Thinking | 10 kinds of thinking for data-driven decision-making

Many people say yes Data analysis People are smarter than others. In fact, they are "smart" and have analysis thinking Today, we will talk about common data analysis thinking.

The following 10 data analysis thoughts may not instantly upgrade your thinking mode, but may bring you a "flash" feeling for your future work. Please read patiently, and don't forget me, your data hunter DataHunter.

In daily work, customer segmentation, product classification, market classification... many things need classification thinking. The key is that the classified things need to be able to distance themselves from the core key indicators! In other words, the results after classification must be significant.

As shown in the figure, the horizontal axis and vertical axis are often you operate The focus is on the core indicators (not limited to two-dimensional, of course). After classification, you can see that their distribution is not random, but has a significant tendency to cluster.

For example, classic RFM model Build a user clustering system based on the three core indicators of charging: the latest consumption time (Recency), consumption frequency (Frequency), and consumption amount (Monetary).

In terms of the three indicators of R/M/F, we divide the actual users into the following eight regions through experience (as shown in the figure above). What we need to do is to promote the transfer of different users to more valuable regions. That is to match each paying user to different user value groups according to their consumption behavior data, and then adopt different strategies according to the value of different paying user groups (see the table below)

One of the development of classification thinking is Matrix thinking , matrix thinking is no longer limited to the use of quantitative indicators to classify. In many cases, when we have no data to support us and can only make subjective inference through experience, we can combine some important factors into a matrix, roughly define the good and bad directions, and then analyze them. You can use Baidu's classic management analysis method "Boston Matrix" model.

This way of thinking has become more popular. Funnel analysis can be divided into long funnels and short funnels. The characteristics of the long funnel are that it involves many links and has a long time cycle. The commonly used long funnel includes channel attribution model, AARRR model, user life cycle model, etc; A short funnel has a clear purpose and takes a short time, such as order conversion funnel and registration funnel.

However, it seems that the more universal the model is, the easier it is to understand, and its application should be more cautious and careful. In funnel thinking, we should pay particular attention to the length of the funnel.

Where does the funnel start and end? There should not be more than 5 funnel links, and the percentage value of each link in the funnel should not exceed 100 times the magnitude (the conversion rate from the first link of the funnel to the last link should not be less than 1%). If these two numerical standards are exceeded, it is recommended to divide them into multiple funnels for observation.

What is the reason? If there are more than 5 links, there will often be multiple key links, so analyzing multiple important issues in a funnel model is prone to confusion. The difference of numerical magnitude is too large, and the fluctuation relationship between numerical values is difficult to be detected, and information is easy to be missed.

When we observe indicators, we not only need to see the changes of individual indicators, but also need to observe the relationship between indicators. There are positive correlation (red solid line in the figure) and negative correlation (blue dotted line). It is better to calculate the correlation coefficient between indicators frequently and observe the changes regularly.

Related thinking The application is too wide and is often ignored by everyone. The problem faced by many enterprise managers today is not that there is no data, but that there is too much data, but too little useful data. One of the applications of relevant thinking is to help us find the most important data and eliminate the interference of too many messy data.

How? You can calculate the correlation among multiple indicators that can be collected, select the data indicators that have relatively high correlation coefficients with other indicators, analyze their generation logic and corresponding problems, and if all meet the standards, this indicator can be positioned as a core indicator.

It is suggested that you should make a habit of calculating the correlation coefficient between indicators, and carefully consider the logic behind the correlation coefficient. Some are obvious common sense, such as the number of orders and the number of buyers. Some may surprise you! In addition, "no correlation" is often a source of surprise.

In general, when describing the branching of a logical tree, the concepts of "decomposition" and "summary" will be mentioned. I change it here to make it closer to data analysis, which is called "drilling down" and "rolling up".

Drilling down is to analyze the changes of indicators and continuously decompose them according to certain dimensions; The roll up is a summary in the opposite direction.

Drill down and roll up are not limited to one dimension. They are often nodes of multidimensional combination for bifurcation. The logic tree extends to the algorithm field policy decision Trees. The key is when to make a decision (judgment). When branching, we often choose the dimension with the greatest difference for splitting. If the difference is not big enough, the branch will not be subdivided. Nodes that can produce significant differences will be retained and continue to be subdivided until no differences can be distinguished. Through this process, we can find out the factors that affect the change of indicators.

