Series
Residuals
in-sample errors on the training window. They should look like noiseResidual autocorrelation
bars outside the band = structure left on the tableSearch landscape
Surface height · validation MAE
Trials · order tried
bestsampled on X–Yfailed
Dots are the configurations actually tried, placed at their real score. They also differ in the parameters not shown, so they can sit above or below the surface. Drag to rotate; click a dot to load it.
Explored values, every parameter
each dot is a trial: position = value tried across the full allowed range, height = validation MAE (lower is better), colour = order tried. Black line = best value, blue mark = current setting.Feature importance
share of total split gainError vs. number of trees
train validation (one-step)
04
Score on the hidden data
The chart above forecasts from the end of your training window. Scoring keeps your parameters fixed, lets the model also read the validation zone, then forecasts the hidden last 20% and compares it with the real values. Score = mean absolute error there (lower is better).