# Trend Analysis > Separates durable trends from noise using a baseline, seasonal decomposition, and a signal test requiring persistence, magnitude beyond 2 standard deviations, 3+ independent corroborating data points, and a plausible mechanism - then characterizes the trend's stage and forecasts trajectory with a confidence range. Use when someone asks "is this a real trend or a blip", "where is this metric or market heading", "should we bet on this shift", or must defend a forecast to stakeholders. Do NOT use for scouting what content formats are trending on Instagram - use instagram-trend-scout instead; for tracking brand mentions and sentiment shifts in news and social - use media-monitor instead. - Skill: `skillmedev/trend-analysis` (Agent Skill) - Install (CLI): `npx skillmds@latest add skillmedev/trend-analysis` - Raw SKILL.md: https://api.skillmd.com/api/skills/skillmedev/trend-analysis/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: SkillMedev (https://skillmd.com/u/skillmedev) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/skillmedev/trend-analysis --- # Trend Analysis Distinguish durable trends from temporary fluctuations and project where they are headed. The hardest part is not spotting movement - it is deciding whether movement matters. The costly failure this skill prevents is the strategy built on a spike: a seasonal bump or a hype-cycle blip read as a secular shift, funded for three quarters, then quietly unwound. ## Inputs to collect 1. **The series or phenomenon** under study, and the metric that operationalizes it. 2. **History**: at least 2 full seasonal cycles of data where cycles exist (2 years for annual seasonality); with less, say the seasonality adjustment is unreliable and lower confidence. 3. **Candidate corroborating sources**: related metrics, independent datasets, external indicators. 4. **The decision at stake** and its horizon - a 6-month product bet and a 5-year infrastructure bet tolerate very different forecast uncertainty. 5. Defaults where the user has none: baseline window of 8+ periods, forecast horizon no longer than one-third of the observed history. Label assumed values as guesses. ## Operating procedure ### Step 1: Establish the baseline Before calling anything a trend, define what "normal" looks like: compute a baseline level and expected variance over a relevant window. A data point is only notable relative to this. No baseline, no trend claim. ### Step 2: Decompose the series Separate the data into components: - **Trend**: the long-run direction. - **Seasonality**: predictable recurring cycles (weekly, quarterly, yearly). - **Noise**: random, unexplained fluctuation. Many "trends" are just seasonality or noise. Adjust for seasonality before concluding anything - a December spike in a retail metric is not a trend, it is December. ### Step 3: Apply the signal-vs-noise test Treat a movement as signal only if it passes the gate. The first three criteria are mandatory; mechanism is required before the trend can drive a decision: - **Persistence**: sustained across multiple consecutive periods (3+ for the direction to be meaningful), not a single spike. - **Magnitude**: exceeds normal variance - more than 2 standard deviations from the baseline. - **Corroboration**: visible in **at least 3 independent data points or sources** before the movement is called a trend. Independence matters: three dashboards fed by the same pipeline are one source; a usage metric, a survey result, and an external market indicator are three. - **Mechanism**: a plausible causal story for why the movement exists. A statistical pattern with no mechanism is a candidate, not a conclusion - it may still be watched, but not bet on. A movement that fails persistence and corroboration is noise. Log it and move on; do not promote it because it is exciting. ### Step 4: Characterize confirmed trends For each survivor of the gate, describe: - Direction and rate of change - and whether the rate is linear or exponential (plot on a log scale to check; eyeballing raw exponentials misleads). - Acceleration or deceleration. - Stage: **emerging** (few adopters, high variance), **growing** (accelerating adoption), **maturing** (decelerating growth), **declining**. The stage sets the strategic clock - an emerging trend rewards early positioning; a maturing one rewards efficiency plays. - Underlying drivers, and whether each driver is strengthening or weakening. ### Step 5: Forecast the trajectory - Project forward with the simplest model that fits the decomposed trend; do not over-fit - a model with more parameters than the data supports predicts the past, not the future. - Give a **range, not a point estimate**, and state confidence. - Identify inflection risks that could bend the curve: saturation (adoption ceilings), substitution, regulation, backlash. - Distinguish interpolation (safe) from extrapolation far beyond the observed data (risky). Rule: do not extrapolate further forward than one-third of the history observed without flagging the forecast as speculative. - State the horizon explicitly - confidence decays sharply with distance. ### Step 6: Report with the invalidators Deliver the trends found, the evidence per trend, the forecast range, the key assumptions, and - critically - the observable events that would invalidate the forecast, so the reader knows what to watch rather than when to blindly trust. ## Worked artifact: trend-evidence table Claim under test: "Developer adoption of AI coding assistants is a durable growth trend, not hype." | Test | Evidence | Independent? | Pass | |---|---|---|---| | Persistence | Weekly-active usage up 8 consecutive quarters in vendor telemetry | - | Yes | | Magnitude | Growth rate 4.1 SD above the pre-period baseline for dev-tool adoption | - | Yes | | Corroboration 1 | Vendor-reported paid-seat growth across 3 competing products | Yes - commercial data | Yes | | Corroboration 2 | Independent developer survey: majority report weekly use, up from a small minority two waves earlier | Yes - survey, different collection method | Yes | | Corroboration 3 | Job postings requiring assistant proficiency rising across two job boards | Yes - labor-market exhaust data | Yes | | Mechanism | Falling inference cost + measured task-time savings give users and buyers a sustained reason to adopt | - | Yes | Verdict: signal - 3 independent corroborations from different data-generating processes, persistent, large, with a mechanism. Stage: growing. Invalidators to watch: plateau in survey wave-over-wave usage, seat-growth deceleration across 2+ consecutive quarters, or a cost/regulation shock to the mechanism. ## Deliverable Produce a trend report containing: the baseline and decomposition summary, a filled trend-evidence table per candidate trend, the verdict (signal / noise / watch) with the failed criteria named for rejects, stage and drivers for confirmed trends, a forecast range with horizon and confidence, and the explicit list of events that would invalidate each forecast. ## Do NOT - Do not call a trend from fewer than 3 independent data points - one source repeated is one data point, and independence is what kills most trend claims. - Do not skip seasonal adjustment; a recurring cycle read as a trend is the most common false positive. - Do not treat a co-moving metric as a driver - correlation is not causation, and a trend "explained" by a fellow-traveler has no mechanism. - Do not issue point forecasts; ranges with stated confidence, always. - Do not extrapolate past one-third of the observed history without labeling it speculative. - Do not ignore survivorship and selection bias in the underlying data - trends measured only on survivors always look better. - Do not let hype cycles inflate early signals; discount novelty-driven spikes and demand the corroboration anyway. ## Quality bar - Every trend claim shows a filled evidence table with all four tests, and the corroboration row counts genuinely independent sources. - Rejected candidates are listed with the criterion they failed - silent drops hide the base rate. - The forecast has a range, a horizon, a confidence statement, and named invalidators. - A skeptical reader could re-run the signal test from the evidence cited.