Nixtla Schema Mapper

Transform data sources to Nixtla-compatible schema (unique_id, ds, y).

Overview

This skill automates data transformation:

• Column inference: Detects timestamp, target, and ID columns
• Code generation: Python modules for CSV/SQL/Parquet/dbt
• Schema contracts: Documentation with validation rules
• Quality checks: Validates transformed data

Prerequisites

Required:

• Python 3.8+
• pandas

Optional:

• pyarrow: For Parquet support
• sqlalchemy: For SQL sources
• dbt-core: For dbt models

Installation:

pip install pandas pyarrow sqlalchemy

Instructions

Step 1: Identify Data Source

Supported formats:

• CSV/Parquet files
• SQL tables or queries
• dbt models

Step 2: Analyze Schema

python {baseDir}/scripts/analyze_schema.py --input data/sales.csv

Output:

Detected columns:
  Timestamp: 'date' (datetime64)
  Target: 'sales' (float64)
  Series ID: 'store_id' (object)
  Exogenous: price, promotion

Step 3: Generate Transformation

python {baseDir}/scripts/generate_transform.py \
    --input data/sales.csv \
    --id_col store_id \
    --date_col date \
    --target_col sales \
    --output data/transform/to_nixtla_schema.py

Step 4: Create Schema Contract

python {baseDir}/scripts/create_contract.py \
    --mapping mapping.json \
    --output NIXTLA_SCHEMA_CONTRACT.md

Step 5: Validate Transformation

python data/transform/to_nixtla_schema.py

Output

• data/transform/to_nixtla_schema.py: Transformation module
• NIXTLA_SCHEMA_CONTRACT.md: Schema documentation
• nixtla_data.csv: Transformed data (optional)

Error Handling

1. Error: No timestamp column detected Solution: Specify manually with --date_col

2. Error: Multiple target candidates Solution: Specify manually with --target_col

3. Error: Date parsing failed Solution: Specify format with --date_format "%Y-%m-%d"

4. Error: Non-numeric target column Solution: Check for string values, use pd.to_numeric(errors='coerce')

Examples

Example 1: CSV Transformation

python {baseDir}/scripts/generate_transform.py \
    --input sales.csv \
    --id_col product_id \
    --date_col timestamp \
    --target_col revenue

Generated code:

def to_nixtla_schema(path="sales.csv"):
    df = pd.read_csv(path)
    df = df.rename(columns={
        'product_id': 'unique_id',
        'timestamp': 'ds',
        'revenue': 'y'
    })
    df['ds'] = pd.to_datetime(df['ds'])
    return df[['unique_id', 'ds', 'y']]

Example 2: SQL Source

python {baseDir}/scripts/generate_transform.py \
    --sql "SELECT * FROM daily_sales" \
    --connection postgresql://localhost/db \
    --id_col store_id \
    --date_col sale_date \
    --target_col amount

Resources

• Scripts: {baseDir}/scripts/
• Templates: {baseDir}/assets/templates/
• Nixtla Schema Docs: https://nixtla.github.io/statsforecast/

Related Skills:

• nixtla-timegpt-lab: Use transformed data for forecasting
• nixtla-experiment-architect: Reference in experiments