-`symbol`: Used primarily as the subfolder name for all generated files within the `data_folder`.
-`description`: A textual description of this configuration. It can be consumed by various pluggable components, such as text or visual output adapters.
-`freq`: Data frequency in `pandas` format (e.g., `1h` for hourly data or `1min` for minutely data).
-`train`: Boolean flag specifying whether the analysis runs in train (`true`) or predict (`false`) mode.
If `true`, all trainable features will fit their models using historical data.
If `false`, existing models will be loaded for prediction.
During analysis, all data is represented as a DataFrame configured via the following parameters.
These parameters define the required shape of the DataFrame and are used primarily by the server:
-`label_horizon`: The minimum number of future rows required to compute a label (i.e., the prediction horizon).
During training, rows lacking sufficient future data are excluded.
-`features_horizon`: The minimum number of past rows required for a feature to be valid. For example, computing a 10-day moving average requires 10 previous rows.
Consequently, the first `features_horizon` rows are considered invalid and excluded from analysis.
This value should be derived from the feature definitions.
-`train_length`: Default upper limit for the training dataset size.
This serves as a global maximum for all ML features, though individual features may override it.
A value of `0` indicates that all available data should be used.
-`predict_length`: The minimum number of rows kept up-to-date and valid in online mode.
To ensure validity, these rows must be preceded by at least `features_horizon` historical rows.
-`append_overlap_records`: In online mode, the server requests additional records beyond those strictly missing.
This parameter specifies the size of that overlap buffer.
Overlapping records are re-evaluated and overwrite previous values, which helps mitigate connection errors or minor discrepancies in recently received data.
