67 lines
2.2 KiB
Python
67 lines
2.2 KiB
Python
|
"""时间序列（MCM C 必备）。
|
|||
|
|
|
|||
|
|
原建议列了 `moving_average` `weighted_moving_average` `holt_trend`
|
|||
|
|
`seasonal_decompose` `lag_features` `forecast_interval` 等一堆函数，
|
|||
|
|
但 statsmodels 已经全部提供，按它自己的原则（"做薄封装，不要重写框架"）
|
|||
|
|
只保留三个真正有封装价值的入口。
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
from __future__ import annotations
|
|||
|
|
|
|||
|
|
from typing import Any
|
|||
|
|
|
|||
|
|
__all__ = ["exponential_smoothing", "fit_arima", "forecast"]
|
|||
|
|
|
|||
|
|
|
|||
|
|
def exponential_smoothing(series: Any, alpha: float | None = None, trend: str | None = None, seasonal: str | None = None, periods: int | None = None) -> Any:
|
|||
|
|
"""指数平滑（底层 `statsmodels.tsa.holtwinters`）。
|
|||
|
|
|
|||
|
|
`alpha=None` 时自动用极大似然估计平滑系数——比自己拍脑袋定 `alpha` 靠谱。
|
|||
|
|
|
|||
|
|
Examples
|
|||
|
|
--------
|
|||
|
|
>>> r = exponential_smoothing(y, trend="add") # doctest: +SKIP
|
|||
|
|
"""
|
|||
|
|
# TODO(L1)
|
|||
|
|
raise NotImplementedError
|
|||
|
|
|
|||
|
|
|
|||
|
|
def fit_arima(series: Any, order: tuple[int, int, int] = (1, 1, 1), seasonal_order: tuple[int, int, int, int] | None = None, auto: bool = False) -> Any:
|
|||
|
|
"""ARIMA 拟合。
|
|||
|
|
|
|||
|
|
Parameters
|
|||
|
|
----------
|
|||
|
|
auto:
|
|||
|
|
`True` 时按 AIC 网格搜索 (p, d, q)，省去手动定阶；
|
|||
|
|
数据量大时较慢，比赛里建议先手动看一眼 ACF/PACF。
|
|||
|
|
order:
|
|||
|
|
`(p, d, q)`——自回归阶数、差分阶数、移动平均阶数。
|
|||
|
|
|
|||
|
|
Returns
|
|||
|
|
-------
|
|||
|
|
`Result`：`metrics={"aic", "bic"}`，`extras["model"]` 是 statsmodels 结果对象
|
|||
|
|
（需要 `summary()`、残差诊断时直接拿）。
|
|||
|
|
|
|||
|
|
Notes
|
|||
|
|
-----
|
|||
|
|
ARIMA 要求序列**平稳**。`d` 是差分阶数，但差分会丢信息，
|
|||
|
|
非平稳序列更推荐先看是不是有结构突变或季节成分。
|
|||
|
|
"""
|
|||
|
|
# TODO(L1)
|
|||
|
|
raise NotImplementedError
|
|||
|
|
|
|||
|
|
|
|||
|
|
def forecast(model: Any, steps: int, alpha: float = 0.05) -> Any:
|
|||
|
|
"""向后预测 `steps` 步，并给出预测区间。
|
|||
|
|
|
|||
|
|
Returns
|
|||
|
|
-------
|
|||
|
|
`Result`：`data` 是含 `forecast / lower / upper` 三列的 DataFrame。
|
|||
|
|
|
|||
|
|
Notes
|
|||
|
|
-----
|
|||
|
|
**预测区间随步长急剧变宽是正常的**，不要只画中心线——
|
|||
|
|
论文里只画一条线而不给区间，是很常见的扣分点。
|
|||
|
|
"""
|
|||
|
|
# TODO(L1)
|
|||
|
|
raise NotImplementedError
|