parameters.py

from dataclasses import dataclass

@dataclass
class ModelConfig:
    """Transformer architecture hyperparameters."""

    n_features: int = 94
    context_len: int = 256
    attn_window: int = 256
    n_layers: int = 4
    n_heads: int = 8
    model_dim: int = 256
    ffn_dim: int = 704
    dropout: float = 0.2
    stock_emb_dim: int = 16
    input_scale: float = 1.0
    cross_stock_attn: bool = True
    target_horizons: tuple = ("fret1", "fret6", "fret12", "fret24")

@dataclass
class TrainingConfig:
    """Training loop hyperparameters."""

    lr: float = 2e-4
    weight_decay: float = 0.5
    grad_clip: float = 1.0
    batch_size: int = 256
    n_epochs: int = 12
    warmup_ratio: float = 0.1
    lr_scheduler_eta_min: float = 1e-5

@dataclass
class PreprocessingConfig:
    """Feature preprocessing parameters and fitted state.

    Fields marked [FITTED] are computed from data by fit_preprocessing().
    Do NOT set them manually — they will be overwritten.
    """

    preprocessing_fit_days: int = 20
    eps: float = 1e-8
    feat_p01: float | None = None
    feat_p99: float | None = None
    feat_log_mask: list[bool] | None = None
    feat_mean: float | None = None
    feat_std: float | None = None
    y_std: float | None = None
    vol20_idx: int | None = None
    vol20_floor: float = 1e-8

MODEL_CONFIG = ModelConfig()
TRAINING_CONFIG = TrainingConfig()
PREPROCESSING_CONFIG = PreprocessingConfig()