trend-analysis

name	trend-analysis
description	Turns a trend claim into an evidence-checked trend statement, quantifies its rate, curvature and S-curve position from a time series, then projects it — naive extrapolation, logistic fit, or event-adjusted Trend Impact Analysis (Gordon) — with an assumption sheet, counter-trend scan and signposts. Use when asked "is {X} actually growing or is it a fad?", "how far can this trend run?", "extrapolate this trend to 2030", "validate this trend claim" or "is {X} a megatrend?". Not for publication and citation curves (use `assess-research-momentum`) or for placing a technology on the hype curve (use `apply-hype-cycle`).
license	MIT
metadata	{"category":"foresight","method":"Trend analysis with Trend Impact Analysis (TIA)","origin":"Theodore J. Gordon (Trend Impact Analysis), The Futures Group, 1970s; Fisher & Pry substitution model, 1971; Rogers' diffusion curve, 1962/2003","version":"2.0.0"}

Trend Analysis

Trend analysis converts a claim that something is changing into a measured trajectory and, where evidence supports it, a projection whose assumptions are visible. Its core principle comes from Trend Impact Analysis, as Theodore J. Gordon set it out for the Millennium Project (2009): extrapolate the series surprise-free, then state the events that could bend it, each with a probability and an impact, so hidden assumptions become a challengeable sheet rather than a silent forecast. Two constraints keep projections honest: adoption saturates, so growth follows S-shaped rather than straight paths (Fisher & Pry, 1971; Rogers, 2003), and a claim is not a trend until multi-source evidence and a mechanism say so. It prevents projecting a one-source spike, and extrapolating a real trend past its ceiling.

When to invoke

Invoke when:

• A trend claim needs testing or projecting: "is {X} actually growing or is it a fad?", "how far can this trend run?".
• A number is needed for a plan and the series exists: adoption share, unit cost, penetration, price.

Do NOT invoke when:

• The series is publication or citation counts for a research field — use assess-research-momentum.
• The question is hype-narrative placement — use apply-hype-cycle.
• Fewer than about five comparable observations exist — say so rather than fit a line to three points.
• The system is reflexive or hinges on a discontinuity (elections, asset prices) — use scenario-planning.
• Nothing is measured yet — use horizon-scanning.

Procedure

Steps 3–4 are arithmetic: use the companion tool. Curve choice and caveats: references/projection-methods.md.

1 — State the trend as a falsifiable claim

Force the claim into {quantity} is {rising | falling | shifting} at {rate} in {population or geography}, horizon {year}. "AI is changing everything" is not analysable; "electric share of global new car sales to 2030" is. If quantity and population cannot be named, that statement is the deliverable. Output: trend statement.

2 — Validate the evidence: trend or fad

Gather series from at least two independent sources — official statistics, industry datasets, registries, studies — ideally spanning five years. Check the fad markers: single source, media-only coverage, spike-shaped series, no mechanism. Name the structural driver that would sustain the change (cost curve, demographics, policy, installed base). Verdict: VALIDATED, UNCERTAIN or FAD MARKERS; the latter two stop the analysis. Output: evidence table and verdict.

3 — Quantify rate, curvature and position on the curve

Compute absolute and relative change per period, then read curvature from the period gains: growing gains mean acceleration, constant gains a linear path, shrinking gains saturation. For adoption trends locate the position on the diffusion curve, first checking which quantity the series is. Rogers' adopter categories partition cumulative adopters, so his boundaries — below roughly 16 % cumulative adoption the innovator and early-adopter phase, 16–50 % the early majority, above 50 % later — can only be read off a cumulative-adoption series. An annual sales share, new-registration count or shipment share is a flow, and the 16 % marker does not apply to it: read a flow in the Fisher–Pry substitution frame instead, as the fraction f of new units the challenger has taken from the incumbent, whose f/(1−f) is linear in time while the substitution runs to completion. Output: trend profile with drivers and inhibitors.

