Chapter 1: Understanding and choosing a model

Single-Touch Attribution (STA) Models

When to use them:

• Suitable for simple customer journeys with few touchpoints, or when you want to evaluate the impact of a single touchpoint.
• Best suited for campaigns where it is easy to identify an initial touchpoint or conversion.
• Single-Touch models are recommended for answering specific questions such as:
• Does affiliate marketing cannibalize the performance of my other channels?



Add a caption...

• Does SEA Brand cannibalize the performance of my other channels?



Add a caption...

• They are also recommended for one-off testing of new media devices.

Choosing the best model:

• “First Paid Click” Model: Use if you want to understand which channels are driving customer engagement, for example by finding out which channels generate the most awareness.
• “Last Paid Click” Model: Use if the last touchpoint is considered most responsible for the conversion, useful in direct response campaigns.
• Custom Model: Customization of the decision tree according to the specific needs of the Client:
• with or without post-printing,
• prioritization or deprioritization of certain interactions (eg: CRM campaign click)
• exclusion of brand keywords
• ....

Multi-Touch Attribution (MTA) Models

When to use them:

• Ideal for complex customer journeys with multiple touchpoints across different channels, where understanding the complete customer journey is crucial.
• Allows you to have a more global vision of how different interactions contribute to conversion.

Choosing the best model:

• Data-driven model (Markov or Shapley): Ideal when you need a custom statistical model that adapts to customer journeys and awards credits based on performance.
• Markov Algorithm: Use when you want to account for the suppression effect, which helps understand the contribution of each channel by simulating its suppression. Markov is particularly suitable for media mixes that are extremely rich in terms of the number of campaigns and touchpoints.
• Shapley Algorithm: Use when the goal is to fairly distribute conversion credit across all touchpoints based on their contribution, taking into account all possible interactions.
• Linear Model: Use when all touchpoints are considered of equal importance.
• U-shaped model: Used when the first and last touchpoints have more value, but the intermediate touchpoints still contribute.
• Time-deprecating model: Ideal when recent interactions are more likely to result in conversions, especially for products with shorter consideration periods.

Marketing Mix Modeling (MMM)

When to use it:

• Effective for measuring the impact of media investments over long periods, including for offline channels.
• Useful when user-level data is not available or privacy concerns limit granular tracking.
• Ideal for analyzing a media mix rich in offline data for high-level strategic decisions, especially for cross-channel budget allocation.
• Model capable of integrating exogenous factors (weather, seasonality, brand awareness) to understand their correlation with business results.

Choosing the best model:

• For each Client, the Eulerian platform will create multiple iterations of models using different combinations of variables, smoothing, and transformations. Each iteration will produce slightly different results.
• The Eulerian platform will evaluate the performance of model iterations using validation metrics:
• R² (Coefficient of Determination): Measures the proportion of variance explained by the model. The higher the R², the better, but avoid overfitting the model.
• NRMSE (Normalized Root Mean Squared Error): Indicator of the accuracy of the model by normalizing the prediction errors. A lower NRS indicates better accuracy.
• Cross-validation : Eulerian uses cross-validation to assess the robustness of the model. The platform ensures that performance remains stable across different cross-validation cuts.
• Decomp.rssd (Decomposition Root Sum of Squared Differences) : The decomp.rssd is a specific metric used in Robyn to assess the stability and quality of the decomposition of contributions of model variables (such as advertising expenditure) relative to the total sum of contributions.
• The decomp.rssd measures the deviation between predicted contributions and expected (or average) contributions over a series of simulations or iterations. A lower decomp.rssd indicates better stability and consistency in the distribution of the effects of the explanatory variables.
• This metric is particularly useful for checking the robustness of the marginal contributions of the model and for ensuring that the model does not exhibit large variability in the decomposition of effects during simulation iterations.

Summary