quickstart

Date: Nov 23, 2021 Version:

divina in 5 minutes

In order to perform your first forecast with divina,

The aim of divina is twofold:

  1. to reduce the complexity of configuration for causal forecasting at scale. This is accomplished by abstracting all configuration to a single JSON file that lets users configure new experiments easily and safely.

  2. to deliver scalable and bidirectionally interpretable causal forcasting models. These models bring transparency and incremental control to the forecasting process using a variety of coefficient calculation tools, binning and interacting of features and set of link functions that enable a linear model to fit various target distributions.

Experiment Definition

A minimal experiment definition supplies the path to the dataset, the name of the column that holds the time index and the name of the column that holds the target to predict. Below is a minimal example using the retail sales data included with divina.

Experiment Persistence

Experiment artifacts are persisted either locally or to S3 depending on the use of the –local flag when running the experiment command and will produce a local output structure as shown below:

experiment path
  |- models
  |    |
  |    \- h_{forecast horizon}
  |           |-fit_model.joblib
  |           |-bootstrap
  |                |
  |                |- bootstrap_model_{random seed}
  |
  |- forecast
  |    |
  |    |- common_meta.parquet
  |    |- forecast_partition_0_meta.parquet
  |    |- forecast_partition_0.parquet
  |    \  ...
  |
  |- validation
       |
       |- metrics.json
       \- {validation split}
              |
              |- validation_partition_0_meta.parquet
              |- validation_partition_0.parquet
              \  ...