Rossmann daily sales forecast · Random forest

Predicts the daily sales (EUR) of any of the 1,115 Rossmann drugstores from what is known in advance: the store, the date, whether a promotion runs and the holidays. It does not use the number of customers of that day.

60 regression trees (max depth 20), each grown on half of the rows and half of the features.

Metric (last 6 weeks, held out, n=40,282 store-days) Value
R² 0.891
RMSPE (Kaggle metric) 0.142
RMSE 1,007 EUR
MAE 684 EUR

For reference, predicting each store's average for that weekday gives R² 0.696 and RMSPE 0.224.

Actual vs predicted

Data: Kaggle "Rossmann Store Sales" (train.csv + store.csv), read from AiiN-aini/rossmann-store-sales. Open days with sales only. Train: 2013-01-01 to 2015-06-19 (804,056 store-days). Test: 2015-06-20 to 2015-07-31: the last 6 weeks, never seen by the model or by the per-store statistics.

Target: log1p(sales); predictions are turned back into euros with expm1.

Features (21): calendar (weekday, day, month, year, week and day of year, month start/end), promotion and holiday flags, store type, assortment, competition distance and age, Promo2 status, and four per-store statistics from the training period (store_stats.csv).

Usage

The features are built by forecast/features.py in the project repository; scikit-learn must be version 1.9.1.

import joblib, numpy as np, pandas as pd
from huggingface_hub import snapshot_download
from forecast.features import build_features   # from the project repository

path = snapshot_download("jramduran/rossmann-sales-random-forest")
model = joblib.load(f"{path}/model.joblib")   # only load pickles you trust
stats = pd.read_csv(f"{path}/store_stats.csv")
day = pd.DataFrame([{"store": 336, "date": "2015-08-07", "promotion": 1,
                     "state_holiday": 0, "school_holiday": 0}])
X = build_features(day, stats)
features = ["day_of_week", "day", "month", "year", "week_of_year", "day_of_year", "is_month_start", "is_month_end", "promotion", "state_holiday", "school_holiday", "store_type", "assortment", "competition_distance_log", "competition_open_months", "promo2_active", "promo2_month", "store_mean_log_sales", "store_dow_log_sales", "store_promo_lift", "store_mean_log_customers"]
print(np.expm1(model.predict(X[features])))   # EUR
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train jramduran/rossmann-sales-random-forest