Repository navigation
The seasonality component in the forecast for the training data differs from that in the future predictions #2662
Description
Hi!
I have encountered an issue while using Facebook Prophet for time series forecasting. The seasonal component included in the forecast for the training data appears to be different from the seasonal component applied when making future predictions. Specifically, the weekly seasonality pattern observed in historical data does not seem to be reflected consistently in the forecasted values for the future.
Ultimately, due to this issue, the forecast for the future appears unrealistic in terms of weekly seasonality. I have not been able to find an explanation for this behavior myself. I would greatly appreciate any explanation from knowledgeable professionals.
Below, I have included a code snippet and a visualization to illustrate the issue. Thank you!
def single_prophet(df, train_dates, seasonality_mode='additive'):
df = df.copy()
df['ds'] = pd.to_datetime(df['ds'])
df = df.sort_values('ds')
future_pred_dates = pd.DataFrame(pd.date_range(start=pd.Timestamp.now(), end=pd.Timestamp('2025-12-31'), freq='D'), columns=['ds'])
model = Prophet(seasonality_mode=seasonality_mode) # , holidays=holidays, holidays_prior_scale=1
model.fit(df[df['ds'] >= train_dates[0]])
fact_pred = model.predict(df[df['ds'] >= train_dates[0]][['ds']])
future_pred = model.predict(future_pred_dates)
fig = go.Figure()
fig.add_trace(go.Scatter(x=df.ds, y=df['y'], name='fact'))
fig.add_trace(go.Scatter(x=fact_pred.ds, y=fact_pred['yhat'], name='fact_pred'))
fig.add_trace(go.Scatter(x=fact_pred.ds, y=fact_pred['trend'], name='fact_trend'))
fig.add_trace(go.Scatter(x=fact_pred.ds, y=fact_pred['weekly'], name='fact_weekly'))
fig.add_trace(go.Scatter(x=future_pred.ds, y=future_pred['yhat'], name='future_pred'))
fig.add_trace(go.Scatter(x=future_pred.ds, y=future_pred['trend'], name='future_trend'))
fig.add_trace(go.Scatter(x=future_pred.ds, y=future_pred['weekly'], name='future_weekly'))
fig.show()
single_prophet(
data,
train_dates=['2025-02-01', '2025-03-01']
)
Activity