There are several classic ultimate problems in the field of data analysis. Multi-factor attribution is definitely the most daunting among them. Especially near the end of the year, the brand, after-sales, customer service, supply chain, operations, product, product management will run, and asked: "This year's performance is pretty good, then the question will come up: This year the company to earn 10 million, in the end a few One hundred million is attributed to the brand, and several hundred million is attributed to the supply ...Please make a quantitative analysis, thank you".
So, how do you analyze it? Today we will explain in detail.
1
The surface of multi-factor attribution
Why this problem is the ultimate problem, as long as you do an experiment, you will know immediately.
You can try it yourself
Step 1: Please close your eyes
Step 2: Recall a recently bought item on Taobao
Step 3: Recall the name, packaging, price, brand, voice of the customer service girl in your mind...
Step 4: Open your eyes and tell yourself: I paid XXX money for this product, of which X% is paid for the name, X% is paid for the packaging, X% is paid for the blingbling advertisement on the product, and X% is paid for the express uncle……
Can you answer it?
Not only is it difficult to answer, it is estimated that many students don’t even remember what they bought recently.
Not all consumers are rational in their consumption.
Not all products are valued by consumers.
The brands, services, and products provided by the merchants themselves are a whole.
Therefore, from the perspective of consumers, this problem does not hold true from the source. Especially FMCG products such as beer melon seeds and mineral water . The price is low, the consumption frequency is high, and there is a lot of impulsive consumption . Buy some when you are in a good mood, and buy some when you are in a bad mood, so it is difficult to make it clear. Large durable goods, such as houses and cars, may think more and have a certain probability of distinguishing, but it is still difficult to quantify the score of each factor (if you doubt this, imagine how entangled you were at the moment of marriage, just understand Now, haha).
So here comes the question: why do you know that the division is not clear, and each department still asks for division again and again?
This touches on the essence of the problem: where is everyone's ass sitting.
2
The nature of multi-factor attribution
On the surface, it is difficult to disassemble the data due to multiple influencing factors.
In essence, the attribution of multiple influencing factors is only the result of the uneven distribution of spoils between departments.
Each department is too eager to prove its value, always trying to connect with performance indicators.
Especially at the end of the year, if you want to fight for bonuses for the department and for the budget for next year, the impulse to separate families is even higher.
Therefore, attribution of multiple influencing factors is essentially a measure of departmental value , which is the core difficulty.
Many students will ignore this core issue and use some simple data methods to deal with it. For example: Set the cost of each department as x, set the performance as y, and use a linear regression model. Then consider the coefficient of each parameter as the contribution size. Not to mention, in doing so, first of all, I completely misunderstood the meaning of the regression model; secondly, I did not consider the issue of categorical variables and continuous variables. Just the result itself will be sprayed to death.
For example, if the coefficient of the calculated sales is 2 and the coefficient of the supply chain is 1, then we will find twice the sales next year, but only provide 1 times the goods. Can we still have this sales performance? It must be impossible, with a gun but no bullets! The division of labor and cooperation between departments is not a simple 1+1=2 relationship . This is common sense . Therefore, forcibly splitting the links between departments and pulling together different types of departments for evaluation is destined to hit the streets.
Therefore, the idea of breaking the game is that from the very beginning, we should directly reject this idea of a formula to beat the world. Proceeding from the nature of departmental work, establish a scientific measurement mechanism to effectively resolve this anxiety about meritorious service .
3
Breaking the game
If you want to break the game, first clarify the work types and contribution methods of each department (as shown below)
After that, you can analyze and investigate in different categories.
Soft support category: Give up directly related sales performance and allocate on demand.
The core problem of soft support is that you cannot prove your innocence . As far as brand promotion is concerned, even if all the promotion is connected to the purchase page, it is impossible to prove how big a user’s purchase is because of the brand, and even these years of popular styles and Internet celebrity products can weaken the brand and highlight Product features and fan effect. Moreover, at least 60% of the brand promotion does not even have a link to the goods (such as pre-market promotion), so it is impossible to talk about it. The same goes for customer service and after-sales service, although these two services are very important when customers come to the door. However, the proportion of proactively initiated customers is small, so it is difficult to correlate overall performance.
