code/methods/_config.py

"""Per-method typed Pydantic configurations for the MacroLens unified API.

Each method class declares its own ``MethodConfig`` subclass with typed,
validated, default-bearing hyperparameters. The runner records
``config.model_dump()`` into ``RunRecord.hyperparams`` for every result.
"""

from __future__ import annotations

from typing import Literal

import pydantic


class MethodConfig(pydantic.BaseModel):
    """Base for all method configs.

    ``model_config = ConfigDict(extra="forbid")`` so unknown kwargs raise at
    construction; methods can override to allow ``extra="allow"`` if they
    intentionally pass through to a wrapped library.
    """

    model_config = pydantic.ConfigDict(extra="forbid", frozen=True)


# ── Naive ──────────────────────────────────────────────────────────────────


class PersistenceConfig(MethodConfig):
    # Index of the "close" feature on the (N, lookback, F) X array. Since
    # X is a numpy ndarray (no column names), the runner must pass this
    # in via config; default 0 matches the convention that close is the
    # first numeric feature returned by the T1 loader.
    close_feature_idx: int = 0
    # Optional explicit horizon override; when None, fit() reads horizon
    # from y.shape[1] and stores it as self._horizon.
    horizon: int | None = None


class SectorMedianConfig(MethodConfig):
    fallback_to_global: bool = True


class MetroMedianConfig(MethodConfig):
    fallback_to_global: bool = True
    metro_key: Literal["city_state", "state", "state_property_type"] = (
        "state_property_type"
    )


class HistoricalAnalogueConfig(MethodConfig):
    fallback_to_global: bool = True


class LogSizeOLSConfig(MethodConfig):
    use_log_assets: bool = True
    sector_dummies: bool = True


# ── Classical ──────────────────────────────────────────────────────────────


class LightGBMConfig(MethodConfig):
    """LightGBM library defaults; only ``n_jobs`` is a system-level flag."""

    n_estimators: int = 100  # LightGBM default
    max_depth: int = -1  # LightGBM default (unlimited)
    learning_rate: float = 0.1  # LightGBM default
    subsample: float = 1.0  # LightGBM default
    colsample_bytree: float = 1.0  # LightGBM default
    n_jobs: int = 8  # system-level flag (not a hyperparameter)
    verbosity: int = -1
    t6_text_handling: Literal["sector_industry_only"] = "sector_industry_only"


class RandomForestConfig(MethodConfig):
    """sklearn RandomForestRegressor library defaults; ``n_jobs`` is system-level."""

    n_estimators: int = 100  # sklearn default
    max_depth: int | None = None  # sklearn default (unlimited)
    min_samples_leaf: int = 1  # sklearn default
    n_jobs: int = 8  # system-level flag (not a hyperparameter)
    t6_text_handling: Literal["sector_industry_only"] = "sector_industry_only"


# ── Sequence ───────────────────────────────────────────────────────────────


class SequenceConfig(MethodConfig):
    """Shared training-loop defaults aligned with each upstream paper's
    reference script (DLinear / iTransformer / ModernTCN — all 3 papers
    use train_epochs=10, batch_size=32, patience=3, no weight_decay)."""

    epochs: int = 10  # all three upstream papers use 10
    batch_size: int = 32  # all three upstream papers use 32
    learning_rate: float = 1e-4  # subclass overrides match each paper
    patience: int = 3  # all three upstream papers use 3
    weight_decay: float = 0.0  # upstream papers don't use weight_decay
    grad_clip: float = 1.0
    target_idx: int = 0


class DLinearConfig(SequenceConfig):
    """DLinear (Zeng et al. AAAI 2023; cure-lab/LTSF-Linear, ETTh1 ref)."""

    moving_avg: int = 25  # paper default
    learning_rate: float = 5e-3  # ETTh1 reference script lr=0.005


class ITransformerConfig(SequenceConfig):
    """iTransformer (Liu et al. ICLR 2024; thuml/iTransformer, ETTh1 ref)."""

