Many people who have read MBA, or who have taken some **chain courses**, as soon as they see the demand, they start to calculate the exponential smoothing method, and what normal distribution calculates the mean and calculates the variance. Why is exponential smoothing good? Why use a normal distribution?
People with this kind of understanding have never answered a reliable answer. It is generally just convenient to apply, because it is more mathematically concise and relatively acceptable to those who take these courses.
For example, the exponential smoothing method is only suitable for the situation that the mean does not change much, if the sales volume has been growing, then it is impossible to get bigger and bigger by giving the historical value of the past period of time according to different weights, so that the obtained ** will always be small. What kind of ** is reliable? We can do experiments!
Let's assume that we use these thousands of **methods** to meet the needs of the next three months, so after three months, let's see which one is more accurate? This is the so-called back-testing, Chinese called post-testing, that is, we take the data from 3 months ago and feed various methods, so that they are respectively ** the demand in the past 3 months, and then compared with the actual demand in the past 3 months, which is closer, then that method is relatively more suitable for our business. Isn't that more reliable than picking a particular method?
Make good use of time series features.
Since we want to demand quantity, and demand is a time series, we can also take advantage of the characteristics of time series. As I mentioned in my previous article, a time series can be broken down into.
Time Series = Trend + Periodic Change + Special Event + Stochastic Fluctuation.
As shown in the figure below, the first one is our time series, which can be broken down into the trend of the second chart, the periodic changes of the third chart, and the special events + random fluctuations of the fourth chart.
The extent to which special events affect the amount of demand, sometimes we can, sometimes not. For example, Double 11, if we experience Double 11 for the first time, it is impossible because the historical data does not. But if we have experienced Double 11 many times, it can be put into the model together.
Let's talk more specifically about these components.
Trends are something we can do, some can build economic models through the analysis of business and mechanism, and some purely use various regression models to do it, whether it is smoothing, fitting, interpolation, or statistical regression, or machine Xi, or even various deep Xi methods.
Cyclical changes are the easiest to do, and they're also very realistic. In real consumption, mid-week and weekend consumption is very different.
Special events are what we have encountered many times, and we have specially marked some outliers on special dates, which is also possible, but it is difficult to be very accurate, but it is better than not. Inventory management requires special treatment for this.
Random fluctuations are impossible, but we can characterize its uncertainty through the magnitude of change and variance, which can become the basis for the formulation of safety stock for inventory management.