| name | nixtla-model-selector |
| description | Automatically selects the best forecasting model between StatsForecast and TimeGPT based on time series data characteristics. Use when unsure which model performs best. Trigger with "auto-select model", "choose best model", "model selection".
|
| version | 1.0.0 |
| allowed-tools | Read,Write,Bash,Glob,Grep |
Nixtla Model Selector
Automatically selects and executes the optimal forecasting model for time series data.
Overview
This skill analyzes time series data characteristics to determine whether StatsForecast or TimeGPT will yield more accurate forecasts. It evaluates data length, frequency, seasonality, missing values, and series count to make an intelligent selection. The skill then executes the chosen model and returns forecasts with selection rationale. This eliminates manual model selection and experimentation cycles.
Decision Logic:
- StatsForecast: Selected for missing values, short data (<30 points), seasonal patterns, or many series (>100)
- TimeGPT: Selected for long, non-seasonal data with complete observations
The skill outputs forecast predictions and a model selection report explaining the decision.
Prerequisites
Tools: Read, Write, Bash, Glob, Grep
Environment: NIXTLA_TIMEGPT_API_KEY (required if TimeGPT is selected)
Packages:
pip install statsforecast nixtla pandas matplotlib statsmodels
Input Format: CSV file with columns unique_id, ds (datetime), y (target values)
Instructions
Step 1: Prepare Data
Claude reads the input CSV file and validates the required schema. The data must contain unique_id (series identifier), ds (timestamp), and y (observations). Claude converts timestamps to datetime format and infers the frequency (daily, hourly, etc.).
Execute using: {baseDir}/scripts/model_selector.py --input data.csv --visualize
The script performs basic exploratory analysis including data shape, date range, unique series count, and optional visualization plots.
Step 2: Analyze Data Characteristics
Claude analyzes the time series to extract decision criteria:
- Data length: Number of observations per series
- Missing values: Presence of null values in target variable
- Seasonality: Detects seasonal patterns using decomposition
- Series count: Number of unique time series in the dataset
The analysis uses statistical decomposition to identify seasonal components and compares seasonal variance against total variance using a heuristic threshold.
Step 3: Select Model
Based on the analysis, Claude applies decision rules:
- Missing values detected → StatsForecast (handles missing data robustly)
- Data length < 30 → StatsForecast (insufficient data for TimeGPT)
- Seasonality present → StatsForecast (specialized seasonal models)
- Series count > 100 → StatsForecast (efficient batch processing)
- Otherwise → TimeGPT (optimal for long, complex patterns)
The script logs the decision rationale including specific metrics that triggered the selection.
Execute using the same command: {baseDir}/scripts/model_selector.py --input data.csv
Step 4: Execute Forecast
Claude runs the selected model with appropriate configuration:
StatsForecast execution:
- Models: AutoETS, AutoARIMA
- Parallel processing enabled (n_jobs=-1)
- Returns forecasts for each model variant
TimeGPT execution:
- Requires NIXTLA_TIMEGPT_API_KEY environment variable
- Calls cloud API with inferred frequency
- Returns single forecast column
Default horizon: 14 periods (customizable via --horizon flag)
Step 5: Generate Output
Claude saves two output files:
- forecast.csv: Predictions with columns
unique_id, ds, model, yhat
- model_selection.txt: Selection report with model name, reason, record count
The forecast CSV uses long format for consistency across model types. StatsForecast produces multiple model columns (AutoETS, AutoARIMA) while TimeGPT produces a single forecast.
Execute complete workflow:
{baseDir}/scripts/model_selector.py --input data.csv --output forecast.csv --horizon 30
Output
forecast.csv: Time series predictions generated by the selected model
- Columns:
unique_id, ds, model, yhat
- Format: Long format with one row per forecast point
model_selection.txt: Model selection report
- Selected model name (StatsForecast or TimeGPT)
- Selection reason with specific data characteristics
- Forecast record count
time_series_plot.png (optional): Visualization of input data when --visualize flag is used
Error Handling
Input file not found
- Cause: Invalid file path or file does not exist
- Solution: Verify file path and check file permissions
Invalid data format
- Cause: Missing required columns (unique_id, ds, y)
- Solution: Ensure CSV contains all required columns with correct names
NIXTLA_TIMEGPT_API_KEY not set
- Cause: TimeGPT selected but API key not configured
- Solution: Set environment variable before execution:
export NIXTLA_TIMEGPT_API_KEY="your_key"
Could not infer frequency
- Cause: Irregular or missing timestamps in data
- Solution: Ensure
ds column has consistent datetime intervals
Seasonality check failed
- Cause: Insufficient data for seasonal decomposition
- Solution: Provide at least 24 observations; script falls back to non-seasonal analysis
Examples
Example 1: Short Seasonal Data
Input (7 observations):
unique_id,ds,y
product_1,2023-01-01,10
product_1,2023-01-02,12
product_1,2023-01-03,15
product_1,2023-01-04,13
product_1,2023-01-05,16
product_1,2023-01-06,18
product_1,2023-01-07,20
Command: {baseDir}/scripts/model_selector.py --input short_data.csv
Output:
- model_selection.txt: "StatsForecast selected due to short data length (<30 observations)."
- forecast.csv: 14-period forecasts from AutoETS and AutoARIMA
Example 2: Long Non-Seasonal Data
Input (365+ observations):
unique_id,ds,y
location_1,2020-01-01,100
location_1,2020-01-02,102
location_1,2020-01-03,105
... (365+ rows)
Command: {baseDir}/scripts/model_selector.py --input long_data.csv --horizon 30 --visualize
Output:
- model_selection.txt: "TimeGPT selected due to long data length and lack of clear seasonality."
- forecast.csv: 30-period TimeGPT forecasts
- time_series_plot.png: Input data visualization
Resources
Script: {baseDir}/scripts/model_selector.py
StatsForecast Documentation: https://nixtla.github.io/statsforecast/
TimeGPT API Reference: https://docs.nixtla.io/
CSV Schema Requirements:
unique_id: Series identifier (string or integer)ds: Timestamp (datetime parseable string)y: Observation values (numeric)