For many problems, we cannot find the method and object of horizontal comparison, so it will become very important to compare with the historical situation. In fact, many times we use time dimension comparison to analyze problems, which is convenient to eliminate some external interference, especially suitable for innovative analysis objects, such as a company in a new industry or a new product.

The thinking of time series has three key points:

First, the closer the time point is, the more attention should be paid to it (the depth in the figure, the more recent the event, the more likely it will happen again);

Second, it is necessary to do year-on-year (the line in the figure indicates that indicators often have some periodicity, which needs to be compared at the same stage of the cycle to be meaningful);

Third, attention should be paid to the occurrence of abnormal values (for example, the historical lowest value or the historical highest value, it is recommended to add the average value line and the average value plus or minus one or two times the standard deviation line when mapping the time series to facilitate the observation of abnormal values).

One sub concept of time series thinking must be mentioned, that is, the concept of "life cycle". The product life cycle theory (PLC model) was proposed by American economist Raymond Vernon, that is, the whole process of a new product from entering the market to being eliminated by the market. Users, products, people and things all have life cycles.

With the improvement of computer computing ability, the method of queue analysis is gradually emerging. From an empirical perspective, queue analysis is to slice the observation objects at the time granularity according to certain rules to form an observation sample, and then observe the changes of some indicators of the sample over time. At present, the most frequently used scenario is retention analysis.

In queue analysis, indicators are actually time series, but the difference is measurement samples. In cohort analysis, the measurement samples vary in time grains, while the samples of time series are relatively fixed.

The concept of cycle/closed-loop can be extended to many scenarios, such as business process closed-loop, user life cycle closed-loop, product function use closed-loop, marketing strategy closed-loop, etc.

The closed loop of business processes is easy for managers to define. It lists all the business processes of the company, sorts out the business processes, defines the indicators that affect each other among the processes, tracks the changes of these indicators, and can grasp the company's operation status from a global perspective, such as the pulse of business processes (see the following figure). The advantage of circular thinking is that you can quickly establish a logically related indicator system.

Logical thinking means understanding the value chain and the relationship in various data, that is, causality.

The key to this method is to understand the relationship among them, which requires you to understand and be familiar with the work, be careful and careful, and be clear about the relationship between sufficiency and necessity. In fact, it means: What data do you need? How to obtain these data? What is the relationship between the data? The most common method here is A/B test.

How to refine this concept? First, when conditions permit, try to do comparative tests before making decisions; Second, when testing, we must pay attention to the selection of reference groups. It is suggested that a group of samples without any change should be kept in any experiment as the most basic reference. Now data acquisition is more and more convenient. On the premise of ensuring data quality, we hope that you can do more experiments and find more rules. You can do experiments according to the following table.

Exponential thinking is the most important of the 10 thoughts shared today. Many managers face the problem of "too much data, too little available", which requires "dimension reduction", that is, to compress multiple indicators into a single indicator. Exponential thinking is a way to combine multiple factors that measure a problem into a comprehensive index (dimension reduction) after quantifying them respectively, so as to keep tracking.

The benefits of indexation are obvious: First, the indicators are reduced, which makes the managers more concentrated; Second, indexation indicators often improve the reliability of data; Third, the index can be used for a long time and easy to understand.

The design of the index is a big question. Here are three key points: First, we should follow the principle of independence and exhaustion; Second, we should pay attention to the units of each indicator, and try to standardize to eliminate the influence of units; Third, the sum of weights should be equal to 1.

PS: The principle of independent exhaustion, that is, when collecting multiple indicators to measure the problem, each indicator should be as independent as possible, and the indicators that can measure the problem should be as exhaustive as possible. For example, when operators consider whether to distribute their own content to other platforms, they can use exponential thinking to score the whole.

We have finished sharing 10 data analysis thinking modes. Of course, there are more than 10 in our work. We will share them with you in the future. In addition, if you have additional good data analysis thinking modes, you can share them with me!

Originated from WeChat official account (DataHunter)