    d_model: int = 128  # ETTh1 reference d_model=128
    n_heads: int = 8  # paper default
    e_layers: int = 2  # ETTh1 reference e_layers=2
    d_ff: int = 128  # ETTh1 reference d_ff=128
    dropout: float = 0.1  # paper default
    factor: int = 1  # paper default
    activation: str = "gelu"  # paper default
    learning_rate: float = 1e-4  # ETTh1 reference lr=0.0001


class ModernTCNConfig(SequenceConfig):
    """ModernTCN (Donghao & Xue ICLR 2024; luodhhh/ModernTCN, ETTh1 ref)."""

    patch_size: int = 16  # paper default
    patch_stride: int = 8  # paper default
    d_model: int = 64  # ETTh1 reference d_model=64
    kernel_size: int = 25  # paper default
    stem_ratio: int = 1  # paper default
    downsample_ratio: int = 2  # paper default
    ffn_ratio: int = 2  # paper default
    num_blocks: tuple[int, ...] = (1,)  # paper default
    large_size: tuple[int, ...] = (51,)  # paper default
    small_size: tuple[int, ...] = (5,)  # paper default
    dropout: float = 0.1  # paper default
    head_dropout: float = 0.1  # paper default
    revin: bool = True  # paper default
    affine: bool = True  # paper default
    learning_rate: float = 1e-3  # ETTh1 reference lr=0.001


# ── TSFM (zero-shot) ───────────────────────────────────────────────────────


class TSFMConfig(MethodConfig):
    """Shared base for Chronos2 / Moirai2 / TimesFM."""

    model_id: str = ""           # subclass overrides the default
    device: Literal["auto", "cpu", "cuda"] = "auto"
    batch_size: int = 32
    target_idx: int = 0          # close-column index in T1 X (N, L, F)


class Chronos2Config(TSFMConfig):
    model_id: str = "amazon/chronos-2"
    num_samples: int = 20


class Moirai2Config(TSFMConfig):
    model_id: str = "Salesforce/moirai-2.0-R-small"


class TimesFMConfig(TSFMConfig):
    model_id: str = "google/timesfm-1.0-200m-pytorch"
    per_core_batch_size: int = 32
    granularity: Literal["daily", "weekly", "monthly"] = "daily"


# ── LLM (frontier) ─────────────────────────────────────────────────────────


class LLMConfig(MethodConfig):
    model_id: str = ""
    tensor_parallel_size: int = 1
    max_model_len: int = 8192
    temperature: float = 0.0
    max_tokens: int = 256
    enable_thinking: bool = False
    # Number of in-context (X_train, y_train) examples to include in the
    # prompt at predict time. ``0`` (default) = pure zero-shot.
    in_context_k: int = 0
    # Smoke-test mode: ``predict`` returns deterministic fake outputs
    # without invoking the engine. Used when ``engine=None`` in CI.
    dry_run: bool = False


class LlamaScoutConfig(LLMConfig):
    model_id: str = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
    tensor_parallel_size: int = 4


class Gemma4Config(LLMConfig):
    model_id: str = "google/gemma-4-31B-it"
    tensor_parallel_size: int = 2


class Qwen35Config(LLMConfig):
    model_id: str = "Qwen/Qwen3.5-27B-FP8"
    tensor_parallel_size: int = 1


# ── LLM-TS multi-task ──────────────────────────────────────────────────────


class LLMTSConfig(MethodConfig):
    model_id: str = ""
    device: Literal["auto", "cpu", "cuda"] = "auto"
    # Smoke-test mode: ``predict`` returns deterministic fake outputs without
    # invoking a real engine. Mirrors :class:`LLMConfig.dry_run`.
    dry_run: bool = False


class ChatTimeConfig(LLMTSConfig):
    model_id: str = "ChengsenWang/ChatTime-1-7B-Chat"
    hist_len: int = 63
    pred_len: int = 21


class TimeMQAConfig(LLMTSConfig):
    model_id: str = "Time-MQA/Qwen-2.5-7B"
    base_model_id: str = "Qwen/Qwen2.5-7B-Instruct"


# ── LLM fine-tune (deferred) ───────────────────────────────────────────────


class LLMFineTunedConfig(LLMConfig):
    lora_r: int = 16
    lora_alpha: int = 32
    epochs: int = 3
    learning_rate: float = 2e-4