Anyone who has done consumer-oriented product solutions knows that before each project starts, customers will put forward some requirements or worry about the current marketing situation, for example, we want to know who our potential consumers are; How to send coupons is the best; In other words, what kind of new products should be launched to win the reputation and favor of consumers? In the quantitative decision analysis method, we call this series of early demand: customer demand or future expectation.
Next, you need to know the current situation of this problem, such as what the consumers of existing products or services are like, what the effect of coupons issued before, what the sales trend of the market is now, and so on.
When we understand the customer's needs and current situation, we need to slowly get rid of the cocoon and find a solution to fill this gap.
Generally speaking, consultants or analysts without any methodology or experience will be at a loss when they hear these expectations of customers. They don't know from which angle to cut in, what data to collect, what assumptions to make and what methods to analyze.
In fact, problems like this are methodological, and we can use a four-step method to build a bridge between the present situation and the future.
Step 1: descriptive analysis-what?
Find the problem. We can use the analogy of seeing a doctor. A patient went to see a doctor and said that he was not feeling well recently. So the doctor asked the patient to describe further what went wrong. It's the same here Take preferential sales promotion as an example. We will first know whether the customer has done similar promotion cases in the past, when and how effective it is. Through these problems, a series of KPIs have been produced.
There are several ways to generate KPIs:
1) We ask questions and customers answer them.
2) Obtaining information from the database of the client company (SQL)
3) Obtaining information from external data (third-party data enhancement)
4) Competing partner information
5) Policy information
6) Semantic analysis
7) Others
Tools for obtaining KPI:
1) Q&A (discussion, phone call, email, SMS, questionnaire)
2) Database (SQL)
3)Excel
4)R, Python and other software
5) Website search information
6) natural language learning
7) Others
Analyze these KPI variables:
These KPIs can be absolute numbers, percentages or indices. It can be the comparative data of different periods in the past, different groups (such as crowd grouping and pattern grouping) or competitors.
Usually KPI analysis methods are:
1) univariate analysis
2) Bivariate analysis
3) Multivariate analysis
4) Hypothesis verification
5) Simple modeling (clustering and grouping)
By analyzing these KPIs, it can help us to form:
1) Portraits of existing consumers
2) Portraits of potential consumers
3) Portrait of loyal customers
4) consumer value grouping
5) Others
Step 2: Diagnostic Analysis (Why)
Answer the question. We also use the example of a doctor seeing a doctor as an analogy. When the doctor asks questions to the patient, the doctor begins to use his knowledge to diagnose the patient's condition through consultation and X-ray.
In analytical methods, this step usually requires:
1) Understanding causality
2) Understand the sensitivity among various factors.
We need to know what caused the current market situation, or what caused it. For example, in the previous stage, we got 50 very useful KPIs, and through causal analysis, it played an important role to determine 10 KPI. After drawing a conclusion, we will ask, among these 10 factors, what is the individual contribution of each factor, some may be high, and some may be relatively low.
For this problem, we can get the contribution of various factors through modeling, and the model can also play the role of eliminating highly correlated variables. Another reason for using the model is that when there are hundreds of factors, it is difficult to find the most useful one among so many factors by traditional methods. In this case, a model is also needed to help select variables. The last reason is that we can determine whether this factor is positive or negative.
Through the results of modeling, we can get the following models about consumers:
1) loyalty model
2) Satisfaction model
3) Price sensitivity model
4) Attribution model
5) Customer churn model
The algorithms behind these models are:
1) linear regression
2) Logistic regression
3) Decision tree
4) Time series
5) random forest, boosting, SVM, PCA, etc.
Step 3: Forecast and Analysis
Predict the right time and get the pre-emptive marketing effect. With the preparation for the first step and the second step, we need to predict what changes and impacts will occur if I make some adjustments.
The models used are:
1) Intention Scoring Model
2) Brand loyalty score
3) Purchase channel preference model
4) Usage habit of catalyst
6) Sales volume forecast
5) Survival analysis model
For example: intention scoring model. We find that if we use the existing factors, the conversion tendency of consumers may be 60%, but if I make some adjustments to some factors, such as: I send two more advertisements to existing customers, the possibility of customers buying will rise to 65%; If five more advertisements are sent to customers, the possibility of customers buying them will rise to 85%. Through this adjustment, I can estimate the future advertising cost, or the income brought by the transformation.
For another example, through the time series model, we can predict that 654.38 million consumers will buy a certain brand model next year, so that we can make a preliminary preparation for next year's production plan and marketing plan.
Step 4: Application of decision analysis
1) provides strategy suggestions.
2) Optimization
3) Market simulation
4) A/B test
The example of the third step mentioned more than two advertisements, with a conversion rate of 65%; The conversion rate of five advertisements is 85%. So what if there are three more? What will happen if four more advertisements are issued? Academic circles have been looking for the best and perfect answer to this problem: how many advertisements can I send to maximize profits?
As we all know, when making a regression model, there are the following assumptions:
1, and the random error term is a random variable with expected value or average value of 0;
2. For all the observed values of explanatory variables, the random error term has the same variance;
3. Random error terms are uncorrelated;
4. Explanatory variables are deterministic variables, not random variables, and have nothing to do with random error terms.
5. There is no exact linear relationship between explanatory variables, that is, the sample observation matrix of explanatory variables is full rank matrix.
6. The random error term obeys the normal distribution.
In fact, it is difficult to achieve this ideal state in real life, and the concept of maximization, from a mathematical point of view, will involve the problem of finding the optimal extreme value. Many times, we actually get the solution of local optimization, not the solution of global optimization.
Therefore, in this case, the market simulation method is derived from management science to determine the final plan. One of the most famous methods is sand table simulation, but when these simulations really land, there will be a gap with the previous results.
So in recent years, more and more companies choose to do A/B testing. When you are not sure about several schemes or are not particularly confident about the prediction results, the emergence of A/B testing solves these concerns. A recent successful case is that Amazon passed the A/B test and put the "order" from the account bar to the menu bar on the home page, which brought considerable revenue growth to the company.
A/B test should pay attention to:
1) number of samples
2) Selection of crowds
3) Time span
4) significance statistics
The whole decision analysis method is a ladder and a closed loop. According to the actual market reaction, further analysis and iterative optimization are carried out.
After reading the whole quantitative decision analysis method, you should have a certain framework for consumer-centric big data solutions.