A single score hides why a model wins or fails.
High dimensional forecasts need inspecting across related series, time windows and covariates, not just aggregate metrics.
-
“Which model wins, and where?”
Side by side comparison
DeepGLO, DeepGLOLite, ACDeepGLO and WeightedDeepGLO on the same series and time window.
-
“Why did this forecast fail?”
Diagnostics by inspection
Hover and click to see how forecasts behave around regime changes, peaks and discontinuities.
-
“Which series move together?”
Dependence made visible
Similarity is computed using covariance, and the most similar series appear as covariates and context plots.
Accuracy is only part of the story.
What the experiments showed.
Run on benchmark datasets with hundreds of correlated time series.
- ACDeepGLO, notably with about 10 covariates, tends to perform best on global and aggregate loss metrics when trained on the full dataset.
- WeightedDeepGLO underperformed relative to DeepGLO in these experiments.
- Under batch wise training, ACDeepGLO and DeepGLO appear more similar, which shows how much training protocol matters in comparative studies.
How it’s built
A React and D3 client with one Flask endpoint that trains, validates and returns predictions. Rolling validation (τ=14, n=4) gives a 56 step horizon. Datasets: Traffic, Electricity, PEMS and Wiki.
Why it matters for your AI
Evaluation becomes more useful when you can inspect behavior across slices, not just aggregate scores. The same discipline sits behind a measured failure rate.
Tell me where your AI’s average score is hiding the failures.
The first 30 minute call is free, and I reply within one business day.