Can't cycle prediction be made?

[zhizhuaa]

What happened + What you expected to happen

When I have trained a model, I cannot make cyclic predictions. How can I implement the loop prediction of the trained model rows? It's impossible to call nf.cross_validation(), conduct real-time training each time and then make predictions, right?

Versions / Dependencies

Reproduction script

        merged_df['unique_id'] = 1
        merged_df = merged_df.loc[:, ['y', 'ds'] + list(merged_df.columns[4:])]
        train_df = merged_df[merged_df['ds'] < eval_start]
        val_df = merged_df[(merged_df['ds'] >= val_start) & (merged_df['ds'] < val_end)]
        test_df = merged_df[(merged_df['ds'] >= eval_start) & (merged_df['ds'] < eval_end)]
        folder_path = folder_path + 'deep/'
        models = [LSTM(input_size=96, 
                       h=1, 
                       scaler_type='standard',
                       max_steps=200,
                       val_check_steps=10,
                       early_stop_patience_steps=3,
                       futr_exog_list=[col for col in merged_df.columns if col not in ['unique_id','ds','y']]
                      )]
        nf = NeuralForecast(models=models, freq='15min')
        file_name ='configuration.pkl'
        if os.path.exists(folder_path+file_name):
            nf = NeuralForecast.load(path=folder_path)
        else:
            nf.fit(df=train_df, val_size=len(val_df))
            nf.save(path=folder_path,
                    model_index=None,
                    overwrite=True,
                    save_dataset=True)
        # cv_df_val_test = nf.cross_validation(merged_df, val_size=len(val_df), test_size=len(test_df), n_windows=None)
        results = []
        for index, row in test_df.iterrows():
            futr_df = pd.DataFrame([row])
            Y_hat_df = nf.predict(futr_df=futr_df)
            results.append(Y_hat_df)
        final_results = pd.concat(results, ignore_index=True)

Issue Severity

None

[marcopeix]

Hello! I'm not entirely sure I understand what you mean.

You can run cross-validation with refit=False. This trains the model on the initial training set (whatever precedes n_windows*horizon), and then the model is fixed and it predicts each windows specified.

Or, you can set refit=True and then the model will be retrained with every new window.

You can read more in our cross-validation tutorial.

Does that answer your question?

[zhizhuaa]

That's not what I mean. What I mean is to train a model well, then directly load this model, call nf.predict(), input the futr_df at a specific future time, and you can get the corresponding predicted value.

[marcopeix]

Yes, you can. Just make sure to pass two things:

• df containing the input sequence preceding your forecasts
• futr_df containing the future values of the exogenous features listed in futr_exog_list over the entire forecast horizon. For example, if you forecast 12 steps ahead, you need the 12 future values of your futr_exog_list.

This code runs:

AirPassengersPanel_train = AirPassengersPanel[AirPassengersPanel['ds'] < AirPassengersPanel['ds'].values[-12]].reset_index(drop=True)
AirPassengersPanel_test = AirPassengersPanel[AirPassengersPanel['ds'] >= AirPassengersPanel['ds'].values[-12]].reset_index(drop=True)

lstm = LSTM(
    h=1,
    input_size=12,
    futr_exog_list=["trend"],
    max_steps=10
)

nf = NeuralForecast(models=[lstm], freq="ME")
nf.fit(df=AirPassengersPanel_train)
nf.save(path="./models", overwrite=True, save_dataset=False)

nf2 = nf.load(path="./models")

futr_exog_df = AirPassengersPanel_test[["unique_id", "ds", "trend"]]
preds = nf2.predict(df=AirPassengersPanel_train, futr_df=futr_exog_df)
preds

[zhizhuaa]

Thanks! It works. I would like to ask another question. In this link #1354 (comment), you said that the model will automatically take into account past values. Do you mean it will only consider the past y values? If we need to consider other past values separately, do we need to add them to the hist_exog_list variable?

[marcopeix]

If you don't use any exogenous features, it will only use the past values of y. If you specify hist_exog_list it will consider the values of any features in hist_exog_list on top of the past values of y.