Forecast model, how to do it is accurate enough

In data analysis, an ultimate problem that will encounter is “predicting not allowed!” So how to solve this problem? The author shared the relevant method, let the predictive model make more precise, let’s take a look.

“It is not predicted!” Is the ultimate problem in the field of data analysis. The predicted algorithm has a lot of pile, and then encounters reality basically being hammered into slag, the business parties are not satisfied.

What should I break? Today system tells.

First, the essence of the prediction algorithm

From essentially, the prediction algorithm is only 2 categories:

1. Based on time series

Smooth: It is used to relatively smooth data. Self-return: Used for trending increments, decreasing data. Self-regression with seasonal factors: Data for cyclical fluctuations.

2. Based on causal relationship

Second classification question: XX will not occur in the future, typical, such as LR. Multi-classification: In the future, which is ABC, typical as decision trees. Continuous problem: how much is the value of the future, typically linear regression.

When it is possible to model, it is not a model package to play the world, but use two phases modeling. For example, predicting a passenger group consumption, you can use the second classification model to predict the consumption amount, then predict the consumption amount, so that the number of consumption households * predict the amount of consumption, you can get the total consumption. This is a typical processing technique.

The book is taught on the book, but it is hammered into a slag for a reality.

Second, the difficulty of predicting algorithms

Because: Book in order to highlight the model effect, deliberately select the quality and data set of data.

The troubles in reality are all constant:

No data. Many times give to the data to be predicted, in order to “total monthly consumption”, other data fart is not … still no data. Many companies can’t take one-on-hand data over Tmall, shake, Amazon, can only use a little bit of data exported in the background … There is no data. Most companies are not the monopoly company of Scott, only very little data. The most common, most companies users are spending money, only one mobile phone number + a discount order …

This leads to a funny situation: many companies use causal relationship models, which have the greatest variables, must be promotional. Even with gradual regression method, the variable of the promotion can be done directly to the other variables directly. The prediction result is turned: the larger the promotion, the more users join, the more purchases.

This result is throwing out, and a quasi-author evaluation is: “TM is nonsense, I know it!”

This is the second big trouble in reality: how the business effect is measured.

For example, the prediction sales is 10 million L business to achieve 9 million, which will say: It is not allowed to predict, and it is necessary to pay back.

The business is 11 million, it will say: It is not allowed to predict, or I am

In short, as long as you are not 100% accurate, he has a reason to go to your head. I can even jump repeatedly. For example: “Original business can meet the standard, see the forecast to meet the standard, we will put in the price, the result is not reached, all blame the interference, business judgment …”

How to break? The problem is born by people, of course, I have to solve it here. Avoiding a prediction of a gambling, from the perspective of business scenarios, and eliminating people is the key to the key.

Third, use business operation to avoid prediction errors

Some scenes can directly eliminate the problem directly through business operations. At this time, you will use your business means and don’t model it.

for example:

Scene 1: “Sales data is small, distributed is very scattered, how to predict sales? Group purchase is to solve this problem.

Scene 2: “Sales data is very small, less to calculate price elasticity, business side want to predict price elasticity, make more money” – use auction! The auction is doing this.

Scene 3: “New product is a new model, no data, what is forecast?” – Do new pre-sale / fan with code purchase. Hunger marketing is doing this.

Scene 4: “How do you predict during the promotion? Is it not allowed to have users?” – 10 yuan deposit, deposit swelling 3 times voucher, just doing this.

Almost all Internet marketing models, from Xiaomi to Tmall to spend a lot, actually in confronting the stockpicies brought about by insufficient data. So don’t look at the model of people, and the operation of people is also learning.

Fourth, use basic analysis to narrow the prediction range

All gambling predictions have a common point: must not be high. For example, typical predictive sales performance, if it is actually 10 million, he is not required to predict 1 million talents. This is the root cause of the model being evaluated as “not allowed”.

Back to business scenarios, most business scenarios don’t need this level of accuracy. Most of the time, the business is afraid of a scene that suddenly incremented / plummeted. The prediction target is set to: “100% precision”, it is better to set as: “Whether it has incremented / plugging over business digestive ability.”

It is easy to predict 100% accuracy, but it is easy to get rising / plunge. Through basic analysis, the unstable factors are distinguished, and the difficulty of predicting problems can be greatly reduced (as shown below).

After making basic analysis, after splitting instability, it is also more convenient to select the model combination to solve the problem (as shown below).

V. Use rolling prediction instead of long-term prediction of all gambling forecasts, the prediction time period is very long. For a long time, there is a short day. When the gambling time is too long, there are few data that can be collected in the previous period, and it is not possible to reflect various intermediate operations in the business unit, so very passive.

Use scrolling prediction to greatly make up this shortcomings. Through day / week scroll prediction, it can supplement data deletion, and reflect the effects of temporary adjustment of the business parties, two (as shown below).

Sixth, users buy to protect themselves from hand

A good problem + rolling prediction, basically meet the actual work needs. But as a prediction, you have to learn to protect yourself, avoid the business side to repeated horizontal strips, jealousy pot.

It is a good way to buy a way. After the prediction results are given, they are sold to the hand, and all relevant business are no longer questioned the forecast results, but based on the prediction results.

Who felt less forecast, who wrote to apply for additional goods and leave a written evidence. It is predicted that it is not allowed, or it is still a lot of applications, so I can’t do it, I can’t see it, I can see it clear (as shown below).

Seven, deep prediction problem

Behind the prediction problem is a very deep business problem: in many companies, the loss of inventory backlog is intuitive and visible, the goods are rotten in the warehouse. However, the potential sales of out-of-stock losses did not care seriously.

It is easy to count the loss of outsider, allowing customers to make a reservation, allowing customer registration requirements and time. But many companies are either lazy or stupid, or do not want to be responsible, in short, there is no.

Pre-sale, group purchase, hunger marketing, deposit expansion, out of stock registration … All business means, not only operate, but also need system construction support.

Obviously, for the business, compare these system construction and complex operational means, or directly to the “model prediction is not allowed” easier. Therefore, if you ask your business forecasting needs, they will tend to predict the “not high”.

But it is clear that this analyst is unfair. Since potential loss is unbearable, reality backlog is intuitive, so as a data analyst can be practical. So there are all kinds of operational methods.

Writing this, there must be many students want to see specific operating cases, if I am interested, pay attention to the public number of the grounding gas collection, the next article Let’s share a case, please look forward to.