Foresight engine
How Horizon computes a signal while it is still forming
A language model answers by continuing text: it returns the most probable version of what has already been written. A weak signal is the opposite. Few people have written about it yet, so a system built to reproduce the common view smooths it away first.
Horizon keeps the two jobs apart. The language model reads your question and writes the report. The numbers come from computations on observed data or from a cited source, and a figure with neither behind it is marked as unconfirmed.
A signal starts as a change in attention
Before anything shows in sales, it shows in what people search for and talk about. The engine measures the direction of that attention and whether the latest surge is unusual. Readings taken from open conversation carry the lowest evidence grade and a plain label: a leading signal, not sales.
Season is removed before anything is called new
Most peaks return every year. The engine splits a series into season and trend, so a peak that comes back each November is read as the calendar, not as news.
An event is a break with the series' own norm
Each observation is compared with what the series' own trend and season predict. Only breaks strong enough to pass a false-alarm check are reported. The engine also tests whether one series moves ahead of another, and says plainly that a lead is not a cause.
Event or shift
A fad is carried by an event, a trend by a shift. The engine tests whether a trajectory is still growing in a straight line or already bending towards a ceiling, and where it sits on its life cycle. When the history is too short to tell, the answer is “not identified”.
The cost of being wrong
A forecast is a range, never a single point. The engine also shocks the trajectory to show what a sudden break would cost.
A forecast has to beat a naive one
Every forecast is replayed on the series' own past against the simplest rule there is: same as last time. The verdict is validated, not proven or no skill, and it is printed next to the forecast.
Memory is not evidence
Historical precedents named by the language model are labelled as unverified. They stay out of every calculation unless each one comes with a source.
What reaches you
Each figure with its source and an evidence grade: A for your own data, B for a measured external series, C for an estimate built on assumptions.
The method is open
Put it to work: ask Horizon a question about your own market and read the calculation behind the answer.
Open ANI Horizon