4 — Project, with the assumption sheet attached

Choose a method and justify it: naive extrapolation for slow, structurally driven quantities; a logistic fit where a ceiling exists, stating that ceiling and the fit quality; Trend Impact Analysis where identifiable events could bend the curve — extrapolate the baseline, list events with probability and impact, show the arithmetic. Report one alternative shape. Output: projection and assumption sheet (drivers, ceiling, events).

5 — Scan for the counter-trend and collisions

Name the counter-trend the change provokes (globalisation and localism, digital and analogue revival) — discontinuities usually enter there — and any collision with other validated trends. Output: counter-trend note.

6 — Verify and emit the trend card

Recompute the rate and projection arithmetic from the quoted series, confirm every number traces to a dated source, then set signposts and a cadence. Output: trend card.

Output template

## Trend card — {trend name}

Trend statement: {quantity} is {direction} at {rate} in {population/geography}; horizon {year}.
Validation: {VALIDATED | UNCERTAIN | FAD MARKERS} — {sources, span, mechanism}

Rate: {absolute and relative change per period}
Curvature: {accelerating | linear | saturating} — {period-gain evidence}
Curve position: {early | mid | late}
Drivers: {forces}   Inhibitors: {frictions}   Counter-trend: {the opposition it generates}

Projection to {horizon}: {naive | logistic | TIA} → {value or range}; alternative shape: {value}
Assumptions: {drivers persist; ceiling; events with P × impact}
Signposts: {indicator} — breaks the projection if {condition}; review every {cadence}
Confidence: {low | medium | high} — {reason}

Mandatory fields: trend statement, validation verdict with sources, assumption sheet, signposts, confidence. A projection without an assumption sheet may not ship.

Worked example

Claim, mid-2025: "EVs are taking over — how far does this run?" Trend statement: electric vehicles (BEV + PHEV) as a share of global new car sales, rising, horizon 2030. Illustrative series, rounded from published EV outlook data: 4 % (2020), 9 % (2021), 14 % (2022), 18 % (2023), 21 % (2024). Validation: VALIDATED — five years of sustained rise, independent sources, a mechanism (battery cost learning plus fleet-emission mandates); not spike-shaped. Fits from scripts/trend.py describe --demo and fit --demo:

Change: 4.00 → 21.00 = +17.00 (+425.0%) over 4 periods
Rate:   +4.25 per period (average absolute change)
Curvature: saturating — period gains +5.00, +5.00, +4.00, +3.00
  naive     R² 0.9766   y = 4.00 + 4.25·(t − 2020)
  linear    R² 0.9898   y = 4.60 + 4.30·(t − 2020) — OLS slope +4.30 per period
  logistic  R² 0.9976   y = 22.84 / (1 + e^(−0.9708·(t − 2021.54))) — K 22.84, last = 91.9% of K

The series is a share of new sales, so it is a Fisher–Pry substitution fraction, not cumulative adoption, and Rogers' 16 % boundary does not read off it — cumulative EV share of the global fleet was far below 21 % in 2024. In the Fisher–Pry frame, f = 0.21 gives f/(1−f) = 0.27: the substitution is still early. The logistic ceiling of 22.84 % is an artefact of five points on one segment — the tool's own caveat — so the naive baseline is used and the fitted ceiling reported as the alternative shape. Baseline to 2030: 21 + 4.25 × 6 ≈ 46.5 %. TIA adjustment, reproduced by the tool below: subsidy rollback (P 0.5, −5), charging bottlenecks (P 0.4, −4), battery cost breakthrough (P 0.3, +3) → 43.3 %, against a logistic alternative near 23 %. Assumptions: no deep recession, mandates not repealed. Counter-trend: hybrid resurgence and pushback on phase-out dates. Signposts: annual share print within ±3 points of track; battery pack prices; the EU 2035 review. Confidence: medium — policy-driven, reflexive to election cycles.

