Skip to main content
Before running this notebook, please visit our dashboard to obtain your TimeGPT api_key.

​
Why TimeGPT?

TimeGPT is a powerful, general-purpose time series forecasting solution. Throughout this notebook, we compare TimeGPT’s performance against three popular forecasting approaches:
  • Classical model (ARIMA)
  • Machine learning model (LightGBM)
  • Deep learning model (N-HiTS)
Below are three core benefits that our users value the most:

Accuracy

TimeGPT consistently outperforms traditional models by accurately capturing complex patterns.

Speed

Quickly generates forecasts with minimal training and tuning requirements per series.

Ease of Use

Minimal setup and no complex preprocessing make TimeGPT immediately accessible for use.

​
TimeGPT Advantage

TimeGPT delivers superior results with minimal effort compared to traditional approaches. In head-to-head testing against ARIMA, LightGBM, and N-HiTS models on M5 competition data, TimeGPT consistently achieves better accuracy metrics (lowest RMSE at 592.6 and SMAPE at 4.94%). Unlike other models which require:
  • Extensive preprocessing
  • Parameter tuning
  • Significant computational resources
TimeGPT provides powerful forecasting capabilities with a simple API interface, making advanced time series analysis accessible to users of all technical backgrounds.
1

1. Data Introduction

This notebook uses an aggregated subset from the M5 Forecasting Accuracy competition. The dataset:
  • Consists of 7 daily time series
  • Has 1,941 observations per series
  • Reserves the last 28 observations for evaluation on unseen data
Data Loading and Stats Preview
import os
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

from nixtla import NixtlaClient
from utilsforecast.plotting import plot_series
from utilsforecast.losses import mae, rmse, smape
from utilsforecast.evaluation import evaluate

nixtla_client = NixtlaClient(
    # api_key='my_api_key_provided_by_nixtla'
)

df = pd.read_csv(
    'https://datasets-nixtla.s3.amazonaws.com/demand_example.csv',
    parse_dates=['ds']
)

# Display aggregated statistics per time series
df.groupby('unique_id').agg({
    "ds": ["min", "max", "count"],
    "y": ["min", "mean", "median", "max"]
})
Below is a preview of the aggregated statistics for each of the 7 time series.
unique_idmin datemax datecountmin ymean ymedian ymax y
FOODS_12011-01-292016-05-2219410.02674.0862665.05493.0
FOODS_22011-01-292016-05-2219410.04015.9843894.09069.0
……………………
Next, we split our dataset into training and test sets. Here, we use data up to “2016-04-24” for training and the remaining data for testing.
Train-Test Split Example
df_train = df.query('ds <= "2016-04-24"')
df_test = df.query('ds > "2016-04-24"')

print(df_train.shape, df_test.shape)
# (13391, 3) (196, 3)
2

2. Model Fitting (TimeGPT, ARIMA, LightGBM, N-HiTS)

TimeGPT is compared against four different modeling approaches. Each approach forecasts the final 28 days of our dataset and we compare results across Root Mean Squared Error (RMSE) and Symmetric Mean Absolute Percentage Error (SMAPE).
TimeGPT offers a streamlined solution for time series forecasting with minimal setup.
TimeGPT Forecasting with NixtlaClient
fcst_timegpt = nixtla_client.forecast(
    df=df_train,
    target_col='y',
    h=28,
    model='timegpt-1-long-horizon',
    finetune_steps=10,
    level=[90]
)

evaluation_timegpt.groupby(['metric'])['TimeGPT'].mean()
# metric
# rmse     592.607378
# smape      0.049403
# Name: TimeGPT, dtype: float64
ARIMA is a common baseline for time series, though it often requires more data preprocessing and does not handle multiple series as efficiently.
ARIMA Forecasting Using StatsForecast
from statsforecast import StatsForecast
from statsforecast.models import AutoARIMA

sf = StatsForecast(models=[AutoARIMA()], freq='D')
fcst_arima = sf.forecast(h=28, df=df_train)
# Evaluation methods omitted here for brevity
LightGBM is a popular gradient-boosted tree approach. However, careful feature engineering is typically required for optimal results.
LightGBM Modeling with AutoMLForecast
import optuna
from mlforecast.auto import AutoMLForecast, AutoLightGBM

mlf = AutoMLForecast(models=[AutoLightGBM()], freq='D')
mlf.fit(df_train)
fcst_lgbm = mlf.predict(28)
# Evaluation methods omitted here for brevity
N-HiTS is a deep learning architecture for time series. While powerful, it often requires GPU resources and more hyperparameter tuning.
N-HiTS Deep Learning Forecast
from neuralforecast.core import NeuralForecast
from neuralforecast.models import NHITS

nf = NeuralForecast(models=[NHITS()], freq='D')
nf.fit(df=df_train)
fcst_nhits = nf.predict()
# Evaluation methods omitted here for brevity
3

3. Performance Comparison and Results

Below is a summary of the performance metrics (RMSE and SMAPE) on the test dataset. TimeGPT consistently delivers superior forecasting accuracy:
ModelRMSESMAPE
ARIMA724.95.50%
LightGBM687.85.14%
N-HiTS605.05.34%
TimeGPT592.64.94%
Performance Chart

Comparative Performance Visualization

Benchmarking Results

Benchmark Results

4

4. Conclusion

TimeGPT stands out with its accuracy, speed, and ease of use. Get started today by visiting the Nixtla dashboard to generate your api_key and access advanced forecasting with minimal overhead.
⌘I