Background
Marketers want to answer questions like, how many additional daily clicks or conversions were generated by an advertising campaign? Answering a question like this can be difficult when a randomized experiment is not available. It’s quite expensive and time-intensive to set up control and test markets. Also, it’s actually hard to measure such impact in the real world because there are many attributes that can influence the outcome. This is called ‘noise’.
Summary
Lift Analysis is an approach to estimating the causal effect on sales, of a designed intervention (marketing). We use Bayesian structural time-series model as a method that looks at a historical sales pattern and builds a prediction of its future course, allowing us to compare what actually happened following our intervention. Given a response and a set of controls (e.g., sales in non-affected markets) this model is used to try and predict how the response metric would have evolved after the intervention if the intervention had never occurred.
Data Inputs
The model requires historical sales data with the following information:
Date (by day)
Number of Events (count of individual sales, not revenue)
If there are only a few SKUs, this can be broken out by SKU.
Market Name
Market Type: control or experiment
If you have data from multiple markets, where marketing was active in one market and wasn’t in another, it would be helpful to provide data of the ‘untreated’ markets but not necessary.
We can use more than one market as a control. More control markets are encouraged, the model will aggregate multiple control markets to reduce noise. If no control markets exist, we can generate predictor controls - datasets that have a high correlation with the sales data provided.
At the minimum, we need at least 2 months of events data before the campaign (more data = more confidence), and then throughout the duration of the campaign.
Obfuscating Data
If you need to obfuscate data for security purposes, you can simply pick a random number, and multiply the number of events by this same random number. When we provide the estimated number of attributed sales events, you can then simply divide by this number to get to the accurate amount.
Data Delivery
We have two main options for data delivery, but if you’d prefer something different then please let us know.
We can create an S3 or Google Cloud bucket and send you the credentials. You can then upload the data via the user interface or via the command line.
If your data is available only via API, we can pull the data ourselves, although we may request a minimum campaign size or custom build fee.
Timing
We can start running the analysis as soon as 2 weeks after the campaign starts, and we can provide a weekly or monthly update after that. Alternatively, we can provide a single report at the end of the campaign. It’s up to you – of course, to run the report, we need the sales data, so it depends on the cadence you’d prefer.
