Time SeriesForecastingTabular

Store Sales Time Series Forecasting

Forecasting daily sales across 1,782 store × product time series

Predict daily sales for every store × product-family combination over a 16-day holdout window — Corporación Favorita's retail data, evaluated by RMSLE (root mean squared log error), which penalizes relative rather than absolute error so a miss on a slow-selling item counts as much as a miss on a bestseller.

Dataset

Source

Kaggle time-series forecasting competition — Corporación Favorita (Ecuadorian grocery retailer) daily sales, 2013–2017

•3,000,888 train rows, 2013-01-01 to 2017-08-15 (~4.6 years daily)
•28,512 test rows: 16-day forecast horizon (2017-08-16 to 2017-08-31)
•54 stores × 33 product families = 1,782 individual time series
•Auxiliary tables: store metadata (22 cities, 16 states, 5 store types, 17 clusters), daily oil price, national/bridge holidays, daily transactions per store
•31.3% of rows have zero sales; sales are heavily right-skewed (median 11, mean 358, max 124,717)

Approach

•LightGBM regression on log1p(sales) — turns RMSLE minimization into plain RMSE minimization
•Lag features at 7/14/21/28/35/42/56 days per (store_nbr, family) group — multiples of 7 to align with weekly seasonality
•Rolling mean (7/14/30-day) and rolling std (7-day), all shifted 7 days back before rolling to prevent leakage from the future
•Promotion and transaction lag/rolling features, plus date features (day-of-week, week-of-year, is_weekend, is_month_start/end, is_payday for Ecuador's 15th/month-end paydays)
•Oil price (Ecuador's economy is oil-dependent) and national/bridge holiday flags merged in as auxiliary signals
•Categorical columns (family, city, state, store type) integer-encoded via pandas cat.codes — plain label encoding, not LightGBM's native categorical splits

Training

•num_leaves=127, learning_rate=0.05, feature_fraction/bagging_fraction=0.8, min_child_samples=50
•Validation: last 16 days of training data held out, matching the actual test horizon length — a time-based split, not random k-fold, to avoid leaking future information into training
•Early stopping (patience 50 rounds) against validation RMSE, best iteration 1,187 of a 3,000-round budget
•Final model retrained on the full training set using that fixed iteration count before predicting on the test period

Results

Validation RMSLE0.3793

•Validation RMSLE: 0.37934
•lag_7 (same weekday one week prior) dominates feature importance by a wide margin — gain of 122M, over 4× the next-highest feature (lag_7_roll3, a 3-week rolling average built on the same lag)
•Weekly seasonality features (lag_7, lag_7_roll3, roll_mean_7) account for the vast majority of the model's predictive power; store transactions, product family, and oil price contribute meaningfully but far behind