[feature] Retraining a fitted model

[Msaleh87]

I know how to do this in fbProphet, you get the parameters from the old model and update them using the new data points, instead of retraining the whole model from scratch. How to do the same with Neural Prophet? I tried the same function but it didn't work out.

[ourownstory]

Hi @Msaleh87
We currently have disabled retraining of an already fitted model.
Do you mind sharing how this would be beneficial to your application?

[Msaleh87]

It was a great help just telling me that the retraining was disabled, it saved me a lot of searching time, thanks!

So I have an electric load data for about a year, I trained the model for a data set from Mar 2021 until Mar 2022. I enabled the holidays. Here is my code:

m = NeuralProphet(n_forecasts=(2*24*12),
    n_lags=(16*24*12),
    yearly_seasonality="auto",
    weekly_seasonality="auto",  
    daily_seasonality="auto", 
    growth="off", seasonality_reg= 0.5, optimizer = 'AdamW',
    learning_rate = 0.1, n_changepoints="auto", changepoints_range = "auto" ) # trend_reg=1
m.add_country_holidays("US", mode="additive") # lower_window=-1, upper_window=1
metrics = m.fit(L12M, freq="5min")

m.highlight_nth_step_ahead_of_each_forecast(step_number=m.n_forecasts)

P_horizon = 576 # 2 days ahead
future = m.make_future_dataframe(df = L12M, periods = P_horizon, n_historic_predictions=576)
forecast1 = m.predict(df=future)

That's the 48 hours ahead I got, I was wondering what hyperparamters I could add/tune to have better results. Or this is as good as it gets since I am using 1 feature (holidays)?

[ourownstory]

Hi @Msaleh87 Do you mind creating a new Question in the forecasting help section?
some pointers:

• Try a lower amount of lags (e.g. just 3-7 days). OR downsample your data to 15min or hourly OR add a hidden layer of a small size (e.g. 8-32) (reason: your current AR model is massive, it has 2*24*12 * 16*24*12 = 2,654,208 parameters, and is likely overfitting)
• instead of regularizing seasonality, try setting the number of fourier terms (instead of "auto")
• use the plot_components and plot_parameters function and see if the make sense
• find external regressors like temp
• at the end, tune learning rate and epochs

[Msaleh87]

I just did, here is the link: #741
So I need to predict 48 hours ahead, which is 576 data points (2 days* 48 hrs*12 datapoints per hr), so I used n_forecast = 576. as for n_lags I tries to look back 2-3 days, but the forecast used to come out bad. When I played with the hidden layers and neurons, it was also not that great, even though I think I used 3 hidden layers and 64 neurons I guess. I also tried to set seasonality like the documentation daily is set to a default of 6 and weekly to 4, in the documentation it mentioned that I should use multiple of those (like for daily, it should be m * 6, and for weekly: n * 4, where m and n are integers) to capture the complex seasonality. I guess I'll try to change seasonality again, but I remember that I did, but I'll give it another try. But I am sure that n_lags that big started to give me reasonable forecasts like the one in the picture. I know since the horizon is big I have to sacrifice some of the accuracy, but I was hoping for more accurate model, or better tune it.
Is there a way, besides grid search to tune a Neural prophet model?

[ourownstory]

replied: #741 (reply in thread)

[tercjak]

I think the best is to enable updating data to trained model.
When we have this feature we could train 3 models (n_forecast1=3 A,n_forecast2,=4 B, n_forecast3=5 C) so we can get forecasts with lower amount of steps (eg 3-5 -1 step 6-10 2steps 11-15 3steps etc. ) 3,4,5 6,7,8 ,9..= A,B,C,AA,AB,AC,BB,BC,CC ... etc. For now the only option to do this is to add lagged Y data as regressor. There is no need to sacrifice accuracy if the horizon is larger; on the contrary, a large amount of data should give us a more accurate model. If the opposite is the case, our methodology is flawed.