0
$\begingroup$

I want to predict conversion rates for an eCommerce store. I have data from Google Analytics with features like averageSessionDuration, bounceRate, numberOfVisitorsBySource etc. and the corresponding conversion rate.

That's the specific background for a generic question: I am not sure how I should select the time units for my data. Should I export the data by day, by week, by month or by year?

  • When I have one row per day, I have a lot of data (a lot of rows), but also a lot of noise, because the conversion data for one day is sparse.
  • When I have one row per year, I have less noise, but also less data (less rows).

In general, what's the right approach to select the right time frame per row in such a case?

$\endgroup$
0
$\begingroup$

I think you should first export it with a timestamp and then add a day, month ... other features to the data. This might help in better accuracy.

For more details, you can refer: This article

$\endgroup$
0
$\begingroup$

At the start, it'll be better to discuss the use-case with a Domain Expert, if possible, and decide if predicting Conversion Rates over a specific future time horizon say daily/weekly/monthly/quarterly/annually would make sense to your Business using historical data. In some use-cases, predicting over a relatively short term horizon of 1 day or week may be more beneficial than predicting over a medium term horizon of 1-2 months. Ofcourse, you'll have to explore the dataset later to see if it has any predictive power to forecast over that intended time horizon.

Based on that discussion, you could either fetch aggregated data/stats as per the desired time granularity or simply fetch data at the lowest available granularity (say daily). In case of a sparse daily dataset, try aggregating data at the weekly level.

Next, visually explore the dataset at the desired or different higher time granularity values and see if there is evidence of relationship of the Dependent variable (Conversion Rate) with the prior lagged Independent variables (lagged by the intended prediction horizon of n days/weeks/etc.).

Depending on the nature of the relationship, you could then build traditional Regression models, Econometric models such as ARIMA or VAR or even build advanced ML Ensemble Models (using XGBoostRegressor or RandomForestRegressor) or Neural Nets. Also, look into Facebook Prophet package for this use-case.

Example of a Linear Regression expression for this problem in which I'm trying to predict the value of conversionRate using 1-period lagged value of independent variables:

conversionRate(t) = B0 + B1 * averageSessionDuration(t-1) + B2 * bounceRate(t-1) ...

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.