Meridian MMM Model Building

Meridian Model Building Skill

This skill guides the user through the process of creating a Meridian model, accumulating the code into a Python script.

Core Workflow

Interactivity Checkpoint Rule

Throughout this workflow, you will encounter CRITICAL INTERACTIVE CHECKPOINTs. At each checkpoint, you MUST:

1. Present the current proposed configurations, parameters, mappings, script path, or status to the user for approval.
2. Ask the user if they are ready to proceed using the available user-interaction tool (e.g., ask_question), structured as a multiple-choice question. Do NOT use raw chat text.
3. Wait for their response before proceeding.
   • MANDATORY: You MUST pause at every checkpoint regardless of the initial prompt instructions (even if the user request contains phrases like "run autonomously", "execute directly", "fix autonomously", etc.). The initial request does NOT bypass these interactive checkpoints.
   • Note: If the user replies to a checkpoint with a generic approval (e.g., "proceed", "do what you think is best"), proceed with the proposed defaults.

---

1. Initial Setup

• Prompt the user for the input CSV file path, the desired path for the generated Python script, the EDA HTML report output path, and the saved model path (meridian_model.binpb by default). If the user does not specify output paths, default to model_build/ in the active project directory (or relative to the input data directory) for the script and all outputs (meridian_model.binpb, eda.html).
• CRITICAL INTERACTIVE CHECKPOINT: Present the gathered paths to the user and obtain confirmation before proceeding to data loading.

2. Add Data Loading & Column Mapping Code

• Target Module: meridian.data.data_frame_input_data_builder
• Action:
  • Check CSV Format: Before loading data, verify if the CSV data is in the right format. Consult the meridian-doc-consultant skill or check the documentation map in skills/meridian_doc_consultant/references/documentation_map.md under "Data Preparation & Loading" to find specific guides (like load-geo-data-without-rf.md, load-geo-data-with-organic-and-non-media.md based on the columns observed in the data) to understand the expected columns and data types. Consult references/csv_format_reference.md for details on expected row/column structure and data quality guardrails. If the format is incorrect or missing required columns, attempt to autonomously convert the dataset to the expected format for the user (e.g., renaming columns, restructuring) unless you are uncertain and need user input.
  • Read the header row of the provided CSV using Python to get the column names.
  • Propose heuristic mappings based on column keywords (e.g., 'sales' -> kpi_col, 'spend' -> media_spend_cols) and infer the kpi_type ('revenue' or 'non_revenue') based on the columns (e.g., 'revenue' or 'sales' implying 'revenue', and 'conversions' or 'leads' implying 'non_revenue').
  • Robust Mapping: If the user prompt specifies mapping a column name that does not exist in the CSV, do not assume it is a literal name if it looks like a description (e.g., 'media_impressions' vs 'ChannelX_impression'). Use heuristics to find matching columns and proceed.
  • Present the proposed mapping to the user.
  • CRITICAL INTERACTIVE CHECKPOINT: Present the proposed column mappings to the user and obtain approval before continuing to model configuration.
  • Accumulate the data loading code using meridian.data.data_frame_input_data_builder.DataFrameInputDataBuilder and its with_* methods (e.g. with_kpi, with_media). See data_builder_template.md.

3. Add Model Configuration Code

• Target Modules: meridian.model.spec, meridian.model.model
• Action:
  • Read the ModelSpec and PriorDistribution definitions in meridian.model.spec.
  • Guide the user through configuration, prompting for relevant values while explaining their purpose based on the source code docstrings.
  • CRITICAL INTERACTIVE CHECKPOINT: Present the proposed model specification parameters to the user and obtain approval before continuing.
  • Accumulate the code to initialize meridian.model.spec.ModelSpec and meridian.model.model.Meridian. See model_spec_template.md.
  • Accumulate code: mmm.sample_prior()

4. Add Exploratory Data Analysis (EDA) Code

• Target Module: meridian.model.eda.meridian_eda
• Action:
  • Read meridian_eda.py or module docstrings to confirm the generate_and_save_report method.
  • Accumulate code to initialize meridian_eda.MeridianEDA and call generate_and_save_report(filepath) using the user's specified path.
  • CRITICAL INTERACTIVE CHECKPOINT: Present the EDA output path configuration and obtain approval before proceeding to the model fitting step.

5. Add Model Fitting Code

• Target Module: meridian.model.model
• Action:
  • Read the sample_posterior method in meridian.model.model to understand its parameters.
  • Prompt the user for MCMC parameters: n_chains, n_adapt, n_burnin, n_keep.
  • CRITICAL INTERACTIVE CHECKPOINT: Present the MCMC parameters to the user and obtain approval before proceeding to compile the model fitting code.
  • Accumulate code: mmm.sample_posterior(...)