What is the correct way to open "data analysis"?

Last week I talked with you about data construction, so a reader asked what approach should be used to do good data analysis , so let’s talk about a 10-dollar approach.

1. Basic principles

When doing data analysis, you must follow certain principles. I have briefly listed these principles, which are the ones that Brother Liang values ​​the most, but they are not necessarily universal.

Let’s first look at the data analysis process.

Liang Ge believes that data analysis should follow the following process:

In other words, data analysis needs to go back and forth from the problem. In fact, data analysis mainly consists of five steps:

Ask questions, find objects, select data, do analysis, and then verify.

1. Ask questions: Data analysis must have a clear purpose

When starting data analysis, there must be a very clear purpose.

This purpose may be accompanied by one or more targeted testing actions.

For example, the purchase conversion rate of the original product landing page is relatively low, and a new landing page is needed to increase the purchase conversion rate after traffic enters.

There are three questions to ask at this time:

1. Is the current product landing page worth optimizing?
2. Is it possible to optimize product landing pages?
3. To what extent can it be optimized?

You will find that at this time, the original purpose (to find a way to increase the purchase conversion rate of the product landing page) has become three problems, and these three problems will guide the entire data analysis process.

2. Find the object: clarify the feasibility and scope of the analysis

For the above example, we actually need to solve three problems 1by1 and figure out the objects.

Question 1: Is the current product landing page worth optimizing?

The object of this question is not actually the product landing page, but the traffic situation of the source channel of the product landing page and the behavior pattern on the product landing page after the traffic arrives.

This is because if the amount of traffic entering is small, the sample itself may not be capable of testing and verification. Instead, we need to first improve our ability to distribute traffic.

Question 2: Is it feasible to optimize product landing pages?

The object of this question can be either a new landing page or an old landing page. As long as the data can prove whether the adjustment of the landing page will affect the shopping conversion rate, a conclusion can be drawn .

Question 3: To what extent can it be optimized?

This question actually cannot be answered now. To answer this question, it is obvious that there is a prerequisite, which is that the adjustment of the product landing page is feasible for improving the purchase conversion rate.

This way you can identify the object.

3. Data selection: mining related data from identified objects

The second half of selecting data is not difficult, but the first half is not easy.

It involves selecting the time span, the dimension of the object data, etc.

In a word, be logical.

4. Analysis: Be objective and neutral, and gain insights from data

Analyzing this matter actually requires naturally drawing conclusions from the data and being objective and neutral.

You can't just put together data with conclusions in mind, but you can't be confused and not know what the data wants to talk to you about.

5. Revalidate: Take the conclusions that the data tells you and look back at the original problem

What I told the kids was that data should be able to bring:

1. in conclusion
2. Assumptions
3. Action Items

In other words, a data analysis should be able to help you first understand the current situation of the problem (conclusion), help you propose possibilities for the next stage (hypothesis), and help you sort out what to do next (action items).

2. How to establish the logic of data extraction

Logic is very important.

How important is it?

I have seen analysis reports from young students who have just started doing data analysis. Basically, they just pile up a bunch of data, talk about it, come to a conclusion, and that's it.

I just asked, why do we need to collect this data? Is there any correlation between these data? What kind of relationship is it?

A few days ago, my good friend Sanshui, who is also a famous Zhihu user, wrote a column article using poison to refute the negative impact of the Three Gorges Dam. It is worth reading. After reading it, you will know that sometimes the data that are piled up in a serious manner actually have no causal relationship at all.

So, after you spend a few days collecting various data, what you stack up is just a conclusion that you think is correct, which has no value.

To establish the logic of data extraction, the first thing is to train your logical thinking ability.

A relatively light and slightly perverted training method is as follows:

Just pick up a phenomenon and start training yourself to "find relationships."

For example, grab a cup of coffee and you can start training:

1. Is the cup of coffee an Americano, latte, or some other type of drink?
2. The difference between Americano and latte is that latte is coffee with milk added, while Americano is coffee with water added. But why does it take 7-11 longer to make an Americano using a machine than to make a latte, or even double the time?
3. Which link causes the extra time?

…

And so on.

Then, you need to get close to the business and be able to understand the business scenarios.

Take, for example, parking lots.

If I ask you, what kind of scenario is the most frequent for cars?

I believe that at least 50% of people will tell me that parking, refueling and maintenance, especially parking and refueling. Some people will even say that parking is the most frequent because as long as the car drives out, it will stop, and to stop, there needs to be a place to park.

But in fact, the parking we are discussing is a high-frequency scenario, because, assuming that the car owner is an office worker, then basically on weekdays, he has no temporary parking needs, and there is no strong demand for public parking lots.

Because both home and company have relatively fixed parking spaces.

However, if you are not a car owner, or you are far away from the parking business line, when you do data analysis, you will have assumptions and logic that seem to make sense but are actually completely wrong.

