使用Prophet做时间序列预测效果差，如何优化模型？

我用Prophet做时间序列预测，但效果极差，MAE和R-squared指标都很低。已经试过调整季节性设置和模式，没明显改善。

数据集概览 #

我的代码 #

# pandas
import pandas as pd
# prophet
from prophet import Prophet
# metrics
from sklearn.metrics import r2_score, mean_absolute_error

# holoviews
import holoviews as hv
from holoviews import opts

# Create a new column to identify if the date is a weekday or weekend
df_prophet['day_type'] = df_prophet['ds'].apply(lambda x: 'Weekend' if x.weekday() >= 5 else 'Weekday')
df_prophet


threshold_date = pd.to_datetime('2023-10-01')
mask = df_prophet['ds'] < threshold_date
# Split the data and select `ds` and `y` columns.
df_train = df_prophet[mask][['ds', 'y']]
df_test = df_prophet[~ mask][['ds', 'y']]

def build_model():
    """Define forecasting model."""
    model = Prophet(
        yearly_seasonality=True,
        weekly_seasonality=True,
        # daily_seasonality=True,
        # holidays = holidays,
        interval_width=0.95,
        mcmc_samples = 1000,
        # seasonality_mode = 'multiplicative',
        seasonality_mode = 'additive',
        growth= 'linear'
    )

    # model.add_seasonality(
    #     name='daily',
    #     period=5,
    #     fourier_order=5
    # )

    return model

model = build_model()

model.fit(df_train)

# Extend dates and features.
horizon = df_test.shape[0]

future = model.make_future_dataframe(periods=horizon, freq='D') # daily days predictions
# Generate predictions.
forecast = model.predict(df=future)

forecast.loc[:, 'yhat'] = forecast['yhat'].clip(lower=0)
forecast.loc[:, 'yhat_lower'] = forecast['yhat_lower'].clip(lower=0)

print('r2 train: {}'.format(r2_score(y_true=df_train['y'], y_pred=forecast_train['yhat'])))
print('r2 test: {}'.format(r2_score(y_true=df_test['y'], y_pred=forecast_test['yhat'])))
print('---'*10)
print('mae train: {}'.format(mean_absolute_error(y_true=df_train['y'], y_pred=forecast_train['yhat'])))
print('mae test: {}'.format(mean_absolute_error(y_true=df_test['y'], y_pred=forecast_test['yhat'])))

预测结果 #

优化建议 #

1. 用上已构造的day_type特征 #

你已经生成了day_type列，但训练时没传入模型。Prophet支持添加额外回归因子，在model.fit()前加上：

model.add_regressor('day_type')

同时，生成future dataframe后，要把day_type列合并进去，否则预测时会缺失特征：

future = future.merge(df_prophet[['ds', 'day_type']], on='ds', how='left')

2. 修正代码变量错误 #

打印指标时用到的forecast_train和forecast_test未定义，需要拆分forecast：

forecast_train = forecast[forecast['ds'] < threshold_date]
forecast_test = forecast[forecast['ds'] >= threshold_date]

否则无法正确计算指标，甚至会报错。

3. 匹配数据特征调整模型增长模式 #

从数据集图看，数据存在明显突变点，默认线性增长可能不适用：

• 调大changepoint_prior_scale让模型更敏感捕捉突变：

  model = Prophet(growth='linear', changepoint_prior_scale=0.1)
  

• 手动指定突变点日期（根据数据实际突变情况调整）：

  model = Prophet(changepoints=['2023-05-01', '2023-08-01'])
  

4. 优化季节性设置 #

• 如果数据波动幅度随数值增大而变大，切换到seasonality_mode='multiplicative'替代加法模式。
• 自定义季节性规则，比如针对工作日/周末设置不同的周季节性：

  model.add_seasonality(name='weekly_working', period=7, fourier_order=3, condition_name='day_type')
  

• 调整fourier_order值：增大该值可让模型捕捉更复杂的季节性，但注意避免过拟合。

5. 关闭不必要的MCMC采样 #

mcmc_samples=1000会大幅增加训练时间，非特殊场景（如需要精确不确定性估计）建议去掉该参数，用默认MAP估计即可。

6. 加入节假日影响 #

如果业务有固定节假日影响数据，收集节假日列表传入模型：

holidays = pd.DataFrame({
  'holiday': 'festival',
  'ds': pd.to_datetime(['2023-01-01', '2023-05-01']), # 替换为实际节假日日期
  'lower_window': 0,
  'upper_window': 1,
})
model = Prophet(holidays=holidays)

内容的提问来源于stack exchange，提问作者AndCh