xgb_config.yaml

# Enhanced XGBoost Configuration for Flood Depth Prediction (Tweedie Objective)

# Objective
objective: reg:tweedie                     # models zero + positive skewed values (ideal for flood depth)
tweedie_variance_power: 1.3                # adjust between 1.1–1.9 depending on data
eval_metric: rmse                          # root mean square error

# Boosting rounds & learning
n_estimators: 500                          # total boosting rounds
learning_rate: 0.5                         # learning rate (try 0.03–0.1 range if tuning)

# Tree structure control using depth
tree_method: hist                          # efficient CPU histogram algorithm (use gpu_hist for GPU)
grow_policy: depthwise                     # grow trees depth-wise (standard method)
max_depth: 8                               # maximum tree depth to control complexity

# Regularization to avoid overfitting
reg_lambda: 3.0                           # stronger L2 regularization
reg_alpha: 1.0                             # stronger L1 regularization
min_child_weight: 10                       # minimum data in a child node before split
gamma: 1.0                                 # minimum loss reduction to allow a split
max_delta_step: 1                          # limits size of leaf value updates (adds numerical stability)

# Subsampling
subsample: 0.8                             # row sampling for each tree
colsample_bytree: 0.8                      # feature sampling for each tree

# Optimization and reproducibility
early_stopping_rounds: 30                 # stop early if no improvement on validation set
random_state: 42                           # reproducibility
n_jobs: -1                                 # utilize all CPU cores
verbosity: 1                               # warnings only (set to 2 for more logs, or 0 for silent)