Verification

• The trend statement names quantity, population or geography, rate and horizon.
• Two or more independent sources cover the series; each quoted number traces to a dated source.
• Rate, curvature and projection arithmetic were recomputed with scripts/trend.py.
• The chosen curve is reported with an alternative shape, its fit quality and the ceiling assumption.
• Every TIA event carries a probability and impact, and the adjustment arithmetic is shown.
• Signposts and a review cadence are set; an UNCERTAIN or FAD verdict stopped the analysis.

Companion tool

scripts/trend.py (stdlib only) does the arithmetic of steps 3–4 on a period,value series (--file series.csv|json or --series "2020:4,2021:9,…"; ≥ 3 points, warns below 5). describe: n, span, CAGR, average change, doubling time, YoY table, curvature read, spike check (Iglewicz & Hoaglin modified z-score). fit: OLS linear, exponential (log-linear) and logistic S-curve (Fisher–Pry linearisation over a K grid), best shape by R² with the Meade & Islam lower-half caveat. project --to 2030 [--model naive|linear|exponential|logistic|auto] [--ceiling K] [--event "label:P:impact"]: per-model projections, ± 2·RMSE band (not a prediction interval), horizon multiple, TIA arithmetic. --json everywhere; --demo reproduces the worked example; --selftest self-checks.

$ python3 scripts/trend.py project --demo --to 2030 --model naive --event "subsidy rollback in a major market:0.5:-5" --event "charging bottlenecks:0.4:-4" --event "battery cost breakthrough:0.3:+3"
TIA adjustment (Gordon): baseline 46.50
     -2.50   subsidy rollback in a major market   (P 0.50 × -5.00)
     -1.60   charging bottlenecks                 (P 0.40 × -4.00)
     +0.90   battery cost breakthrough            (P 0.30 × +3.00)
         = 43.30   adjusted projection (net adjustment -3.20)

The skill is fully usable without the tool; it only removes arithmetic effort.

Pair with adjacent skills

• assess-research-momentum — owns the bibliometric case: publication and citation curves.
• horizon-scanning — upstream: emerging issues become candidates once they have a series.
• scenario-planning — downstream: validated trends enter as predetermined elements.
• triangulate-sources — checks the evidence table when numbers came from web search.
• Methodology counterpart: methodologies/foresight/trend-analysis.md.

Anti-patterns

• Do not project an unvalidated trend; an UNCERTAIN or FAD verdict is itself the deliverable.
• Do not fit a line to three points; below five observations the honest output is a data request.
• Do not extrapolate reflexive systems such as markets or elections.
• Do not pick the curve that gives the wanted answer; report an alternative shape.
• Do not ship a projection without its assumption sheet.
• Do not skip the counter-trend, where discontinuity usually enters.
• Do not treat the card as permanent; a stale trend keeps steering strategy.

Reference

• T. J. Gordon, "Trend Impact Analysis," ch. 8 in J. C. Glenn and T. J. Gordon (eds.), Futures Research Methodology — Version 3.0. Washington, DC: The Millennium Project, 2009. ISBN 978-0-9818941-1-9 — the surprise-free baseline plus probability-weighted events of step 4.
• J. C. Fisher and R. H. Pry, "A simple substitution model of technological change," Technological Forecasting and Social Change, vol. 3, pp. 75–88, 1971. https://doi.org/10.1016/S0040-1625(71)80005-7
• E. M. Rogers, Diffusion of Innovations, 5th ed. New York: Free Press, 2003. ISBN 978-0-7432-5823-4 — adopter categories, which partition cumulative adopters; step 3 uses the 16 % boundary only on a cumulative series.
• N. Meade and T. Islam, "Modelling and forecasting the diffusion of innovation — a 25-year review," International Journal of Forecasting, vol. 22, no. 3, pp. 519–545, 2006. https://doi.org/10.1016/j.ijforecast.2006.01.005