​
Introduction

Forecasting intermittent demand (demand that contains many zero values) can be challenging, especially when performance and accuracy are paramount. In this tutorial, we demonstrate how to handle intermittent demand data using TimeGPT. We use a subset of the M5 dataset that captures food item sales in a California store, including exogenous variables like selling price and daily events.

TimeGPT delivers:

  • An impressive Mean Absolute Error (MAE) of 0.49, a 14% improvement over the best statistical model specifically designed for intermittent series.
  • Quick inference times (approximately 6.8 seconds for prediction), slightly slower (~1 second) than statistical models but highly accurate.

TimeGPT significantly enhances accuracy for intermittent demand forecasting, outperforming specialized statistical alternatives.

Open in Colab:

Key Features

Performance

​
Data Preview

Here’s a quick look at the data structure. The dataset tracks food item demand along with exogenous variables (e.g., selling price, event types):

unique_iddsysell_priceevent_type_Culturalevent_type_Nationalevent_type_Religiousevent_type_Sporting
FOODS_1_0012011-01-2932.00000
FOODS_1_0012011-01-3002.00000
FOODS_1_0012011-01-3102.00000
FOODS_1_0012011-02-0112.00000
FOODS_1_0012011-02-0242.00000

Figure 1: Visualization of intermittent demand data

1

Step 1: Environment Setup

You’ll need packages like pandas, numpy, and nixtla before starting.

Environment Setup
import time
import pandas as pd
import numpy as np

from nixtla import NixtlaClient
from utilsforecast.losses import mae
from utilsforecast.evaluation import evaluate

nixtla_client = NixtlaClient(api_key='my_api_key_provided_by_nixtla')

Using an Azure AI endpoint
To use an Azure AI endpoint, specify the base_url parameter:

Azure AI Endpoint Setup
nixtla_client = NixtlaClient(base_url="your azure ai endpoint", api_key="your api_key")
2

Step 2: Load the Dataset

Load Dataset
df = pd.read_csv("https://raw.githubusercontent.com/Nixtla/transfer-learning-time-series/main/datasets/m5_sales_exog_small.csv")
df['ds'] = pd.to_datetime(df['ds'])
df.head()

Confirm that the dataset columns match unique_id, ds, and y, plus any exogenous variables (e.g., sell_price and event details).

3

Step 3: Transform the Data

4

Step 4: Forecast with TimeGPT

Forecast Execution
start = time.time()

fcst_df = nixtla_client.forecast(
    df=input_df,
    h=28,
    level=[80],
    finetune_steps=10,
    finetune_loss='mae',
    model='timegpt-1-long-horizon',
    time_col='ds',
    target_col='y',
    id_col='unique_id'
)

timegpt_duration = time.time() - start
print(f"Time (TimeGPT): {timegpt_duration}")

# Inverse transform predictions
cols = [col for col in fcst_df.columns if col not in ['ds', 'unique_id']]
fcst_df[cols] = np.exp(fcst_df[cols]) - 1

Available Models in Azure AI
To use Azure AI directly, set model="azureai".
Public API models include timegpt-1 and timegpt-1-long-horizon.

5

Step 5: Evaluate the Forecasts

Forecast Evaluation
# Visualize the TimeGPT forecasts
nixtla_client.plot(
    test_df,
    fcst_df,
    models=['TimeGPT'],
    level=[80],
    time_col='ds',
    target_col='y'
)

# Compute MAE
test_df = pd.merge(test_df, fcst_df, how='left', on=['unique_id', 'ds'])
evaluation = evaluate(
    test_df, 
    metrics=[mae], 
    models=['TimeGPT'], 
    target_col='y', 
    id_col='unique_id'
)
average_metrics = evaluation.groupby('metric')['TimeGPT'].mean()
average_metrics

MAE for TimeGPT: approximately 0.49

6

Step 6: Compare with Statistical Models

Statistical Models Comparison
from statsforecast import StatsForecast
from statsforecast.models import CrostonClassic, CrostonOptimized, IMAPA, TSB

sf = StatsForecast(
    models=[CrostonClassic(), CrostonOptimized(), IMAPA(), TSB(0.1, 0.1)],
    freq='D',
    n_jobs=-1
)

start_stats = time.time()
sf.fit(df=input_df)
sf_preds = sf.predict(h=28)
stats_duration = time.time() - start_stats

# TimeGPT outperforms these specialized intermittent demand models by ~14%.
print(f"Time (Statistical models): {stats_duration}")

TimeGPT’s MAE is about 14% lower than the best-performing statistical model.

7

Step 7: Use Exogenous Variables

Exogenous Variables Forecast
# Include holiday/event data as exogenous features
futr_exog_df = test_df[['unique_id', 'ds', 'event_type_Cultural', 'event_type_National', 'event_type_Religious', 'event_type_Sporting']]

fcst_df_ex = nixtla_client.forecast(
    df=input_df,
    X_df=futr_exog_df,
    h=28,
    level=[80],
    finetune_steps=10,
    finetune_loss='mae',
    model='timegpt-1-long-horizon',
    time_col='ds',
    target_col='y',
    id_col='unique_id'
)

Incorporating exogenous event data can further refine predictions.

​
Final Results

TimeGPT provides a robust solution for forecasting intermittent demand:
• ~14% MAE improvement over specialized models
• Supports exogenous features for enhanced accuracy

By leveraging TimeGPT and combining both internal series patterns and external factors, organizations can achieve more reliable forecasts even for challenging intermittent demands.