autogluon-timeseries-strategy

AutoGluon TimeSeriesPredictor Strategy

The Critical Difference

TabularPredictor ≠ TimeSeriesPredictor. They are separate APIs with different:

• Data formats (TimeSeriesDataFrame vs DataFrame)
• Cross-validation (multi-window backtesting vs k-fold)
• Evaluation metrics (RMSLE/MAE/MAPE vs RMSE/accuracy)
• Models (DeepAR/TFT/Chronos vs LightGBM/XGBoost)

# Tabular (independent rows):
from autogluon.tabular import TabularPredictor
predictor = TabularPredictor(label='target').fit(df)

# Time series (sequential, grouped by entity):
from autogluon.timeseries import TimeSeriesPredictor, TimeSeriesDataFrame
predictor = TimeSeriesPredictor(target='target', prediction_length=16, freq='D').fit(ts_df)

Data Format (Critical)

TimeSeriesPredictor requires TimeSeriesDataFrame:

from autogluon.timeseries import TimeSeriesDataFrame
import pandas as pd

# Required columns: item_id, timestamp, target
# Each item_id is ONE time series (e.g., one store-product combination)
df = pd.DataFrame({
    'item_id': ['store_1', 'store_1', 'store_1', 'store_2', 'store_2'],
    'timestamp': pd.to_datetime(['2024-01-01', '2024-01-02', '2024-01-03',
                                  '2024-01-01', '2024-01-02']),
    'target': [100, 110, 105, 50, 55],
    'onpromotion': [0, 1, 0, 0, 1]  # optional covariates
})

ts_df = TimeSeriesDataFrame.from_data_frame(
    df,
    id_column='item_id',
    timestamp_column='timestamp'
)

Common mistake: passing pd.DataFrame directly to predictor.fit() with a date column. AG will reject it.

Core Parameters

Parameter	Required	Default	Purpose
target	Yes	None	Column name to forecast
prediction_length	Yes	1	How many steps ahead to forecast
freq	Yes (or inferred)	None	'D', 'H', 'M', 'Q', etc.
eval_metric	No	'WQL'	RMSLE, MAE, MAPE, WQL, MASE
known_covariates_names	No	None	Future-known variables (e.g., holidays)
quantile_levels	No	[0.1, ..., 0.9]	For probabilistic forecasts
path	No	None	Model save directory