Forecasting playground
A live demonstration of the Laplace algorithm — so named because it places a weak prior over trees of transforms and lets the data speak.
observation
1-step mean
1-step 95%
k-step forecast fan
—
The paths that fit
— each transform composition (y → transforms → leaf) judged by its full predictive likelihood, K steps ahead
The model sees one observation at a time and emits a full predictive distribution for each of the next k steps. The dashed orange fan is that k-step forecast projecting from the latest point.
The bars above are live: laplace searches a
tree of transform compositions — every candidate is a path
y → [transforms] → leaf (e.g. seasonal → EMA,
fractional differencing → EMA, Yeo-Johnson coordinate,
OU reversion).