This repository is designed to make it as easy as possible to add new model or library. First, look at the structure of the project. It includes several levels of abstraction. If you want to add a model, you need to look at the folder with models (obligatory) and serializers (optional).
Forecaster is a class that generates time series forecasts.
Example of implementation: ARIMAForecaster
You will need to implement all the methods listed below.
**params - parameters for model or library you want to set.
Fit model for univariate time series forecasting
historical_values - historical data for model fitting. Example of expected table:
| datetime | value |
|---|---|
| 01-01-2022 | 254 |
| 02-01-2022 | 223 |
forecast_horizon - number of elements to predict.
**kwargs - additional parameters.
The method should allow to make a prediction for the future on the basis of historical values. The forecast horizon will not differ from that during fit stage.
historical_values - pandas Dataframe with historical values that is the last
known historical values at the time the forecast was generated.
forecast_horizon - number of elements to predict.
**kwargs - additional parameters.
The method returns a special dataclass ForecastResults. Check the fields in this class,
in addition to predictions as a numpy array, it is able to pass any additional information
you want to keep during the experiments (it can be handled in the serializer).
Fit model for multivariate time series forecasting
historical_values - historical data for model fitting. Example of expected table:
| datetime | target_ts_1 | ts_2 |
|---|---|---|
| 01-01-2022 | 254 | 123 |
| 02-01-2022 | 223 | 154 |
where 'target_ts_1' is target column for forecasting and other columns are exogenous variables.
forecast_horizon - number of elements to predict.
target_column - name of target column to predict.
predictors_columns - name of additional columns which can be used as exogenous time series and target series.
**kwargs - additional parameters.
Perform forecasting with several time series.
historical_values - pandas Dataframe with historical values that is the last
known historical values at the time the forecast was generated.
forecast_horizon - number of elements to predict.
target_column - name of target column to predict.
predictors_columns - name of additional columns which can be used as exogenous time series and target series.
**kwargs - additional parameters.
You should then think of a name for your model and
add to the dictionary in the Validator class.
Serializer is a class for saving the results of experiments.
Example of implementation: FedotSerializer
It is not necessary to implement this class. For new models, the default serializer will automatically run and save the model predictions. But if you want to store additional information (for example), such as a model description, you can add own implementation.