paulozip / arauto Goto Github PK

At the moment the frequencies managed are:

• hourly
• daily
• monthly
• quarterly
• yearly
  It would also be useful to add the biannually frequency

Vector Autoregressions it's a great tool to analyze and forecast multivariate time series, chiefly when multiple time series influence each other. For example, one might want to forecast the Air Quality by comparing this time series with others like humidity, temperature, and tungsten oxide.

VAR models may be added to Arauto as an alternative to ARIMA models.

Yeah, we don't have tests for Arauto yet. Please, don't judge me.

Currently, Alchemy will try to find the transformation function (First Difference, Log transformation, Log Difference, Seasonal difference, etc.) that returns the lower score in the Dickey-Fuller Augmented Test.

It would be nice to let the user choose the proper transformation function for her problem. A dropdown menu should do the thing.

Autoregressive Conditional Heteroscedasticity (ARCH) and Generalized Autoregressive Conditional Heteroscedasticity (GARCH) are good models to analyze and forecast volatility in time series. Arauto might use ARCH and GARCH models to:

• Model volatility;
• Use it as a way to reduce errors in ARIMA models. e.g.: by modeling the volatility, we could use it as a variable for ARIMA e deduce the forecasting values;

Exponential Smoothing has been shown as a good model that is able to forecast non-seasonal and seasonal data. It could be used as an alternative to ARIMA models

It would be great with we could upload or read a dataset from S3.

Maybe adding the path to a bucket. Of course, the machine running Arauto should have the appropriate policies for this bucket.

Instead of use a cURL post to upload new datasets, it would be nice to have and interface where this could be done by the user.

It would be useful to integrate models from,
https://github.com/firmai/atspy

And the forecast combination algorithms from,
https://rdrr.io/cran/ForecastComb/man/auto_combine.html

Sometimes the generated code missing the line:
df = np.log1p(df)

This make some errors to forecasting with np.expm1() transformation.

This problem occurred with my dataset but it is possible to replicate with the "yearky_lynx_trapping.csv" dataset and FREQUENCY equals to Daily with default configuration.

In the end the Augmented Dickey-Fuller test stays like this:

# Applying Augmented Dickey-Fuller test
dftest = adfuller(df.diff().dropna(), autolag='AIC')

The test_stationarity function also contains the transformation function. We could split these functions. The transformation function should be a Class containing all the transformations, while the test_stationarity function would only handle the statistical test.