Eternet Portal Documentation

Forecasting and Evaluation

Shared data foundation

Forecast reuses Analytics queries and datasets rather than owning a separate datasource connector system. When given an Analytics reference, the Forecast Service asks Analytics to resolve the authorized source and return Arrow IPC data; the Analytics service owns source resolution and query processing.

Forecast workflow

The workspace supports choosing a saved Analytics query or dataset, preparing input columns and series, configuring a forecasting or evaluation mode, selecting/reviewing models, running the pipeline, and inspecting predictions and metrics. CSV input is also supported as a separate import path.

Outputs

Forecast responses can include predictions, model status, evaluation metrics such as MAE/RMSE/MAPE, and anomaly results depending on the selected workflow and backend response. Saved forecasts can be listed and exported. Prediction intervals/uncertainty depend on model capability and the returned output.

Boundary and requirements

Analytics remains responsible for datasource access and query results; Forecast is responsible for preparation and modeling. A successful run requires an accessible saved source, valid time/value columns, sufficient observations, and a reachable Forecast Service. Do not treat a model recommendation as a trained model until a run completes successfully.

Preparing a trustworthy forecasting dataset

Forecast quality starts before model selection. The input must contain a usable time column, a target value, and, when applicable, a series identifier that separates independent entities. The operator should confirm the time zone, sort order, sampling cadence, missing intervals, duplicate timestamps, and numeric interpretation of values. Data preparation may include selecting the relevant period, choosing a series, handling missing values, and deciding whether additional regressors are available. A model can produce a result from an unsuitable dataset, so successful execution is not equivalent to a meaningful forecast. The validation step should compare the model output with the source history and check whether the selected horizon is operationally reasonable.

When the user selects a saved Analytics query or dataset, Forecast requests that source through the Analytics-owned boundary. The Analytics service applies its datasource permissions, resolves the stored definition, and returns Arrow IPC data; Forecast converts that result at the modeling boundary for its supported pipeline. This arrangement keeps connectors and source authorization centralized instead of duplicating them in Forecast. Direct CSV import remains a separate path for data that is not already stored as an Analytics asset. The customer should know whether the result came from a governed shared dataset or a one-time file upload.

Models, evaluation, and uncertainty

The workspace can inspect model metadata and configure forecast or evaluation workflows. Recommendations are suggestions based on available data and model capabilities; they are not evidence that a model has been trained for production. A run should be reviewed for its status, selected model, forecast horizon, prediction values, and evaluation output. Backtesting or cross-validation helps estimate how a model would have performed on withheld historical intervals. Metrics such as mean absolute error, root mean squared error, and percentage error communicate different kinds of mistakes and can behave differently when values are near zero.

Forecast ranges and uncertainty are not automatically available for every model. They depend on the chosen algorithm and the response fields the backend returns. Present intervals only when they are actually present, and explain whether they are prediction intervals, confidence intervals, or another model-specific quantity. Saved results should retain enough context—source reference, target series, time range, model, and run status—to make a later review interpretable.

Using forecast output elsewhere

Forecast results are tabular time-series outputs and can be exposed back to Analytics so existing result and visualization components can display actual and predicted values together. This makes the prediction part of the broader analytical workflow without making Analytics responsible for model training. Before showing a forecast on a dashboard, verify that the saved result is accessible to the intended audience, its timestamps align with the actual series, and its model/version metadata can be understood. Customer value comes from informed planning and evaluation—not from a prediction presented as certainty.