Therefore, daily training alone is not enough, you must also fully understand and know the business.

3. Data analysis only takes three pages

This is really not a joke. Of course, due to the different depths of analysis work and the different amounts of data required, the three pages can actually be replaced by three parts.

So, the first part is what Brother Liang has repeatedly emphasized, which is the background, purpose, selected sample size, reference objects, and assumptions about the conclusions you hope to draw.

For example, on the original product landing page, the user purchase conversion rate was less than 0.01%. We needed to study whether this situation could be improved. Therefore, we selected 2,000 users and randomly divided them into two groups. Group A saw the new product landing page, and Group B maintained the original product landing page. Over a 30-day period, we compared the changes in the purchase conversion rates of 1,000 users in each of the two groups. The reference objects were:

Average daily conversion rate of a single person in Group A vs Group B within 30 days;

Average daily conversion rate of users in groups A and B 30 days before and after seeing the page

It is assumed that the low conversion rate of the product landing page is due to unreasonable page design and unsatisfactory page content. After adjusting the new version of the landing page, the page design and page content presentation have been improved. It is hoped that through this data analysis, the possibility of optimizing the product landing page and improving the purchase conversion rate can be found.

The second part is very simple. List all the samples and data obtained for reference and comparison. The necessary data interpretation work can be noted in advance.

The third part is even simpler. Based on the results of data analysis, you can provide feedback, propose hypotheses and organize verification.

Therefore, in fact, each data analysis report should only consist of these three parts.

4. Understand conclusions, assumptions, and action items

In the last part, I will give an example and talk about the conclusion, hypothesis and action items.

Data specialist Xiao Wang made a 26-page data analysis report, which extracted and analyzed data on the operational quality of existing internal channels.

The report detailed the multiple existing internal channels, what content these channels had delivered in more than half a year, and the statistics on related display rates and click-through conversion rates, and then it ended.

So Xiao Wang's leader Lao Mao chatted with Xiao Wang:

“Is this your data analysis report? Why did you do such an analysis?”

“Because the colleagues in charge of channels have raised the need to know the relevant data of the channels in the past six months.”

“Do you want data or analysis?”

“We need data, but we need assistance in analysis.”

"If you want data, you can just provide an Excel spreadsheet, but if you want analysis, this PPT is not up to standard."

"What's not qualified? I worked overtime for three days for this report...the channel said it's pretty good..." Xiao Wang felt aggrieved.

“Maybe they think you have helped them beautify their data, and they themselves have not thought enough about the purpose of data analysis, so they think it is good, but in my opinion, it is not up to standard.” Lao Mao explained to Xiao Wang in detail:

“Look, you listed the half-year data of so many channels. Let me ask you a few questions about one of them:

1. In the past six months, there have been 36 campaigns launched through channel A. What are the differences in the targets and contents of these 36 campaigns? To whom should we deliver what content and at what time will the best effect be achieved? On the contrary, when, in what form, and to what people will the content be least effective? Where is your conclusion?

2. If you have concluded from the data that a certain type of content delivered in a certain form to a certain group of people at a certain time is very effective, then, I would like to ask, is this channel only effective for this type of content delivered in this form to this type of people at this time? Can it also be effective to deliver different content in a different form to other groups of people at other times? What hypothesis did you come up with?

3. If you have already put forward the hypothesis, what do you suggest your colleagues in the channel should do next? Should we conduct a series of tests to collect data, improve channel quality, or conduct a round of communication on user selection and content screening before launching in the future? What action items do you suggest? "

The more Xiao Wang listened, the more pressure he felt, but he also slowly understood what Lao Mao meant by "unqualified".

When we do data analysis, we must always be curious and responsible. We can gain insights into many problems through a small piece of data. These problems cannot be limited to being raised or even seen but ignored. Otherwise, the data analysis will be worthless and can be ignored.

Let’s take a practical example.

The business model of Company A is to charge a fixed membership fee on an annual basis, so the core data of Company A is: paid membership rate.

When breaking down this rate, there are two related indicators:

1. Members’ first payment ratio
2. Member renewal rate

If the data analysis shows that the renewal rate is high but the first payment ratio is low, the first conclusion is:

The membership service provided by Enterprise A has certain value, but there may be certain defects in the description of the membership service and the form, copywriting , and transaction process that stimulate users to become paying members.

Next, we will retrieve the processes and data involved in these possible defects to observe whether there are any problems in these areas.

If so, formulate a hypothesis and verify it through action.

Therefore, once data analysis is started, it is not a one-time action. It is more like a starter for an experiment. By starting an experiment, we can adjust the status of different business developments, find opportunities and eliminate risks.

I hope that everyone can consciously draw conclusions, propose hypotheses and formulate action items. This is why I never mention the concept of " data operation " alone, because every operator, as long as you are involved in the business module, should have the responsibility and initiative of "data query" and "data analysis". This will help you improve faster.

That’s it, call it a day.