Scope

What autonomous univariate distributional forecasting is for, and the objections it is not refuted by.

What it is

A skater turns a single numeric stream into a full predictive distribution, online, automatically, in Python or the browser. Not a point and an error bar bolted on afterwards. A distribution: a mean for an estimate, a spread for risk, tails for sizing, and a density for scoring and combination. No human tends any one series, and nothing is configured per series.

What it is good for

A distribution, not a point, is what downstream decisions actually consume. Risk limits read the spread, threshold alarms read the tail probability, position sizing reads the whole shape, simulation reads draws, and a likelihood-based evaluation reads the density at what happened. A skater supplies all of these for any stream it is pointed at.

It is built to run unattended and at scale: thousands of streams, updated on every observation, on a server or in a browser tab, with no fitting job to babysit. That makes it a good baseline, a good prior, and a good monitor — the always-on distributional layer under a system that has better things to spend attention on.

Anomaly detection falls out for free. Because every prediction is a density, every arriving point gets a probability integral transform against the forecast that was made for it: the state carries state["pit"] (roughly Uniform when the stream is behaving) and state["z"] (roughly N(0,1)) at every horizon 1..k, resolved online with no extra compute. A |z| of five is an anomaly by construction — scored against a model that has already adapted to the stream's drift, volatility clock, seasonality, and coordinate, which is precisely what separates an anomaly from a regime the model should have learned. Multi-horizon agreement (the same point surprising at h = 1, 5, and 20) separates one-off spikes from genuine breaks. See the anomaly-detection skill.

The objections it is not refuted by

“But you need more data, or exogenous features.” Yes, and univariate forecasting never claimed otherwise. Real series are driven by outside data, and a serious model brings it in. Univariate distributional prediction is a component, not a pipeline. It is what you run on the series you have, and it plugs in before, during, or after the rest of the modeling. A skater is a transformation by construction, so it composes into a larger system rather than standing alone; exogenous structure enters as another stage that conditions on it and refines the residual.

“You should just hire an analyst, or build a bespoke model.” For one series, sure. The point is autonomy and scale. A hand-built model per stream does not survive thousands of streams updated online, and a point forecast from it is not the object most decisions need. A skater is cheap, fast, dependency-free, and runs anywhere, which is exactly what you want for the long tail of series nobody will ever staff.

Chaining into larger systems

This is the point the Microprediction book is built around. Forecasting can be organized as a supply chain: one stream's distributional output becomes the input, the residual, or the settled outcome of the next, each stage contributed by whoever is best at it. A univariate distributional forecaster is the base layer of that chain, the reusable density that residual markets, dependence models, and conditioning forecasters build on.

That is also why a skater is scored by held-out log-likelihood rather than by a threshold or interval metric. Likelihood is the score under which such refinements compose: the log-likelihood of a chained forecast is the base score plus the contribution of each stage. The reasoning is laid out in why skaters ranks by likelihood.

The short version

Univariate distributional forecasting is not the whole system. It is the honest, autonomous, composable base of one: a full distribution for every stream, at scale, ready to be chained.