Generative AI for forecasting is powered by transformer architectures which are neural networks designed to model long-range dependencies in sequences. A time-series transformer is pre-trained on massive, diverse datasets, learning a general “language of time series” (recurring structures, seasonality, and signal vs. noise). Because this knowledge is broadly applicable, the foundation model can forecast new, unseen series immediately often with strong accuracy without task-specific training. This is zero‑shot forecasting: you bring data, the model brings general knowledge.

When you need domain‑specific nuance, you can fine‑tune the foundation model on your own historical data. Fine‑tuning adapts the pre‑trained knowledge to your context, typically improving accuracy and stability for your use cases while keeping training cost far lower than training from scratch.

TimeGPT

TimeGPT is once such transformer based foundation model for time series developed by Nixtla. It delivers:

  • Zero‑shot forecasts with prediction intervals.
  • Support for multiple series.
  • Optional use of both historical and future exogenous variables.
  • A clear path to fine‑tuning when deeper domain adaptation is required.

Authentication and Setup

You can setup TimeGPT via either of the following:

Azure-hosted TimeGPT

Set the Azure base URL and API key.

Option A: Environment variables

# In R (or add to .Renviron)
Sys.setenv(NIXTLA_BASE_URL = "https://your-azure-deployed-timegen.azure.com/")
Sys.setenv(NIXTLA_API_KEY = "your_api_key_here")
# Note that the current version of nixtlar requires base url to end with "/"

Option B: Using nixtlar helper

# devtools::install_github("Nixtla/nixtlar")
library(nixtlar)
nixtla_client_setup(
  base_url = "Base URL here",
  api_key = "API key here"
)

Nixtla API

If you use API directly from Nixtla, base url is not required.

Option A: Environment variable

Sys.setenv(NIXTLA_API_KEY = "your_api_key_here")

Option B: Using nixtlar helper

# devtools::install_github("Nixtla/nixtlar")
library(nixtlar)
nixtla_set_api_key(api_key = "Your API key here")

Quick start with nixtlar (standalone)

# devtools::install_github("Nixtla/nixtlar")
library(nixtlar)

df <- nixtlar::electricity

# Forecast next 8 steps
fcst <- nixtla_client_forecast(
  df,
  h = 8,
  level = c(80, 95),
  # if using azure deployed timegen
  # model = "azureai"
)

head(fcst)

Using TimeGPT within finnts

TimeGPT is available as a model within finnts unified workflow and can run as both a local (per time series) or global (across all time series) model. When running as a global model, TimeGPT learns patterns across all your time series simultaneously.

library(finnts)

# setup timegpt api keys

# Checkout these data requirements provided by nixtla
# https://www.nixtla.io/docs/data_requirements/data_requirements
hist_data <- timetk::m4_monthly %>%
  dplyr::filter(
    date >= "2010-01-01",
    id == "M2"
  ) %>%
  dplyr::rename(Date = date) %>%
  dplyr::mutate(id = as.character(id))

run_info <- set_run_info(
  project_name = "finnts_fcst",
  run_name = "finn_sub_component_run"
)

prep_data(
  run_info = run_info,
  input_data = hist_data,
  combo_variables = c("id"),
  target_variable = "value",
  date_type = "month",
  forecast_horizon = 6
)

prep_models(
  run_info = run_info,
  models_to_run = c("timegpt"),
)

train_models(
  run_info = run_info,
  run_global_models = FALSE
)

final_models(run_info = run_info)

finn_output_tbl <- get_forecast_data(run_info = run_info)
head(finn_output_tbl)

# A tibble: 6 x 17
#  Combo id    Model_ID      Model_Name Model_Type Recipe_ID Run_Type Train_Test_ID Best_Model Horizon
#  <chr> <chr> <chr>         <chr>      <chr>      <chr>     <chr>            <dbl> <chr>        <dbl>
# 1 M2    M2    timegpt--loc~ timegpt    local      R1        Future_~             1 Yes              1
# 2 M2    M2    timegpt--loc~ timegpt    local      R1        Future_~             1 Yes              2
# 3 M2    M2    timegpt--loc~ timegpt    local      R1        Future_~             1 Yes              3
# 4 M2    M2    timegpt--loc~ timegpt    local      R1        Future_~             1 Yes              4
# 5 M2    M2    timegpt--loc~ timegpt    local      R1        Future_~             1 Yes              5
# 6 M2    M2    timegpt--loc~ timegpt    local      R1        Future_~             1 Yes              6
# i 7 more variables: Date <date>, Target <dbl>, Forecast <dbl>, lo_95 <dbl>, lo_80 <dbl>,
#   hi_80 <dbl>, hi_95 <dbl>

Data Size Requirements and Automatic Padding

Azure AI TimeGPT endpoints have minimum data size requirements based on data frequency:

  • Daily data: Minimum 300 rows
  • Weekly data: Minimum 64 rows
  • Monthly/Quarterly/Yearly data: Minimum 48 rows

When using Azure AI endpoints, finnts automatically pads your data with zeros (backward from the earliest date) to meet these requirements. This ensures your forecasts work even with smaller datasets. The padding only affects the data sent to the API, your original training data remains unchanged in the model fit object.

Note: Padding is only applied for Azure AI endpoints. The default Nixtla API does not have these constraints.

Long-Horizon Forecasting

For forecasts exceeding two seasonal periods, TimeGPT automatically uses the timegpt-1-long-horizon model, which is optimized for longer forecast horizons. The threshold for “long horizon” depends on your data frequency:

  • Daily: > 14 days (2 weeks)
  • Weekly: > 104 weeks (2 years)
  • Monthly: > 24 months (2 years)
  • Quarterly: > 8 quarters (2 years)
  • Yearly: > 2 years

This selection happens automatically based on your forecast_horizon and date_type parameters—no additional configuration needed.

Hyperparameter Tuning

TimeGPT supports hyperparameter tuning for fine-tuning parameters:

  • finetune_steps: Number of fine-tuning steps (range: 0-200)
  • finetune_depth: Fine-tuning depth/layers (range: 1-5)

These parameters are automatically tuned when you set num_hyperparameters in prep_models(). The tuning process uses validation splits to find optimal values, then refits the model with the best parameters.

prep_models(
  run_info = run_info,
  models_to_run = c("timegpt"),
  num_hyperparameters = 4 # Will tune finetune_steps and finetune_depth
)

Global Model Support

TimeGPT can run as both a local (per time series) or global (across all time series) model. When running as a global model, TimeGPT learns patterns across all your time series simultaneously.

# Run TimeGPT as a global model
train_models(
  run_info = run_info,
  run_global_models = TRUE # TimeGPT will train on all combos together
)