EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
4181. Estimation of Online Advertising Effectiveness using a Coarsened Exact Matching Framework
Contributed abstract in session MA-31: Recommender systems, stream Analytics.
Monday, 8:30-10:00Room: 046 (building: 208)
Authors (first author is the speaker)
1. | Vamanie Perumal
|
Engineering Design, Indian Institute of Technology Madras | |
2. | Palaniappan Ramu
|
Engineering Design, Indian Institute of Technology Madras |
Abstract
Geo experiments are a method of choice to estimate online advertising effectiveness. It is challenging to relate ad spend (treatment) in each geo and the ensuing time series responses, such as sales. The return on ad spends (ROAS) is the difference between responses obtained during treatment (after-data) and the counterfactual responses if treatment had not occurred (before-data). Since we obtain only after-data for treated geos, evaluating the ROAS depends on how well the before-data are modelled. A typical approach for modelling before-data includes estimating the behavior of treated geos using the behavior of control geos. An inevitable problem is to find control geos homogeneous to treated geos so that any difference can be directly attributed to the treatment. To tackle the challenges mentioned above, we propose a time series based statistical matching framework. The selected time series features used as covariates vary naturally among all geos without being influenced by the treatment. We achieve a balanced covariate distribution between control and treated geos using Coarsened Exact Matching (CEM). The weights obtained from CEM are used to build a time-varying negative binomial regression model for counterfactual estimation of before-data for treated geos. We demonstrate results using real-time online advertising data. We observe significant improvement in homogeneity between control and treated geos, thereby improving the accuracy of incremental impact measurement.
Keywords
- Analytics and Data Science
- Forecasting
Status: accepted
Back to the list of papers