src/forecasting/config.py

"""
TFT model configuration.
Centralises all hyperparameters so the Streamlit app and training script share one source of truth.
"""

from dataclasses import dataclass, field
from typing import List


@dataclass
class TFTConfig:
    # Temporal structure
    input_size: int = 60        # lookback window (days)
    h: int = 28                 # forecast horizon (days)

    # Architecture
    hidden_size: int = 64
    n_head: int = 4
    attn_dropout: float = 0.1
    dropout: float = 0.1

    # Training
    max_steps: int = 500
    learning_rate: float = 1e-3
    batch_size: int = 32
    val_check_steps: int = 50

    # Quantiles for probabilistic output (confidence bands)
    loss_quantiles: List[float] = field(
        default_factory=lambda: [0.1, 0.5, 0.9]
    )

    # Covariates
    # Static: features that don't change over time per series
    stat_exog_list: List[str] = field(
        default_factory=lambda: ["category_enc", "store_enc"]
    )
    # Historic: known past values
    hist_exog_list: List[str] = field(
        default_factory=lambda: ["price", "event_flag"]
    )
    # Future: known future values (calendar events we schedule in advance)
    futr_exog_list: List[str] = field(
        default_factory=lambda: ["event_flag", "day_of_week", "month", "is_fashion_week"]
    )

    # Persistence
    model_dir: str = "models"
    model_name: str = "tft_luxury"


DEFAULT_CONFIG = TFTConfig()