Overview

‍

Forecasting website performance is a reoccurring deliverable we provide to clients. Therefore, understanding the history behind what was tested, what worked and what didn't was crucial. Each year we analyze what would be a reasonable scope to improve the data science model used.

In FY19, the following was used:

• Multilinear regression model used for primary website traffic
• FY18 rates for secondary visit types were applied on predicted visit traffic

For the FY20 models, we needed to consider the limitations on context in the data as well as forecasting primary and secondary KPIs that were missed in the FY19 model.

Solution

‍

• Mutlilinear regression model continued to be used for primary website traffic. This model approach was also applied to secondary visit types.
• For paid media budgets we considered Inflation rate
• Correlation and p-value tested conducted for model accuracy
• Normality, linearity and homoscedascity were evaluated for model stability

‍

Results

‍

• Improved primary visit RMSE by -10% and r-squared by +6%
• Improved secondary visit type 1 RMSE  -10% and establishing an r-square of 0.68 as this was the first year building a multilinear regression model for this KPI.
• Improved secondary visit type 2 RMSE -10% and establishing an r-square of 0.70 as this was the first year building a multilinear regression model for this KPI.

‍