Nixtla TopDown (forecast proportions) returns mostly 0 as reconciliation result

[hfc25]

Dear community,

While trying out Nixtla's TopDown reconciliation with the method "forecast_proportions", I encountered the following issue: Apart from the top node in my hierarchy, all other nodes resulted in a reconciled value of 0. Since I was not sure if I had made any mistake with my inputs, I also tried out the same reconciliation approach with a test set from "Forecasting: Principles and Practice, the Pythonic Way" (aus_tourism.csv => https://otexts.com/fpppy/data/aus_tourism.csv ) and just replaced the BottomUp / MintShrink reconciliation approaches by TopDown(method="forecast_proportions"). However, once again, most of the reconciliation results are equal to 0. Is this a bug or do you have any hints on how I can implement a TopDown reconciliation that will break down the values of different top nodes to lower nodes in a proportionate way? (For context, in my actual use case, I will not have a Y_fitted_df, so the proportions should be purely based on the forecast proportions, not on historical proportions). Thanks a lot in advance :)

Here is the sample code for the aus_tourism dataframe:

First some processing steps that I added:

Split unique_id into Country, State, and Purpose

aus_tourism[['Country', 'State', 'Purpose']] = aus_tourism['unique_id'].str.split('-', expand=True)

Convert ds to datetime

aus_tourism['ds'] = pd.to_datetime(aus_tourism['ds'])

Then the official code from the website:

from hierarchicalforecast.core import HierarchicalReconciliation
from hierarchicalforecast.methods import BottomUp, TopDown, MiddleOut, MinTrace
from hierarchicalforecast.utils import aggregate, HierarchicalPlot
from statsforecast import StatsForecast
from statsforecast.models import AutoETS, Naive

2. Define the spec and aggregate the data

spec = [
["Country"],
["Country", "State"],
["Country", "Purpose"],
["Country", "State", "Purpose"],
]

Y_df, S_df, tags = aggregate(aus_tourism.drop(columns=["unique_id"]), spec)

Y_train_df = Y_df.query("ds.dt.year <= 2015")
Y_test_df = Y_df.query("ds.dt.year > 2015")

3. Specify a set of models and forecast

sf = StatsForecast(models=[AutoETS(season_length=4)], freq="Q", n_jobs=-1)

Y_hat_df = sf.forecast(h=8, df=Y_train_df, fitted=True)
Y_fitted_df = sf.forecast_fitted_values()

4. Specify the reconcilers

reconcilers = [TopDown(method="forecast_proportions")]

5. Reconcile the forecasts

hrec = HierarchicalReconciliation(reconcilers=reconcilers)

Y_rec_df = hrec.reconcile(Y_hat_df=Y_hat_df, Y_df=Y_fitted_df, S=S_df, tags=tags)