Applied modeling is the core of the Meridian workflow, where you configure, run, and interpret the model to gain insights into your marketing performance. This section provides a comprehensive guide to the components of the Meridian model, from understanding the input data and core concepts to advanced customization with priors and calibration.
The Meridian Model
This section covers the fundamental components of the Meridian model, from the data it requires to the core mathematical structure.
Page | Description |
---|---|
Input Data | This page specifies the format and structure for the data you provide to Meridian. It details the different kinds of input data, like your main KPI, media activity, and control variables, and explains how the model internally transforms this data for analysis. |
Control Variables | This page explains control variables, which are factors that can influence your results but aren't part of your marketing treatments. It covers how to choose the right ones to ensure your model's estimates are accurate and unbiased. |
Organic Media and Non-Media Treatment Variables | This page explains how to classify treatment variables that don't have a direct cost. It helps you decide whether to classify an activity as "organic media" (like a blog post) or a "non-media treatment" (like a price change) for accurate modeling. |
Paid Search Modeling | Discover the best practices for modeling paid search campaigns. This guide highlights the importance of using Google Query Volume as a control variable to get an unbiased understanding of your search ads' true impact. |
Holdout Observations (Train and Test Split) | This guide covers the practice of holding out a portion of your data to test the model's performance. It explains how this helps in comparing different model versions and provides recommendations on the best way to select this holdout data for reliable results. |
The Meridian Model Specification | This page provides a look at the core mathematical equation behind the Meridian model. It breaks down the main components, including how it handles different geographies, adjusts for trends over time, and models the non-linear effects of media. |
Reach and Frequency | Go beyond impression counts by modeling with reach and frequency. This guide explains how this approach can provide a more nuanced view of your media's effectiveness by accounting for how many people you reach and how often they see your ads. |
Media Saturation and Lagging | Understand the two fundamental concepts of marketing effects. This guide explains "lagging effects" (Adstock), where a campaign's impact continues over time, and "saturation" (Hill function), which accounts for the diminishing returns as you spend more. |
Time Based Parameters
Learn how to effectively model time effects, including trend and seasonality, within Meridian.
Page | Description |
---|---|
Setting Knots | Understand how Meridian models time effects using "knots." This guide explains what knots are, how they create a flexible baseline for your model, and provides advice on choosing the right number of knots for your data, and describes Automatic Knot Selection in Meridian. |
Set the max_lag Parameter |
This page explains the max_lag setting, which controls how long the model assumes an ad's impact will last.It discusses the trade-offs involved in choosing a longer or shorter lag period and how it affects your model. |
Set the adstock_decay_spec Parameter |
Fine-tune how your model understands the lagging effects of media. This page describes the different decay shapes (geometric and binomial) you can choose for the Adstock function and provides advice on which to use based on your marketing channels. |
Geo versus National level modeling
Learn how National and Geo level modeling work in Meridian
Page | Description |
---|---|
National-Level Modeling | Learn how Meridian adapts when you only have national-level data instead of data broken down by region. This page explains that it's treated as a special case of the geo-level model and details the automatic adjustments made. |
Priors
Priors are a key feature of Bayesian modeling that allow you to incorporate existing knowledge into the model. This section explains how to use them effectively.
Treatment Prior Types
Meridian offers several ways to define priors, each with different implications for the model.
Page | Description |
---|---|
ROI, mROI, and Contribution Parameterizations | This page dives into the mathematical details of how Meridian re-parameterizes the model to let you set priors directly on business metrics like ROI, mROI, or contribution. |
mROI Priors and Comparison to ROI Priors | This guide compares two advanced prior options: ROI (overall return) and mROI (return on the next dollar spent). It explains the differences and helps you decide which is more suitable, especially when your goal is budget optimization. |
How to Choose Treatment Prior Types | Learn how to select the best type of prior for each of your marketing activities. This page provides guidance on whether to set a prior based on ROI, mROI, contribution, or the underlying model coefficient, helping you match the model's setup to your business knowledge. |
Default Priors
Understand the out-of-the-box assumptions Meridian makes to get you started.
Page | Description |
---|---|
Default Prior Parameterizations | Meridian can set priors on different metrics like ROI, contribution, or a raw coefficient. This page explains which of these options Meridian uses by default for various marketing types (paid, organic, etc.) and in different situations. |
Default Prior Distributions | This page details the default "prior distributions" that Meridian uses out-of-the-box. Priors are the initial assumptions the model makes about its parameters before it analyzes your data. This guide shows you the statistical shapes of these default assumptions. |
When the KPI is not Revenue (Default) | This page explains Meridian's default priors when the key performance indicator (KPI) is not revenue. It describes the concept of a "total paid media contribution prior," which helps anchor the model when a direct ROI is not applicable. |
Custom Priors
Learn how to tailor the model to your specific business context by setting custom priors based on experiments, benchmarks, or other domain knowledge.
Page | Description |
---|---|
ROI Priors and Calibration | Learn about "calibration," the process of incorporating existing knowledge into your model using ROI priors. This guide explains how you can use results from past experiments or industry benchmarks to make your model's outputs more accurate and reliable. |
Set Custom ROI Priors Using Past Experiments | This guide provides practical steps for translating the results of past A/B tests or incrementality studies into custom ROI priors for your model. It covers important considerations to ensure the translation is meaningful. |
When the KPI is not Revenue (Custom Priors) | When your main business goal isn't measured in revenue (like user sign-ups), setting priors can be tricky. This page offers several strategies for creating custom priors that are meaningful for non-revenue KPIs. |
Set Custom Priors from a Combination of Distribution Families | This guide shows an advanced technique for creating custom priors. It explains how you can combine different statistical distributions to more precisely reflect your prior beliefs about how a marketing channel should perform. |