LightGBM for Time Series Forecasting

LightGBM is another powerful gradient-boosting algorithm, designed to be faster and more memory-efficient than XGBoost. In real production systems, LightGBM is often preferred when datasets are large and forecasts must be generated quickly.

You will see LightGBM heavily used in:

• Retail demand forecasting
• Energy consumption prediction
• Ride-sharing demand estimation
• Financial risk modeling

Real-World Scenario

Suppose you manage inventory for a grocery delivery company. You must forecast daily item demand to avoid stockouts and food waste.

Demand depends on:

• Recent sales history
• Weekly shopping behavior
• Short-term demand spikes

LightGBM is well-suited for this type of fast, pattern-driven forecasting.

Step 1: Create a Time Series Dataset

We simulate daily demand with trend, weekly seasonality, and noise.

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

np.random.seed(7)

days = 320
time = np.arange(days)

trend = time * 0.25
weekly = 18 * np.sin(2 * np.pi * time / 7)
noise = np.random.normal(0, 6, days)

demand = 60 + trend + weekly + noise

df = pd.DataFrame({"demand": demand})
df.head()

What this plot represents:

• Gradual growth in demand over time
• Clear weekly shopping cycles
• Random day-to-day fluctuations

Step 2: Feature Engineering

Just like XGBoost, LightGBM needs engineered features to understand time.

We create:

• Short-term lags
• Weekly lag
• Rolling averages

df["lag_1"] = df["demand"].shift(1)
df["lag_2"] = df["demand"].shift(2)
df["lag_7"] = df["demand"].shift(7)

df["rolling_7"] = df["demand"].rolling(7).mean()
df["rolling_14"] = df["demand"].rolling(14).mean()

df = df.dropna()
df.head()

Why this works:

• Lag features give memory
• Weekly lag captures recurring behavior
• Rolling means smooth random noise

Step 3: Time-Aware Train-Test Split

We respect the time order to avoid data leakage.

X = df.drop("demand", axis=1)
y = df["demand"]

split = int(len(df) * 0.8)

X_train, X_test = X.iloc[:split], X.iloc[split:]
y_train, y_test = y.iloc[:split], y.iloc[split:]

Step 4: Train LightGBM Model

LightGBM grows trees leaf-wise, which allows it to capture complex patterns quickly.

from lightgbm import LGBMRegressor

model = LGBMRegressor(
    n_estimators=400,
    learning_rate=0.05,
    max_depth=6,
    subsample=0.8,
    colsample_bytree=0.8,
    random_state=7
)

model.fit(X_train, y_train)

preds = model.predict(X_test)

Step 5: Forecast Visualization

This visualization tells us whether the model truly learned the demand behavior.

Key observations:

• Predictions closely follow real demand
• Weekly oscillations are preserved
• Short-term noise is partially smoothed

Prediction Error Behavior

Error plots reveal model stability and bias.

What this shows:

• No long-term drift in errors
• Random spread around zero
• Model reacts well to demand changes

XGBoost vs LightGBM (Intuition)

• XGBoost: more conservative, robust
• LightGBM: faster, sharper learning
• LightGBM scales better for very large data

In practice, teams often try both and compare results.

Practice Questions

Next lesson: Time-series cross-validation and why normal CV fails.