ts-day-specific-forecasting

Day-Specific Direct Forecasting for Multi-Step Time Series

Problem

In multi-step time series forecasting (predicting N days ahead), a single unified model suffers from stale lag features at test time. All N days get the same ffill'd lag values (constant), causing the model to underpredict by 5-10x. Even with geometric mean post-processing (see ts-lag-stale-underprediction), the model only provides "ranking signal" — the magnitude comes from target encoding, not the model itself.

Context / Trigger Conditions

Use when:

• Predicting N days (N > 1) into the future with lag/rolling features
• A single unified model underpredicts at test time due to stale lag features
• Post-processing blends (geometric mean with TE) are needed but feel like a workaround
• CV score is good but LB score is 3-10x worse
• You want the model itself to produce correct-magnitude predictions

Common in: Kaggle time series competitions (Store Sales, M5, Web Traffic), retail demand forecasting, any multi-step prediction with autoregressive features.

Solution

Core Idea: Train N Separate Models

Instead of 1 model that predicts all N days, train N separate models where:

• Model_d predicts "sales d days from the reference date"
• Features are ALWAYS computed from the last known training date
• At test time, ALL lag features reference real, known data — no ffill needed

Implementation

# For each prediction day d (1 to N):
for d in range(1, N + 1):
    # Training: features from date t → target = sales on date t+d
    ref_dates = [t for t in all_dates if t + d is still in training data]
    target_dates = [t + timedelta(days=d) for t in ref_dates]

    # Reference features (from date t, always known)
    ref_data = train_features[train_features["date"].isin(ref_dates)]

    # Target sales (d days ahead)
    target_data = train_features[train_features["date"].isin(target_dates)]

    # Merge on (store, family, target_date)
    merged = ref_data.merge(target_data, on=["store_nbr", "family", "target_date"])

    # Add TARGET-DATE features (calendar, holidays, promotions on target date)
    merged["target_day_of_week"] = merged["target_date"].dt.dayofweek
    merged["target_month"] = merged["target_date"].dt.month
    merged["target_is_weekend"] = (merged["target_day_of_week"] >= 5).astype(int)

    # Add TARGET-DATE target encoding
    # te_sf_dow_mean for the TARGET day_of_week, not the reference day
    merged = merged.merge(te_sf_dow, on=["store_nbr", "family", "target_day_of_week"])

    # Train model_d on this data
    model_d = lgb.LGBMRegressor(...)
    model_d.fit(X_train, np.log1p(y_train), eval_set=[(X_val, np.log1p(y_val))])