Store Sales — darts + Chronos-2 Cross-Family Blend

The Core Insight (why this beats single models)

A single strong model hits a ceiling. Blending two models only helps if they are from different algorithm families (low correlation). Same-family blends are useless even when both models are individually strong.

* Even at 0.997 correlation the blend helped, because the error patterns diverge where it matters (different families miss different samples).

The Two Routes

Route 1: AutoGluon Chronos-2 ensemble (neural foundation model)

• TimeSeriesPredictor with Chronos-2 (zero-shot + LoRA fine-tune) + Chronos-Bolt + DirectTabular
• num_val_windows=5, local model paths to bypass HF download errors
• Standalone LB: 0.39387
• See autogluon-timeseries-strategy skill for Chronos-2 details

Route 2: darts LightGBM per-family (tree model, top-1 public method)

• darts library's LightGBMModel with output_chunk_length=1 + predict(n=16)
• darts handles recursive prediction AUTOMATICALLY — this is the critical advantage. Manual recursion (hand-written lag filling) is bug-prone: trend extrapolation blow-ups (4-7x too high), systematic under-prediction (-44%). darts' built-in recursion avoids all of these.
• Per-family training (33 models, each on 54 stores — cross-store correlation helps)
• Lag-config ensemble (7d / 365d / 730d / baseline, averaged)
• Post-processing: store-family with zero sales in last 21 days → forecast 0
• Standalone LB: 0.39953

from darts import TimeSeries
from darts.models import LightGBMModel

ts = TimeSeries.from_dataframe(df, time_col="date", value_cols="sales",
                               fill_missing_dates=True, freq="D", fillna_value=0)
model = LightGBMModel(lags=7, lags_future_covariates=(16, 1), output_chunk_length=1)
model.fit(series=[ts_store1, ts_store2, ...],   # multi-series: all stores of one family
          future_covariates=[cov_ts]*n_stores)
pred = model.predict(n=16, series=ts_store, future_covariates=cov_ts)  # auto-recursive

The Blend (geometric mean in log space)

RMSLE is a log-space metric, so blend in log space (geometric mean):

import numpy as np
# w = weight on Chronos-2, (1-w) on darts
blended = np.exp(w * np.log1p(chronos_pred) + (1-w) * np.log1p(darts_pred)) - 1

Optimal weight found by sweep: w=0.55 (slightly favor Chronos-2, the stronger single model). The optimum is flat — w in [0.5, 0.6] all give ~0.3844-0.3846.

Empirical Results (Store Sales, 2026-07-26)

What Did NOT Work (recorded to save future effort)

1. Same-family blend (Chronos v1 + v2): 0.3941, no gain. Correlation too high.
2. Hand-written LightGBM recursion (Route 2 without darts): systematic under-prediction (-44% on GROCERY I). Manual lag filling is the root cause.
3. Hybrid Ridge-trend + LGB-residual: Ridge linear extrapolation blew up on test (PRODUCE predicted 4.3x too high, LB 0.71). Validation looked fine because validation has no extrapolation.
4. Per-family scaling calibration: only +0.8% theoretical gain. Error isn't simple bias.
5. 3rd model (darts XGBoost): correlation with LightGBM 0.996 → minimal additional blend gain. Same-family tree models don't add diversity.

Key Lessons

1. Read the forum FIRST. The darts top-1 method was found in the competition discussion after hours of failing with hand-written recursion. Forum research is higher-ROI than model tuning.
2. Use the right tool for recursion. darts output_chunk_length=1 + predict(n=horizon) is the correct abstraction. Never hand-write recursive lag filling for >1-step forecasts.
3. Cross-family blend > same-family stacking. Two mediocre independent models beat two correlated strong models.
4. Geometric mean for RMSLE. Since RMSLE operates in log space, blend predictions with a log-space linear combination (geometric mean), not arithmetic mean.
5. Post-processing matters. The 21-day-zero rule (force forecast=0 for store-families with all-zero last 21 days) is a free, safe gain from the winning writeup.

References

• Competition: https://www.kaggle.com/competitions/store-sales-time-series-forecasting
• Top-1 method (darts): https://inside-machinelearning.com/en/top-1-kaggle-my-method/
• AutoGluon Chronos-2 tutorial: https://auto.gluon.ai/stable/tutorials/timeseries/forecasting-chronos.html
• darts docs: https://unit8co.github.io/darts/
• Related skill: autogluon-timeseries-strategy (single Chronos-2 route details)