baseline problem

[ICDI0906]

Hi~ I have some question, please help me! for the baseline model ARIMA, how to split the data for multi-step prediction?,is there validation data? and how to train ARIMA? does it use only the past 1 hour's data? for a sensor, is there only one ARIMA model to make prediction? looking forward to your reply~

[still2009]

Hi~ I have some question, please help me! for the baseline model ARIMA, how to split the data for multi-step prediction?,is there validation data? and how to train ARIMA? does it use only the past 1 hour's data? for a sensor, is there only one ARIMA model to make prediction? looking forward to your reply~

Hi, I am also curious about this question, I wonder if you finally solved this. Since the results for different horizon differs in Table 1 of raw paper. The predcition is not made by simply invoking forecast function with steps=n_test, I guess its probabily similar to the way adopted in eval_var function in eval_baseline_methods.py (as follows).

# Do forecasting.
    result = np.zeros(shape=(len(n_forwards), n_test, n_output))
    start = n_train - n_lags - max_n_forwards + 1
    for input_ind in range(start, n_sample - n_lags):
        prediction = var_result.forecast(scaler.transform(df.values[input_ind: input_ind + n_lags]), max_n_forwards)
        for i, n_forward in enumerate(n_forwards):
            result_ind = input_ind - n_train + n_lags + n_forward - 1
            if 0 <= result_ind < n_test:
                result[i, result_ind, :] = prediction[n_forward - 1, :]

Here, the forecast function of var_result accepts a parameter of prior values, which is natural to make Seq2Seq prediction. However, such case didn't hold for the ARIMA and SARIMA models (doc link for VAR model, doc link for ARIMA model)

But I found apply and extend function for ARIMAResult, I wonder whether the mutli-step forecasting is made by continually extending the result object with the size of horizon?

Some other issues were also mentioned related to the question about reproducing ARIMA (#28, #12), but they were not fully addressed. I would be very thankful if dear author @liyaguang could reply; As a high impact work for traffic flow prediction, this would really help the community to push things forward.