This guide shows how NeuroRadar isolates the impact of each marketing channel on revenue using statistical transformations, regularized regression, and automated model selection.

Step 1: Raw Spend This is the actual marketing spend by channel and week. It's the input before any transformation or modeling takes place.

Step 2: Adstock (Carryover Effect) Robyn uses adstock transformation to capture the lagged effects of media spend across time. The carryover effect is represented as:

Adstock_t = Spend_t + θ × Adstock_{t-1} (θ = 0.7)

Step 3: Saturation (Diminishing Returns)

This step models diminishing returns of media, using a Hill function:

Response = x^α / (γ^α + x^α)

Step 4: Ridge Regression

Robyn estimates the beta coefficients using ridge regression (L2 regularization). Channels with low predictive value get coefficients closer to 0.

Step 5: Apply to All Channels

The above steps are repeated for each marketing variable: TV, OOH, Facebook, Search, Newsletter, etc.

Step 6: Hyperparameter Optimization

Robyn runs thousands of model configurations using Nevergrad to tune adstock, saturation, and regularization parameters. Evaluation is based on NRMSE, Decomp.RSSD, and optionally MAPE.

Step 7: Model Selection

Models on the Pareto front (best trade-offs) are selected and clustered. Each cluster selects a representative winner used in simulation and reporting.

Step 8: Budget Simulations

With the final model, NeuroRadar can simulate multiple scenarios: maximizing response for a given spend, or minimizing spend for a target efficiency (ROAS/CPA).

A: NeuroRadar primarily focuses on providing advanced Marketing Mix Modeling (MMM) capabilities using open-source tools and machine learning, specifically built on Meta’s Robyn framework. This allows businesses to make data-driven media decisions across both online and offline channels with transparency, automation, and predictive power.

However, while Multi-Touch Attribution (MTA) models, which track and assign credit to individual advertising touchpoints, are not a core feature of NeuroRadar, MTA can be suported if the necessary data is provided