At the redemption level, the influencing factors mainly include the week factor, credit card repayment, holidays, active consumption, Ali activities, yield (and capital market conditions), transfer out to card transfer, etc.

Week factor: Similar to the subscription, the redemption on weekends is deferred to the account. Therefore, the overall weekend redemption is far less than the subscription;

Credit card repayment: The credit card repayment date is often fixed at several relatively large dates each month, especially in the latter half of the year when the card repayment amount is large;

Holidays: Before many holidays, users will have more capital needs and tend to transfer out;

Active consumption: Many users pay with Yu’e Bao. The active increase in overall consumption will lead to an increase in the redemption of Yu’ebao;

Ali activities: mainly promotions such as Double Eleven, which have a huge impact on the redemption of Yu’e Bao;

Capital market: When the stock market skyrocketed and the IPO restarted, new funds and funds flowing into the stock market had a great impact on the redemption of Yu’e Bao.

Regarding the factors affecting the application and redemption of Yu’e Bao, the above are some of the more important ones, which are temporarily shared here.

Third, the establishment of the model

The core logic of the model adopts the decomposition method of time series, and considers the daily redemption results to be affected by four factors: long-term trend factors, seasonal factors, weekly factors, and other factors.

Finally, use the multiplicative model to multiply each factor parameter:

How to predict the redemption amount of Yu’e Bao based on time series analysis?

The specific modeling process is as follows:

1. Data preprocessing

First, the preprocessing of subscription and redemption data is carried out, and outliers are eliminated, mainly including the subscription and redemption of large-value institutions and the Double Eleven promotion.

The large-amount redemption of whitelisted institutions is often hundreds of millions of dollars in a single transaction, which has an impact on the model, so it must be eliminated first; in addition, the double 11 activities are too low-frequency and the __country email list__ situation is different every year. culling.

In addition, due to the rapid development of Yu’ebao in 2013, in terms of data usage, I directly started modeling with the data of 14 years, excluding the impact of the business outbreak period, so there are about 8 months of data as modeling data.

Determination of long-term trend factors

The long-term factors here are actually mainly the rate of return factors mentioned in the above analysis, the active use of users' Yu'ebao, etc.

It can also be seen from the screenshot above that the total amount of Yu’e Bao’s redemption is in an upward trend; what we want to predict here is this trend.

In terms of specific implementation, on the basis of data preprocessing, a linear regression model is established.

How to predict the redemption amount of Yu’e Bao based on time series analysis?

The above picture is actually a schematic diagram, and the data has reached 17 years. In the modeling situation at the time, the data I used was the data of the past six months as a long-term trend forecast.

Predicting with too long data will lose its meaning and cannot reflect the current situation well, because historical data will have great interference.

Determination of Periodic Factors

After the long-term trend factors are determined, the daily data is divided by the long-term factors to exclude the influence of the long-term factors, leaving only the cyclical factors and other factors.

The cycle factor here actually includes the weekly and monthly factors.

There are many ways to determine the impact factor of the week factor. I calculated the dynamic average of the same period of history, that is, the average value of the last 3 months.

How to predict the redemption amount of Yu’e Bao based on time series analysis?

After obtaining the parameters of the week, divide the parameters that exclude the long-term trend by the parameters of the week again to exclude the influencing factors of the week. At this time, it is mainly the influencing factors of the month.

Do the same for the month factor, and get the monthly factor parameter, as follows:

How to predict the redemption amount of Yu’e Bao based on time series analysis?

It is obvious from this figure that No. 1, No. 10, and No. 15 are indeed small peaks for subscription.