Nixtla TimeGPT Fine-Tuning Lab

Adapts the TimeGPT model to specific datasets for enhanced forecasting performance.

Purpose

Improves forecasting accuracy by fine-tuning the pre-trained TimeGPT model on custom time series data.

Overview

This skill guides users through the process of fine-tuning TimeGPT on their own datasets. It handles data preprocessing, training configuration, and evaluation. Use when domain-specific accuracy is crucial or when the general TimeGPT model underperforms. It outputs a fine-tuned model and performance metrics.

Prerequisites

Tools: Read, Write, Bash, Glob, Grep

Environment: NIXTLA_TIMEGPT_API_KEY

Packages:

pip install nixtla pandas scikit-learn matplotlib

Instructions

Step 1: Prepare data

Load and preprocess time series data into Nixtla format (unique_id, ds, y).

python {baseDir}/scripts/prepare_data.py \
  --input data.csv \
  --train_output train_data.csv \
  --val_output val_data.csv

Requirements:

• CSV file with columns: unique_id, ds, y
• ds column in recognizable datetime format
• y column with numeric values
• No missing values (will be dropped automatically)

Output:

• train_data.csv (80% split)
• val_data.csv (20% split)
• time_series_visualization.png (first 3 series)

Step 2: Configure training

Create or validate training configuration.

# Create default config
python {baseDir}/scripts/configure_training.py \
  --create_default \
  --output config.json

# Or validate existing config
python {baseDir}/scripts/configure_training.py \
  --config config.json

Configuration parameters:

• learning_rate: 0.001 (default)
• epochs: 10 (default)
• batch_size: 32 (default)
• num_workers: 4 (default)
• model_name: TimeGPT
• freq: D (daily), H (hourly), etc.

Step 3: Execute fine-tuning

Run the fine-tuning process using prepared data and configuration.

export NIXTLA_TIMEGPT_API_KEY=your_api_key

python {baseDir}/scripts/finetune_model.py \
  --train_data train_data.csv \
  --config config.json \
  --output finetuned_model.pkl

Output: finetuned_model.pkl (serialized fine-tuned model)

Step 4: Evaluate and save

Evaluate the fine-tuned model on validation data and save metrics.

python {baseDir}/scripts/evaluate_model.py \
  --val_data val_data.csv \
  --model finetuned_model.pkl \
  --config config.json \
  --output metrics.json

Output: metrics.json (MAE, RMSE metrics)

Output

• train_data.csv: Training dataset (80% split)
• val_data.csv: Validation dataset (20% split)
• time_series_visualization.png: Visualization of training data
• config.json: Training configuration
• finetuned_model.pkl: Fine-tuned TimeGPT model
• metrics.json: Evaluation metrics (MAE, RMSE)

Error Handling

1. Error: NIXTLA_TIMEGPT_API_KEY not set Solution: export NIXTLA_TIMEGPT_API_KEY=your_api_key

2. Error: Invalid data format Solution: Ensure data has columns: unique_id, ds, y

3. Error: Insufficient training data Solution: Provide a larger training dataset (at least 50 time series points per series)

4. Error: Fine-tuning failed Solution: Adjust the fine-tuning parameters (learning rate, epochs) in config.json

5. Error: Missing required column Solution: Verify CSV has unique_id, ds, y columns with correct data types

Examples

Example 1: Fine-tuning on Retail Sales Data

Input:

unique_id,ds,y
store_1,2023-01-01,100
store_1,2023-01-02,110

Command:

python {baseDir}/scripts/finetune_model.py \
  --train_data train_data.csv \
  --config config.json \
  --output finetuned_model.pkl

Output: finetuned_model.pkl (a serialized, fine-tuned TimeGPT model)

Example 2: Fine-tuning on Energy Consumption Data

Input:

unique_id,ds,y
building_1,2023-01-01 00:00,50
building_1,2023-01-01 01:00,55

Command:

python {baseDir}/scripts/evaluate_model.py \
  --val_data val_data.csv \
  --model finetuned_model.pkl \
  --config config.json \
  --output metrics.json

Output: metrics.json (performance metrics of the fine-tuned model)

Resources

• Data preparation: {baseDir}/scripts/prepare_data.py
• Training configuration: {baseDir}/scripts/configure_training.py
• Fine-tuning execution: {baseDir}/scripts/finetune_model.py
• Model evaluation: {baseDir}/scripts/evaluate_model.py