Trend Spotter

Two points is not a trend. The skill requires at least three consistent data points before calling something a trend, and five before applying linear regression. Seasonality decomposition needs at least two full cycles. Correlation analysis needs at least twelve paired observations. Forecasts beyond three periods get wider confidence bands; beyond six periods, they become directional language only. If the data is insufficient for what's being asked, the skill says what's needed and refuses to stretch thin data into false conclusions.

Decompose before concluding. The most common analysis mistake is calling a December spike a "trend" when it's Christmas, or panicking at a January dip that's the same dip as every January. The skill separates the data into trend, seasonal, and residual components before reporting anything as a directional finding. A "growing 15%" claim that disappears when you remove seasonality wasn't growth; it was the calendar.

The DRIFT framework runs the analysis. Direction (up, down, flat over the window, with the slope calculated). Rate of change (is the trend accelerating or decelerating — the second derivative often matters more than the first). Irregularities (outliers beyond two standard deviations, structural breaks where the pattern shifts permanently, missing data gaps). Factors (correlations and leading indicators, where one metric moves one or two periods before another, with correlation strength and lag reported but never overclaimed as causation). Trajectory (where this is heading if the trend continues, with optimistic, baseline, and pessimistic ranges).

Forecasts come with the assumptions that have to hold for them to be useful. Point estimate, range, confidence level (GREEN/AMBER/RED based on data quantity, stability, and how far out the projection runs), the explicit assumptions, and an invalidation signal — what would tell you the prediction is wrong before the forecast horizon closes. Confidence interval language gets translated for non-technical readers ("the metric will likely land between X and Y") and traffic-light framing replaces statistical jargon. A point estimate without bounds creates false precision and never appears alone.

You get the data overview with sufficiency notes per analysis type, the baseline (central tendency, variability, seasonality pattern), the trends detected and ranked by magnitude and acceleration (not just whatever's loudest), the classified anomalies with type (spike/drop/shift/trend break/missing) and risk level, correlations with lead/lag relationships flagged for further investigation rather than concluded as causal, the forecast table across 3/6/12 periods with confidence, and a recommended-actions list keyed to the detected patterns. When the user provides many metrics, the top three to five trends become headlines and the rest go to a summary table — nobody parses fifteen trend descriptions to find the three that matter.