A simulated monthly metric (with configurable trend, seasonality and noise) is rendered as a bar chart in 3D. A vertical divider marks "today": everything to the left is training data the model can see, everything to the right is a hidden future shown only faintly so you can judge each method's accuracy. Pick a model to fit the same training bars and project a forecast tube whose radius encodes the width of the prediction interval.
Tree ensembles like random forests and gradient boosting routinely beat linear models on messy, non-linear tabular data — but because every leaf predicts an average seen during training, they are structurally unable to extrapolate a trend past the range of their training data, unlike regression or trend-aware time-series models.
A 3D bar chart of a simulated monthly metric, split into a visible training range and a faint hidden future, with a live forecast tube whose color, curve shape, and error-band radius change as you switch between regression, time-series and tree-ensemble models.
The same training bars can be fit three different ways: a straight regression line, a trend-plus-seasonal decomposition, or a bucketed tree ensemble. Their forecasts diverge most sharply in how each one extrapolates trend and seasonality beyond the training range.
Pick a model, then adjust noise, trend strength, seasonality and forecast horizon. Watch the prediction tube and its translucent error band reshape, and compare the forecast RMSE against the hidden future bars.
Tree-based ensembles like random forests can't produce values outside the range seen during training — their forecasts flatten out beyond "today," which is why practitioners often pair them with a trend-aware model when the future is expected to keep climbing or falling.