nixtla-experiment-architect

intent-solutions-io's avatarfrom intent-solutions-io

Generate production-ready forecasting experiments with StatsForecast and TimeGPT. Use when setting up model benchmarking or cross-validation. Trigger with 'scaffold experiment' or 'compare models'.

⭐0stars🔀0forks📁View on GitHub🕐Updated Jan 10, 2026
[Experiment Analysis]

When & Why to Use This Skill

The Nixtla Experiment Architect is a specialized Claude skill designed to automate the creation of production-ready time-series forecasting experiments. By leveraging industry-leading libraries like StatsForecast and TimeGPT, it streamlines the process of model benchmarking, cross-validation, and performance evaluation. This tool eliminates manual boilerplate by scaffolding complete experiment harnesses, allowing data scientists and developers to rapidly transition from raw data to rigorous model comparison and deployment-ready forecasting pipelines.

Use Cases

  • Model Benchmarking: Automatically compare the accuracy of traditional statistical models (like AutoARIMA and AutoETS) against modern foundation models (TimeGPT) to determine the best fit for specific business data.
  • Automated Cross-Validation: Quickly set up rolling-origin or expanding-window cross-validation schemes to ensure forecasting robustness and prevent overfitting in temporal datasets.
  • Standardized Experiment Scaffolding: Generate consistent, YAML-based configuration files and runnable Python harnesses for retail sales, energy demand, or financial market forecasting to ensure reproducibility across teams.
  • Rapid Prototyping: Use the 'scaffold experiment' trigger to instantly create a full directory structure and script set for new forecasting projects, reducing setup time from hours to seconds.
namenixtla-experiment-architect
description"Generate production-ready forecasting experiments with StatsForecast and TimeGPT. Use when setting up model benchmarking or cross-validation. Trigger with 'scaffold experiment' or 'compare models'."
allowed-tools"Read,Write,Glob,Grep,Edit"
version"1.0.0"
author"Jeremy Longshore <jeremy@intentsolutions.io>"
licenseMIT

Nixtla Experiment Architect

Design and scaffold complete forecasting experiments using Nixtla's libraries.

Overview

This skill creates production-ready experiment harnesses:

  • Configuration management: YAML-based experiment config
  • Multi-model comparison: StatsForecast + MLForecast + TimeGPT
  • Cross-validation: Rolling-origin or expanding-window
  • Metrics evaluation: SMAPE, MASE, MAE, RMSE

Prerequisites

Required:

  • Python 3.8+
  • statsforecast, utilsforecast

Optional:

  • mlforecast: For ML models
  • nixtla: For TimeGPT
  • NIXTLA_API_KEY: TimeGPT access

Installation:

pip install statsforecast mlforecast nixtla utilsforecast pyyaml

Instructions

Step 1: Gather Requirements

Collect experiment parameters:

  • Data source path
  • Target column name
  • Forecast horizon (e.g., 14 days)
  • Frequency (D, H, W, M)
  • Unique ID column (optional)

Step 2: Generate Configuration

python {baseDir}/scripts/generate_config.py \
    --data data/sales.csv \
    --target sales \
    --horizon 14 \
    --freq D \
    --output forecasting/config.yml

Step 3: Scaffold Experiment

python {baseDir}/scripts/scaffold_experiment.py \
    --config forecasting/config.yml \
    --output forecasting/experiments.py

Step 4: Run Experiment

python forecasting/experiments.py

Step 5: Review Results

cat forecasting/results/metrics_summary.csv

Output

  • forecasting/config.yml: Experiment configuration
  • forecasting/experiments.py: Runnable experiment harness
  • forecasting/results/: Metrics and forecasts (after running)

Error Handling

  1. Error: Data file not found Solution: Verify data source path in config

  2. Error: Column not found Solution: Check column names match your data

  3. Error: Missing required package Solution: Install missing dependencies with pip

  4. Error: Cross-validation failed Solution: Ensure enough data for n_windows

Examples

Example 1: Daily Sales Forecast

python {baseDir}/scripts/generate_config.py \
    --data data/sales.csv \
    --target revenue \
    --horizon 30 \
    --freq D \
    --id_col store_id

Output config.yml:

data:
  source: data/sales.csv
  target: revenue
  unique_id: store_id
forecasting:
  horizon: 30
  freq: D
models:
  - SeasonalNaive
  - AutoETS
  - AutoARIMA

Example 2: Hourly Energy Forecast

python {baseDir}/scripts/generate_config.py \
    --data data/energy.csv \
    --target consumption \
    --horizon 24 \
    --freq H

Resources

Related Skills:

  • nixtla-timegpt-lab: Core forecasting guidance
  • nixtla-schema-mapper: Data transformation
  • nixtla-prod-pipeline-generator: Production deployment