This study presents a comparative evaluation of the LightGBM and XGBoost algorithms for the task of next-day (J+1) daily precipitation forecasting. The analysis utilizes a comprehensive dataset of 8,765 daily meteorological observations spanning the entire continental United States over a 24-year period (2000-2023). The research focuses on assessing predictive performance in relation to seasonal climatic variables and evaluates model robustness against interannual variability. With a pedagogical objective, the study aims to identify the most influential climatic determinants for short-term hydrometeorological prediction.

Dataset: https://www.kaggle.com/datasets/shivamshinde1904/weather-data2000-2023

Acknowledgment: shivamShinde1904

Import necessary libraries

import pandas as pd import numpy as np from sklearn.model_selection import train_test_split from sklearn.metrics import mean_squared_error, r2_score, mean_absolute_error import lightgbm as lgb import xgboost as xgb import matplotlib.pyplot as plt import seaborn as sns

Load dataset (copy-paste method)

df = pd.read_clipboard() df['Date'] = pd.to_datetime(df['Date'], dayfirst=True) df['Precipitation_Tomorrow'] = df['Precipitation_Sum'].shift(-1) df = df.dropna(subset=['Precipitation_Tomorrow'])

Descriptive statistics

print("Descriptive statistics of precipitation:") print(df['Precipitation_Sum'].describe()) print(f"\nNumber of dry days: {(df['Precipitation_Sum'] == 0).sum()}") print(f"Number of rainy days: {(df['Precipitation_Sum'] > 0).sum()}")

Prepare data for modeling

Features (X): remove target and date

X = df.drop(['Precipitation_Tomorrow', 'Date'], axis=1)

Target (y): next day precipitation

y = df['Precipitation_Tomorrow']

7. Split data into training (80%) and test (20%) sets

shuffle=False to preserve temporal order

X_train, X_test, y_train, y_test = train_test_split( X, y, test_size=0.2, random_state=42, shuffle=False ) print(f"\nDataset dimensions:") print(f"Train: {X_train.shape}, Test: {X_test.shape}")

TRAINING AND EVALUATION OF LIGHTGBM

print("\n" + "="*50) print("LightGBM") print("="*50)

lgb_model = lgb.LGBMRegressor( random_state=42, n_estimators=100, learning_rate=0.1, max_depth=5 ) lgb_model.fit(X_train, y_train)

lgb_pred = lgb_model.predict(X_test)

Evaluation

lgb_rmse = np.sqrt(mean_squared_error(y_test, lgb_pred)) lgb_mae = mean_absolute_error(y_test, lgb_pred) lgb_r2 = r2_score(y_test, lgb_pred)

print(f"RMSE: {lgb_rmse:.4f}") print(f"MAE: {lgb_mae:.4f}") print(f"R²: {lgb_r2:.4f}")

TRAINING AND EVALUATION OF XGBOOST

print("\n" + "="*50) print("XGBoost") print("="*50)

xgb_model = xgb.XGBRegressor( random_state=42, n_estimators=100, learning_rate=0.1, max_depth=5 ) xgb_model.fit(X_train, y_train)

xgb_pred = xgb_model.predict(X_test)

Evaluation

xgb_rmse = np.sqrt(mean_squared_error(y_test, xgb_pred)) xgb_mae = mean_absolute_error(y_test, xgb_pred) xgb_r2 = r2_score(y_test, xgb_pred)

print(f"RMSE: {xgb_rmse:.4f}") print(f"MAE: {xgb_mae:.4f}") print(f"R²: {xgb_r2:.4f}")

VISUAL COMPARISON OF PERFORMANCE

plt.figure(figsize=(12, 6))

Real vs Predicted

plt.subplot(1, 2, 1) plt.scatter(y_test, lgb_pred, alpha=0.5, label='LightGBM', s=10) plt.scatter(y_test, xgb_pred, alpha=0.5, label='XGBoost', s=10) plt.plot([y_test.min(), y_test.max()], [y_test.min(), y_test.max()], 'r--', lw=2) plt.xlabel('Actual Values') plt.ylabel('Predicted Values') plt.title('Actual vs Predicted Precipitation') plt.legend() plt.grid(True, alpha=0.3)

Feature importance (LightGBM)

plt.subplot(1, 2, 2) feature_importance = pd.DataFrame({ 'feature': X.columns, 'importance': lgb_model.feature_importances_ }) feature_importance = feature_importance.sort_values('importance', ascending=True) plt.barh(feature_importance['feature'], feature_importance['importance']) plt.xlabel('Importance') plt.title('Feature Importance - LightGBM') plt.tight_layout() plt.show()

ERROR ANALYSIS ON FIRST 100 TEST DAYS

plt.figure(figsize=(15, 6)) sample_size = min(100, len(y_test)) plt.plot(y_test.values[:sample_size], label='Actual', marker='o', markersize=3) plt.plot(lgb_pred[:sample_size], label='LightGBM', marker='x', markersize=3) plt.plot(xgb_pred[:sample_size], label='XGBoost', marker='s', markersize=3) plt.xlabel('Days') plt.ylabel('Precipitation (mm)') plt.title('Comparison of Predictions on First 100 Test Days') plt.legend() plt.grid(True, alpha=0.3) plt.show()

DISPLAY BEST AND WORST PREDICTIONS

results = pd.DataFrame({ 'Actual': y_test.values, 'LightGBM': lgb_pred, 'XGBoost': xgb_pred }) results['LightGBM_Error'] = np.abs(results['Actual'] - results['LightGBM']) results['XGBoost_Error'] = np.abs(results['Actual'] - results['XGBoost'])

print("\nTop 5 LightGBM predictions (lowest error):") print(results.nsmallest(5, 'LightGBM_Error')[['Actual', 'LightGBM', 'LightGBM_Error']])

print("\nBottom 5 LightGBM predictions (highest error):") print(results.nlargest(5, 'LightGBM_Error')[['Actual', 'LightGBM', 'LightGBM_Error']])

Residual plot

plt.figure(figsize=(10, 6)) plt.scatter(y_test, y_test - lgb_pred, alpha=0.5, label='LightGBM', s=10) plt.scatter(y_test, y_test - xgb_pred, alpha=0.5, label='XGBoost', s=10) plt.hlines(y=0, xmin=y_test.min(), xmax=y_test.max(), colors='r', linestyles='dashed', lw=2) plt.xlabel('Actual Precipitation (mm)') plt.ylabel('Residuals (Actual - Predicted)') plt.title('Residual Plot') plt.legend() plt.grid(True, alpha=0.3) plt.show()

plt.figure(figsize=(12, 6)) sns.histplot(y_test - lgb_pred, bins=50, color='blue', alpha=0.5, label='LightGBM') sns.histplot(y_test - xgb_pred, bins=50, color='green', alpha=0.5, label='XGBoost') plt.xlabel('Prediction Error (mm)') plt.ylabel('Frequency') plt.title('Distribution of Prediction Errors') plt.legend() plt.grid(True, alpha=0.3) plt.show()