Trends, Forecasts & Outliers

Run a model over the data already on a visual, without moving it anywhere else.

PersivX ships four model types that attach directly to a visualization. They read the visual's existing result — its X axis as the independent variable and its Y axis as the dependent one — so there is no separate dataset to prepare.

ℹ️ Models Are Saved With the Visual A model you add becomes part of the visual's definition. It re-runs whenever the visual refreshes, so a forecast stays current as the data behind it moves.

Trend Line

Fits a line through the visual's existing points.

Model	Use when
Linear	The relationship is roughly a straight line.
Polynomial	The shape curves — growth that accelerates, or a rise and fall.

Only one trend model applies at a time; choosing a different one replaces it.

Forecast

Projects the series forward beyond the last point you have.

Setting	Meaning
Model	Prophet — recommended for seasonally changing data.
Create forecast for (periods)	How many periods ahead to project. Defaults to 30.

Only the points beyond your last actual observation are shown as forecast, so the projection is visibly distinct from history. PersivX also generates a short written summary of what the forecast says.

⚠️ Forecasts Need a Time Axis and Enough History The independent variable has to be a date, and a handful of points won't support a meaningful projection — least of all a seasonal one, which needs enough history to have seen the season before.

Clustering

Groups records that resemble each other across the attributes you choose.

Setting	Meaning
Model	K-Means or DBScan.
Number of Clusters	Leave blank for automatic, or set a specific count.
Attributes	Which columns the clustering considers.

💡 Choosing Between K-Means and DBScan K-Means asks how many groups you want and assigns every record to one. DBScan finds groups by density instead — useful when you don't know the number in advance and don't want unusual records forced into a cluster where they don't belong.

Outlier Detection

Flags records that don't fit the pattern.

Setting	Options
Model	Time Series Anomaly for a series over time, or Isolation Forest for multi-attribute records.
Get outliers in	Most Likely — return the N most anomalous records. Percentile — return everything past a threshold.
Show outliers (most likely)	How many to return in Most Likely mode. Defaults to 3.
Attributes	Which columns are considered when judging how unusual a record is.

💡 Most Likely vs. Percentile Most Likely gives a fixed-size list — right for a review queue someone works through. Percentile gives a variable-size list based on a threshold — right for monitoring, where a quiet week should produce fewer flags, not the same number regardless.

Combining Models

A visual can carry more than one model type at once — a trend line and outlier detection on the same chart, for instance. Adding a second model of the same type replaces the first.

⚠️ Models Follow the Filters Models run over the data the visual currently holds, after filters. Change a filter and the model is refitted on the narrower set — which is usually what you want, but means a forecast a reader sees may not be the one you fitted.