Basic steps for forecasting:

[sumaiyakhatun27]

We divide the cohorts into groups: “old” (older than 8 months), “relatively new” (from 8 months before the start of the forecast period) and “new” (no data);

we move from the metric itself to the coefficient: coefficient = orders / cohort size. This is necessary to bring all forecasts to the same scale;

deseasonalize the coefficients. That is, remove seasonal fluctuations to obtain more macedonia mobile database accurate forecasts;

for each cohort we construct a basic forecast based on the model described above;

We restore seasonality and return to the original metric.

Optimizing customer acquisition

Now comes the fun part. Knowing how different customer groups behave, we can calculate how many new users we need to attract each month. Solving a common nonlinear optimization problem can help us do this.

For "new" cohorts, you need to set its size, that is, the first point on the graph. We feed the model with data on the expected number of new clients for each channel for each month. Together with our marketing colleagues, we know approximately how the distribution by channels will change and focus on the upper limits for the total number of new clients in different months.