Add Developments folder via upload

[facundoallia]

Developments folder contain Ensemble notebook

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

MMenchero commented on 2022-11-09T16:54:13Z
----------------------------------------------------------------

Suggested title: How to do an ensemble model for time series forecasting.

StatsForecast allows you to create ensemble models in a very easy way. First we need to import the data that we are going to use, in this case the M4 Dataset. After that, we´ll to (remove "to") generate the forecasts via the generate_forecast() function. Finally, we´ll create the ensemble models using the forecasts previously generated. In this notebook we’ll implement and benchmark a ensemble model of AutoARIMA, AutoETS and AutoCES:

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

MMenchero commented on 2022-11-09T16:54:13Z
----------------------------------------------------------------

Let's try to summarize the information in a table

|Frequency| Min # observations training set| Forecasting horizon|

|Yearly | 13 | 6 |

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

MMenchero commented on 2022-11-09T16:54:14Z
----------------------------------------------------------------

Give a brief introduction to this section, saying that we'll now generate the forecasts and the ensemble for every frequency.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

MMenchero commented on 2022-11-09T16:54:15Z
----------------------------------------------------------------

Use overall instead of total.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

MMenchero commented on 2022-11-09T16:54:16Z
----------------------------------------------------------------

Expand the conclusions and add the table comparing our results with the other competitors. Mention the ease of use of StatsForecast for generating multiple models in one go and for creating the ensembles. 

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

MMenchero commented on 2022-11-15T07:20:37Z
----------------------------------------------------------------

Hay que checar esta tabla. Para el Naive2, se obtuvieron los siguientes valores:

• sMAPE = 13.564
• MASE = 1.912

Entonces

OWA = 1/2*( sMAPE_Nixtla/sMAPE_Naive2 + MASE_Nixtla/MASE_Naive2) = 0.841 != 0.853

Tomando los valores del Naive2 de arriba, para el primer lugar de M4 obtenemos

OWA = 1/2*(11.374/13.564+1.536/1.912) = 0.8209

Este es el valor que aparece en la tabla de resultados

https://www.sciencedirect.com/science/article/pii/S0169207019301128

Sugiero también agregar a los autores de los métodos ya que hay varias afiliaciones que aparecen como Individual.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

MMenchero commented on 2022-11-18T06:27:51Z
----------------------------------------------------------------

Line #1.    # Get trarin dataframe:

Typo: Train. Seems to be in other frequencies as well.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

MMenchero commented on 2022-11-18T06:27:52Z
----------------------------------------------------------------

I think we can compute the accuracy of every frequency and model more efficiently since the code for it seems to be the same, just with different names. One way of doing this is with this function. The only argument it takes is the name of the frequency.

def compute_accuracy(freq): 
   
  """""Computes the accuracy of a given frequency and model"""
   
  data = pd.read_csv(freq.lower()+'_df_forecast.csv')
  data = data.drop(columns=['Unnamed: 0']) # add 'Unnamed: 0.1' if requiered
  data['ds'] = data.groupby('unique_id')['ds'].transform(lambda data: np.arange(1, len(data) + 1))
  data = pd.melt(data, id_vars=['unique_id', 'ds'], var_name='model')
  data = pd.pivot(data, index=['unique_id', 'model'], columns='ds', values='value').reset_index()
  models = ['Ensemble_median', 'AutoARIMA', 'ETS', 'CES', 'AutoTheta']

  res = {}
  for k in range(len(models)):

    df_mask=data['model'] == models[k]

    df = data[df_mask]

    df.rename(columns={'unique_id': 'id'}, inplace=True)

    df = df.set_index(df['id'])

    df = df.drop(['id', 'model'], axis=1)

    y_metrics = M4Evaluation.evaluate('data2', freq, df.sort_values('id').values)

    res[models[k]] = y_metrics

   

  metrics = [res[k] for k in models]

  total_metrics = pd.concat(metrics)

  total_metrics['model'] = models 

   

  return total_metrics 

We can call it using

compute_accuracy('Yearly')

This should produce the same table as above.

I think the following are good practices to keep in mind: 

- Keep the variable names as short as possible. 

- If you need to do a process more than once, write a function for it if possible. 

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}