Nixtla Batch Forecaster

Process multiple time series forecasts in parallel with optimized throughput.

Overview

Leverages TimeGPT API to generate forecasts for many time series concurrently. Features parallel batch processing with rate limiting, automatic fallback for failed batches, and optional portfolio-level aggregation. Produces individual forecasts per series plus combined outputs.

Prerequisites

Required:

• Python 3.8+
• nixtla, pandas, tqdm packages

Environment Variables:

• NIXTLATIMEGPTAPI_KEY: Your TimeGPT API key

Installation:

pip install nixtla pandas tqdm

Instructions

Step 1: Prepare Input Data

Your CSV must have the Nixtla schema columns:

Analyze your data:

python {baseDir}/scripts/prepare_data.py your_data.csv

Step 2: Set API Key

export NIXTLA_TIMEGPT_API_KEY=your_api_key_here

Step 3: Run Batch Forecast

Execute the batch forecasting engine:

python {baseDir}/scripts/batch_forecast.py your_data.csv --horizon 14 --freq D

Available options:

• --horizon: Forecast horizon (default: 14)
• --freq: Frequency D/H/W/M (default: D)
• --batch-size: Series per batch (default: 20)
• --output-dir: Output directory (default: forecasts)
• --aggregate: Create portfolio aggregation
• --delay: Rate limit delay in seconds (default: 1.0)

Step 4: Generate Report

Create a summary report:

python {baseDir}/scripts/generate_report.py forecasts/

Output

• forecasts/all_forecasts.csv: Combined forecasts for all series
• forecasts/{series}_forecast.csv: Individual series forecasts
• forecasts/summary.json: Processing metadata
• forecasts/aggregated_forecast.csv: Portfolio aggregation (if --aggregate)
• forecasts/batch_report.md: Human-readable summary

Error Handling

1. Error: NIXTLATIMEGPTAPI_KEY not set

Solution: export NIXTLATIMEGPTAPIKEY=yourkey

1. Error: API Rate Limit Exceeded

Solution: Increase --delay or reduce --batch-size

1. Error: Missing required columns

Solution: Ensure CSV has unique_id, ds, y columns

1. Error: Batch failed, falling back to individual

Solution: Normal behavior - some series may have issues

Examples

Example 1: Forecast 50 Daily Contracts

python {baseDir}/scripts/batch_forecast.py contracts.csv \
    --horizon 14 \
    --freq D \
    --batch-size 10 \
    --output-dir forecasts/

Output:

Batch Forecast Complete
Series forecasted: 50/50
Success rate: 100.0%

Example 2: Hourly Portfolio with Aggregation

python {baseDir}/scripts/batch_forecast.py portfolio.csv \
    --horizon 24 \
    --freq H \
    --aggregate \
    --output-dir portfolio_forecasts/

Resources

• Scripts:

- {baseDir}/scripts/preparedata.py - Data validation and analysis - {baseDir}/scripts/batchforecast.py - Main forecasting engine - {baseDir}/scripts/generate_report.py - Report generation

• Nixtla Docs: https://nixtla.github.io/