Python Nixtla实现时间序列预测的完整指南

import pandas as pd
from statsforecast import StatsForecast
from statsforecast.models import AutoARIMA, AutoETS, MSTL

# 1. 准备数据（df需包含 unique_id, ds, y）
df = pd.read_csv('sales.csv')
df['ds'] = pd.to_datetime(df['ds'])

# 2. 定义模型列表
models = [
    AutoARIMA(season_length=12),      # 月度数据的年周期
    AutoETS(season_length=12),
    MSTL(season_length=[7, 365.25])    # 多季节性
]

# 3. 初始化预测器
sf = StatsForecast(models=models, freq='D', n_jobs=-1)

# 4. 拟合模型
sf.fit(df)

# 5. 生成未来30天预测，含95%置信区间
forecast_df = sf.predict(h=30, level=[95])

from neuralforecast import NeuralForecast
from neuralforecast.models import NHITS, NBEATS, PatchTST

# 定义模型
models = [
    NHITS(h=14, input_size=28, max_steps=500),
    NBEATS(h=14, input_size=28, max_steps=500),
    PatchTST(h=14, input_size=28, max_steps=500)
]

# 统一接口，和StatsForecast完全一致
nf = NeuralForecast(models=models, freq='D')
nf.fit(df)
forecast_df = nf.predict(h=14)

from neuralforecast.models import VanillaTransformer, Informer, Autoformer, FEDformer

models = [
    VanillaTransformer(h=14, input_size=28, max_steps=500),
    Informer(h=14, input_size=28, max_steps=500),
    Autoformer(h=14, input_size=28, max_steps=500)
]

from nixtlats import TimeGPT
import pandas as pd

# 读取数据
df = pd.read_csv('sales.csv')
df['timestamp'] = pd.to_datetime(df['timestamp'])

# 初始化（需API Key，官网注册即可获取）
timegpt = TimeGPT(token='YOUR_API_KEY')

# 三行代码生成预测
forecast = timegpt.forecast(df, h=30, freq='D')
timegpt.plot(df, forecast)  # 可视化

from mlforecast.lag_transforms import LagTransform
transforms = [LagTransform(lags=[1,7], func_name='rolling_mean', window_size=7)]
fcst = MLForecast(models=..., lags=[1,7], lag_transforms=transforms)

from mlforecast import MLForecast

lag_transforms = {
    1: ['rolling_mean:7'],   # 对滞后1做7天滚动平均
    7: ['rolling_mean:7', 'exp_smoothing:0.3']  # 可组合多个转换
}

fcst = MLForecast(models=..., freq='D', lags=[1,7], lag_transforms=lag_transforms)

# 第一步：快速建立多模型基线
sf = StatsForecast(models=[AutoARIMA(), AutoETS()], freq='D')
sf.fit(df)
baseline = sf.predict(h=30)

# 第二步：加入滞后特征提升精度
from mlforecast import MLForecast
mf = MLForecast(models=[LightGBM()], lags=[1,7,14], 
                lag_transforms={1:['rolling_mean:7']})
mf.fit(df)
enhanced = mf.predict(h=30)

# 第三步：复杂场景用深度学习或TimeGPT收尾
nf = NeuralForecast(models=[PatchTST()], freq='D')
nf.fit(df)
final = nf.predict(h=30)

pip install statsforecast neuralforecast mlforecast nixtla