Engineers using ModelMill to train Logic-Based Networks for time-series forecasting can now validate model accuracy directly from the Training Results page — without leaving the platform or exporting data to external tooling.
What's new
The Forecasting Backtest block is a new module within the Training Results page. It runs a rolling back-test against a held-out window of your training data and surfaces accuracy metrics for each model trained in a given session.
For each LBN trained for a forecasting objective, the block shows:
- MAE (Mean Absolute Error) — the average magnitude of prediction errors across the back-test window
- MAPE (Mean Absolute Percentage Error) — percentage-based error, useful for comparing across different scales
- RMSE (Root Mean Squared Error) — penalises larger errors more heavily, giving a clearer signal on outlier performance
- Back-test window — the date range used for the held-out evaluation, derived from your dataset configuration
Who this is for
This module is available to any engineer using ModelMill to train LBNs against a forecasting or time-series prediction objective. If your training session includes at least one model with a forecasting target, the Forecasting Backtest block will appear automatically in Training Results.
How to use it
- Open a completed training session from the Training Results page.
- Scroll to the Forecasting Backtest block — it appears below the primary results summary.
- Select the model you want to evaluate from the model selector at the top of the block.
- Review the back-test metrics and the prediction vs. actuals chart for the held-out window.
- Use the Compare toggle to overlay results from multiple trained models side by side.
Notes
The back-test window is determined by the evaluation split configured in your dataset. If no explicit evaluation split was set, ModelMill defaults to holding out the final 20% of the time series for back-testing.
Back-test results are saved alongside the training session and can be revisited at any time from Training History.