如何在面向英国市场的FB Prophet时序模型中纳入COVID-19影响？

Great question—adding COVID-19 effects to your existing Prophet model is crucial for accurate post-2020 forecasts, especially since your target metric depends on early 2020 data. Below are practical, actionable methods tailored to your UK market setup:

1. Add a COVID-19 Regressor (Direct Intervention) #

This is the most straightforward approach: create a binary flag to mark periods where COVID-19 restrictions or impacts were active, then feed this as an external regressor to Prophet.

Steps: #

• Define the COVID start date: For the UK, the first national lockdown started on March 23, 2020—use this as your cutoff (adjust if your metric was affected earlier, e.g., late February 2020).
• Add the regressor to your historical data: Create a column like covid_regressor where values are 1 for dates on/after March 23, 2020, and 0 before.
• Include it in your Prophet model: Use add_regressor() to tell Prophet to account for this variable.
• Set values for the forecast period: If your 30-day forecast falls during ongoing restrictions (or specific COVID conditions), set covid_regressor to 1 (or adjust based on expected policy changes).

Code Example: #

import pandas as pd
from prophet import Prophet

# Load your existing data (df has 'ds' and 'y' columns)
df = pd.read_csv('your_data.csv')

# Add COVID regressor
df['covid_regressor'] = (df['ds'] >= '2020-03-23').astype(int)

# Initialize and fit model with the regressor
m = Prophet(seasonality_mode='additive', holidays=your_existing_holidays)
m.add_regressor('covid_regressor')
m.fit(df)

# Create future dataframe and set regressor values for the 30-day forecast
future = m.make_future_dataframe(periods=30)
future['covid_regressor'] = (future['ds'] >= '2020-03-23').astype(int)  # Adjust based on expected conditions

# Generate forecast
forecast = m.predict(future)

2. Tune Changepoints to Highlight COVID's Sudden Impact #

Prophet automatically detects trend changepoints, but you can manually specify the COVID onset date to force the model to prioritize this critical shift. Increase the changepoint_prior_scale to make the model more responsive to this abrupt change (default is 0.05; try 0.1–0.2).

Code Example: #

# Initialize model with explicit COVID changepoint
m = Prophet(
    changepoints=['2020-03-23'],  # Mark the lockdown start as a key changepoint
    changepoint_prior_scale=0.1,  # Make trend more flexible around this date
    seasonality_mode='additive',
    holidays=your_existing_holidays
)
m.fit(df)

# Generate forecast as usual
future = m.make_future_dataframe(periods=30)
forecast = m.predict(future)

3. Add COVID-Specific "Holidays" for Policy Events #

Treat major COVID-related events (lockdowns, tiered restrictions, reopenings) as custom holidays in Prophet. This works well if certain policies had clear, time-bound impacts on your metric.

Steps: #

• Create a DataFrame with COVID events, including holiday, ds (event date), and lower_window/upper_window to capture the duration of the impact (e.g., a lockdown might have a 7-day upper window to account for gradual adoption).
• Pass this DataFrame to the holidays parameter when initializing Prophet.

Code Example: #

# Define COVID-related events for the UK
covid_holidays = pd.DataFrame({
    'holiday': 'uk_national_lockdown_1',
    'ds': pd.to_datetime(['2020-03-23']),
    'lower_window': 0,
    'upper_window': 90,  # Capture 90 days of lockdown impact
})

# Add other events (e.g., tier 3 restrictions, reopening dates) as needed
covid_holidays = pd.concat([covid_holidays, pd.DataFrame({
    'holiday': 'uk_first_reopening',
    'ds': pd.to_datetime(['2020-06-01']),
    'lower_window': -7,  # Impact starts 7 days before reopening
    'upper_window': 14,
})])

# Initialize model with combined holidays (existing + COVID)
m = Prophet(
    seasonality_mode='additive',
    holidays=pd.concat([your_existing_uk_holidays, covid_holidays])
)
m.fit(df)

4. Segment Data and Combine Models (For Extreme Shifts) #

If COVID caused a permanent structural change in your metric, consider splitting your data into pre-COVID (2019-01 to 2020-02) and post-COVID (2020-03 onwards) segments:

• Train a pre-COVID model to capture baseline seasonality/holidays.
• Train a post-COVID model to learn the new trend and COVID-specific patterns.
• For forecasts, use the post-COVID model (since your metric relies on 2020+ data) or blend predictions if the pre-COVID baseline is still relevant.

Key Tips for All Methods: #

• Backtest rigorously: Use your 2020-03 onwards historical data to test which method produces the lowest forecast error (e.g., MAE, RMSE).
• Update dynamically: If your 30-day forecast includes new COVID policy changes, adjust your regressor or holiday data accordingly.
• Validate with business context: Check Prophet's component plots (m.plot_components(forecast)) to ensure the COVID impact aligns with what you know about your metric and the UK market.

内容的提问来源于stack exchange，提问作者Django0602