene, aren't you doing such ridiculous things?

The second method is strategic thinking. You made this scoring strategy based on the most important information.

When you use model thinking to solve the problem, have you found that you can solve any problems. You can use the things you caught to predict whether a person is rich, GAY, whether you have the potential for derailment, whether there is a tendency to violence, whether it is very rolled, do you want to change jobs, is it yellow gambling poison? , Do you like to say that q Q, will you help the old grandmother cross the road, whether the toothpaste is squeezed from the middle or the two ends squeezed, whether the bowl is clockwise or counterclockwise ...

We are here to solve the problem. It is not so easy to solve the problem. How can we do a black ax? You have to know the second ax, three ax ...

When you use strategic thinking, you will think of various ways to solve the problem. For example, using a problem to replace the original problem is a good strategy to solve the problem. In other words, if you can't solve a problem, you can solve another simple problem. Go find this simple problem.

Still use risk control as an example. Of course, credit risk can be directly predicted, or the problem can be disassembled. Low -income people have high risk of people, high -liability people with high risk, etc. Therefore, solving these two sub -problems can be beneficial to the ultimate problem.

The problem of scoring primary school learning ability above, you think the model thinking is ridiculous. But in fact, the same is true for income and liabilities. It is strange to predict, and of course it is not that it must not be.

As for why risks can predict, income and liabilities cannot be predicted? I think that in a prediction with risk as the evaluation goal, because the forecast and the true value are always errors, there is an error in the prediction of income and liabilities, and the relationship between income liabilities and risk This error is to be overlapped. The prediction results are inevitable.

The more suitable approach of income and liability modules is to define not to predict. Use all the relevant data you can get, according to the degree of reliability, try your best to align, and write the case when.

To sum up, the difference between the model thinking and strategic thinking mentioned above is how to define this Y. The model thinking is to define a small sample y casually, and then find X to predict; the strategic thinking is to do everything possible to define the full Y as much as possible, and then there is no more.

The former is easy or the latter is easy, friends? Don't bully yourself. The former is much more likely to be, and there is no difficulty in the aspect of a model prediction. Ask yourself, will you only be the former and not the latter?

To make models, we all like end -to -end solutions, because end -to -end solutions are the simplest. But sometimes it does not work.

You don't know what you do at all. You said you want to test, can you test it anyway? You think it is always useful, but the fact is useless to do casually. A good strategy should be simple and logical. The model made by model thinking cannot be done. You can only believe in data and only test. Then test is useless.

Another problem is that the end -to -end solution always needs to be based on historical data. If your history is done well, if it is not possible, what is the meaning of learning history to develop the future?

Moreover, historical data is often conditional. Remove this condition, your model is likely to have big problems. For example of wealth management, historical users have high -stick users with 100W+positions. Now they want to find more high net worth high dive customers, so you complete the modeling and remove the conditions of 100W+in history. Find out a batch of batch The so -called "high dive" user with high scores. Is this score good in historical data? Is it necessary to be effective?

If you see a person reading "Reference News" in the subway, which situation is more in line with the reader? He has a PhD, and he has no college diploma. You may choose a doctorate degree, but this is not necessarily wise. You should fully consider the second option, because there are more people in the subway.

If a person is marked as a poetry enthusiast and guess whether she learns ancient Chinese literature or business management, you should also choose the second answer. Because although female students studying ancient Chinese literature are more likely to be poetry lovers, they can almost have more female students who have more business management and poetry.

"Zhang San is a 30 -year -old man, a master's degree, married, but no children. He works in a large factory, has strong ability, motivation, and is very popular with colleagues." Such people show high conversion rate in historical data. , But aside from the 100W+position, the proportion of such people's high net worth can be low to scary.

Regardless of the problem, no matter how the model is a model, I think it is a tactical diligence to cover up strategic laziness. This model is not only effective but not worth. I have a way to judge the value of work. If experience is not important, then work is worthless.

Finally remind you. Don't think that inserting a thermometer in the soil can measure the temperature of the entire earth.

Columnist

Lei Shuai, WeChat public account: Lei Shuai is fast and slow, everyone is a product manager columnist. Risk control algorithm engineers, understand risk control, some business, and life. I always believe that experience makes work easier, and then find that risk control makes life more free.

This article was originally published in everyone's product manager. Reprinting is prohibited without permission.

The question map is from UNSPLASH, based on the CC0 protocol.

The point of view of this article only represents the author himself, and everyone is the product manager platform that only provides information storage space services.

- END -