# Curve-Fitting and Cherry-Picked Windows Given any noisy time series, you can produce a trend of nearly any sign or magnitude by choosing where to start and stop. Nothing is fabricated. The data is real, the arithmetic is right, and the conclusion is manufactured entirely by selection. ## The endpoint move The classic form: **start at an unusual high, end at an unusual low, and report the slope.** The durable example is the "global warming pause." Starting a temperature trend at 1998 — an exceptionally strong El Niño year and a large positive anomaly — and ending in the early 2010s produced a flat or slightly negative slope, presented for roughly a decade as evidence that warming had stopped. Shift the start year by two in either direction and the trend reappears. The claim depended entirely on a single anomalous starting point, and it dissolved as the record extended. See [[Manufacturing Doubt]] for how this fits the broader denial toolkit. The move recurs anywhere a series is noisy: crime rates (pick a peak year and report the decline, or a trough and report the rise), economic indicators (align the window to an administration boundary), and immigration and enforcement statistics. ## Why it works **Short windows are dominated by noise.** In a series with real variance, the trend estimated over a short interval is mostly sampling variation. The Signal vs Noise ratio is the whole issue: over-short windows report noise as if it were signal, and the shorter the window the more freedom the selector has. **Nothing is falsifiable.** Every number quoted is accurate. There is no fabrication to expose, which is what makes this different from ordinary misinformation and much harder to counter — you cannot fact-check a true statement. **The audience cannot see the unselected windows.** The counterfactual — what the trend looks like over other reasonable ranges — is not presented, and reconstructing it requires access to the underlying series. ## The model-selection variant The same freedom exists in choosing the functional form. Fit a high-degree polynomial to any short series and it will track the data closely and then diverge wildly outside the fitted range. Overfitting produces confident extrapolation from noise. The COVID period generated many instances: curves fitted to a few weeks of case data, extrapolated forward, and presented as forecasts. Cubic fits to early case counts predicted the epidemic ending on specific dates. **The fit was excellent and the forecast was meaningless, because a model with enough parameters fits any short series and the fit quality says nothing about predictive validity.** The technical term for the required check — holding out data the model has not seen — is exactly what a rhetorically-motivated fit omits. ## Legitimate versus illegitimate selection The distinction is real and worth stating, because "you cherry-picked" is itself sometimes an unfair charge. **Legitimate:** the window is chosen for a stated substantive reason (a policy took effect, a definition changed, a data series began), the reason is disclosed, and the result is reported alongside other reasonable windows. **Illegitimate:** the window is chosen because of the result it produces, the choice is undisclosed, and alternative windows are not shown. The operational tell is **whether the analyst discloses the sensitivity.** Someone confident in a finding shows that it survives reasonable variation in the window. Someone who has selected for a result will present exactly one range and no robustness check — which is the same discipline that separates honest from motivated use of [[Markov Chain Ensemble Methods]] and of [[Dynamic Scoring]] elasticities. ## The diagnostic Ask what the series looks like over a longer range. Ask why these endpoints. Ask what the trend is under three other plausible windows. Almost every cherry-picked claim fails the first question, and the reason it survives in public argument is that almost nobody asks it. ## Related [[Statistical Impossibility Arguments]] · [[Benford's Law and Election-Fraud Claims]] · [[Pseudo-Precision as Rhetoric]] · Signal vs Noise · Statistics · [[Manufacturing Doubt]] · [[Framing]] · [[Mathematics and Power Hub]]