pyhts

A python package for hierarchical forecasting, inspired by the hts package in R.

Features

• Support pupular forecast reconciliation models in the literature, e.g. ols, wls, mint et al. Forecasting with temporal hierarchies will be supported in the future.
• Multiple methods for the construction of hierarchy.
• Use different base forecasters for different hierarchical levels.
• Sklearn-like API.

Quick Demo

• Load the Australia tourism flows data.

from pyhts import load_tourism

tourism_data = load_tourism()
train = tourism_data.iloc[:, 4:-12].T.values
test = tourism_data.iloc[:, -12:].T.values

• Define the hierarchy.

from pyhts import Hierarchy

hierarchy = Hierarchy.new(tourism_data, [('state', 'region', 'city')])
print(hierarchy.node_name)

• Create an ols forecasting reconciliation model with sklearn-like API.

from pyhts import HFModel

model_ols = HFModel(hierarchy=hierarchy, base_forecasters="arima",
                    hf_method="comb", comb_method="ols")

• Fit the model and produce forecasts.

model_ols.fit(train)
ols = model_ols.predict(horizon=12)

• model.fit() fits the baseforecasters and computes the weighting matrix used to reconcile the base forecasts.

• model.predict() calculates the base forecasts for all levels and reconciles the base forecasts.

• Obtain coherent forecasts of all the hierarchical levels.

all_level_ols = hierarchy.aggregate_ts(ols)

• fit other methods using fitted base forecasters

model_wlss = HFModel(hierarchy, base_forecasters=model_ols.base_forecasters,
                     hf_method="comb", comb_method="wls", weights="structural")
model_wlss.fit(train)
wlss = model_wlss.predict(horizon=12)

model_wlsv = HFModel(hierarchy, base_forecasters=model_ols.base_forecasters,
                     hf_method="comb", comb_method="mint", weights="variance")
model_wlsv.fit(train)
wlsv = model_wlsv.predict(horizon=12)

model_shrink = HFModel(hierarchy, base_forecasters=model_ols.base_forecasters,
                       hf_method="comb", comb_method="mint", weights="shrinkage")
model_shrink.fit(train)
shrink = model_shrink.predict(horizon=12)

• Evaluate the forecasting accuracy.

# accuracy of reconciled forecasts
accuracy = [hierarchy.accuracy(test, fcast, hist=train, measure=['mase', 'rmse'])
            for fcast in (ols, wlss, wlsv, shrink)]

# accuracy of base forecasts
base_forecasts = model_ols.generate_base_forecast(horizon=12)
accuracy_base = hierarchy.accuracy_base(test, base_forecasts, hist=train, measure=['mase', 'rmse'])

Because of the incoherence of base forecasts, base_forecasts are forecasts of all time series in the hierarchy, while coherent forecasts are forecasts of the bottom-level time series.

Documentation

See documentation here.