Search code examples
pythonpandaspredictionforecastfacebook-prophet

Prophet - saturation forecast minimum goes below 0

I'm doing a prophet prediction for a dataframe with call count and I tried to do a logistic growth model for it. Dataframe has ds column; 30min datetime bins and y column; call count for each bin

Here's the prediction code

dataframe['cap']=10000
dataframe['floor']=0

m = Prophet(growth='logistic')
m.fit(dataframe)
future = m.make_future_dataframe(periods=periods, freq='H')
    
future['cap']=10000
future['floor']=0
        
forecast = m.predict(future)
fig1 = m.plot(forecast, xlabel='Date-time DD-MM-YY:HH-MM-SS', ylabel='Call Count')
plt.title('Forecast Prediction')
plt.show()

and graph

enter image description here

Even after that the detected pattern graph and forecast both goes below zero. Why is this happening and What can I do?

Solution

  • Generally speaking, a common technique to handle negative values in prediction models is the logarithmic trasformation.

    To transform your target variable , you can use Y=log(x + c) where c is the constant.

    People usually choose something like Y=log(x+1) or any other "very small" positive number.

    After this transformation, it's impossible to get negative values from forecasting.

    To come back to original scale just use the inverse function of y=log(x+1) , that is y = exp(x)-1