If you don’t do it, you don’t know the effect. It’s best to allocate resources directly in proportion to the overall performance and evaluate your own results, instead of forcing related sales performance.
For example, according to the product life cycle/time, the configuration of publicity (as shown in the figure below) can achieve sufficient market awareness and cover enough people (assessment of clicks, forwarding, reading, etc.) to complete the task.
Such as customer service and after-sales service. Allocate resources and manpower according to the total business volume. Assess its own service satisfaction, service coverage, speed from call to response, response speed to serious complaints/risk events, etc. Do your job well and complete tasks.
Hard support category: assessment.
The assessment of hard support is much simpler and clearer: the supply is in place and the loss is reduced. And what we pursue is peak control and long-term decline. Too much care about the gains and losses of one city and one place, but it is easy to come up with a very rigid process and many oolongs (as shown below)
Hard pull category: Introduce the ABtest mechanism to preset goals in advance.
The hard pull type is a method of superimposing buff, so the control variables must be set in advance, otherwise it will be mixed in a bunch of factors and cannot be split afterwards. For example, set the overall goal of pulling in advance, test the effect of the plan in advance, leave the reference group in the event, and collect process data so that you can make a distinction afterwards. The hard pull category can be evaluated. The problem always lies in: not doing work in advance, leaving no reference in the matter, and not stepping on data. Nothing, it's hell if you can analyze it afterwards.
Core process: Establish a grading mechanism and analyze the impact in depth.
In the core process, it is common for sales and products to interact with each other, but this type of interaction can be analyzed to find out who is at fault. As long as the grading mechanism is established, the channel quality and product attributes are labeled and analyzed, and the process conversion rate is monitored, in-depth analysis can be done. Therefore, the core process should be analyzed as much as possible without leaving room for wrangling (as shown below)
4
Reality is always skinny
The above are only suggestions under ideal conditions, and the actual start of construction:
● There are always people who want to fight for more resources, shouting: "Integration of product and effect!" "Mental resources!"
● There are always people who think that after-sales service for customer service is unnecessary, and they will be deducted more this year? (Thus triggering a counterattack from the service sector "I also contribute!")
● There are always people who like to exaggerate their contributions, and the benefits of promotional activities are written so high, even higher than the natural sales.
● There are always people who are afraid of being held accountable and desperately want: no promotion! No support! The product does not push up responsibility.
Therefore, the issue of “how much each department contributes, and whether it can be specific to who comes from the few cents of each dollar” will never stop.
In addition, there are always new data analysts who think that as long as a few data are entered into a linear regression or factor analysis model, a parameter can be calculated to satisfy each department. So this kind of babbling will continue for many, many years, hahaha.
5
summary
There are many similar data analysis problems for the ages:
● Why is the experiment in ABtest effective but not effective when put into production? How can the test be accurate?
● How to calculate the natural growth rate is the most fair and reasonable in the universe!
● How to measure the development of users' mental resources and the changes in cognitive depth!
● Sales forecast, how can the forecast be 100% accurate!
● ……
Every problem is what looks like data analysis on the surface, behind it is the greed of the people, who have pushed back the power. If the performance is good, it is said to be done by itself. If the performance is not good, it will be left to external factors. Various internal factors that cannot be quantified are used to protect themselves. If it doesn't work, it means that the data analysis ability is not good. Why don't you predict it earlier! Students who are doing data analysis should have a clear enough understanding of these issues and don't be fooled easily. For example, what seems to be the simplest: how to calculate the natural growth rate, and how to convince the business department? If you are interested, I'm watching this collection of 60 , let's share the next one.
Original selection:
If you don’t understand the business, how can you communicate with the business department and maintain your neutrality and authority? Teacher Chen gave a detailed explanation in the video course of "Business Knowledge One Stop" . In the course of study, you can also join a group of students to discuss issues with Teacher Chen one-on-one to improve your ability.
Click " Read Original " in the lower left corner to listen to Teacher Chen's lecture!