Skip to content
Detrend · time-series detrending and trend inference

A detrending tool that refuses to guess.

Detrend is a browser-based detrending tool that turns a raw Excel column into a defensible trend statement. It runs a fixed seven-step protocol inside your browser: unit-root diagnostics decide which operations are legal, the trend test is chosen from those diagnostics, and every number is published with the independent oracle that validated it. Nothing is uploaded.

Computation
Parsing, diagnostics, decomposition and tests run in Web Workers on this device. No upload endpoint exists.
Verification
Every user-visible statistic is compared against an independent implementation with a published tolerance.
Protocol
Eight prescribed steps. No step is skipped silently, and a blocked step states its reason and its alternative.
Demonstration: an Excel time series is imported, sentinel and stuck readings are flagged, a change point is detected, a trend with its 95% band is fitted, and the finding is stated in the panel on the right.
One complete run: the data arrives, the record is screened, the structure is found, a model is fitted and the finding is stated.960 × 560 · 3.6 s · 85 kB
01 · Protocol

The protocol

The analysis is a sequence, not a menu. Each stage has an action, a decision rule that can block the next stage, and a defined output. The rules below are implemented literally in the workspace.

  1. 01Ingest

    The workbook is parsed in a dedicated Web Worker. Sheet names are read first; only the selected sheet is materialised.

    Rule A sheet without at least five numeric values is rejected before any statistic runs.

    Output Sheet inventory, column list, usable sample size.

  2. 02Screen

    The parsed column is audited before any statistic runs: the time axis, the sentinel codes, the completeness and the instrument resolution.

    Rule A non-monotonic time axis blocks the analysis. Sentinel codes are excluded and counted, never treated as measurements.

    Output Integrity findings, usable n, missing ratio, sentinel count and the provenance record.

  3. 03Diagnose

    ADF and KPSS are run against each other as complementary unit-root tests.

    Rule Rejecting ADF while keeping KPSS admits a stable trend baseline. The reverse admits differencing.

    Output Stationarity verdict and per-method availability.

  4. 04Clean

    Robust outlier rules remove only points that fail their own stated threshold.

    Rule IQR, modified Z, MAD/Qn and STL-residual rules each publish their efficiency and breakdown point.

    Output The removed set, with the rule and row index for every point.

  5. 05Detrend

    The detrending transform is selected by the verdict rather than by a default, and an explicit trend-plus-seasonality model is fitted alongside it.

    Rule A blocked operation is never executed silently. Its reason and a legal alternative are shown instead.

    Output Trend, seasonal and residual components, or the differenced series; the model slope with its interval and its AICc.

  6. 06Regimes

    PELT with the modified BIC penalty searches for change points, and an interrupted time series is fitted at the strongest boundary.

    Rule The penalty is calibrated so that noise is not cut into pieces, and the candidate models are ranked by corrected AIC.

    Output Change points, the pre- and post-intervention rates with Newey–West standard errors, and the model comparison.

  7. 07Test

    The Mann-Kendall family is applied with the correction the diagnostics support.

    Rule Significant autocorrelation forces the corrected result to be shown beside the uncorrected one.

    Output Tau, p-value, effect size and the Theil-Sen interval.

  8. 08Report

    The finding is stated with its interval, its caveats and the decision trace behind it, and the whole run is exported as a reproducible recipe.

    Rule A statistic without an independent oracle keeps a permanent UNVALIDATED label in the interface.

    Output Finding, interval, caveats, the trace and a downloadable recipe with every parameter in force.

02 · Decision rules

How the detrending tool chooses the method

This is the decision table encoded in the workspace. It is published so that the choice can be argued with, not merely accepted.

How the detrending tool chooses the method
Diagnostic conditionDecisionReason
ADF rejects the unit root, KPSS does not reject stationarityDeterministic detrending is availableA stable trend baseline is plausible, so flattening the series is defensible.
ADF cannot reject the unit root, KPSS rejects stationarityDifferencing is used; a straight line is blockedThe series wanders like a random walk, and a fitted line would manufacture structure that is not there.
The two unit-root tests disagreeDeterministic detrending carries a warningAmbiguity is propagated into every downstream result instead of being resolved by preference.
Lag-1 autocorrelation exceeds the 95% bandThe autocorrelation-corrected test becomes the headlineVariance under independence is overstated when observations persist, so the corrected p-value is the honest one.
A cycle length is detectedSTL replaces the flat trendA single slope cannot represent a seasonal component, and the seasonal test depends on it.
The usable sample is too small or too incompleteTrend inference is blockedDescriptive statistics remain available, but no headline is produced from an unstable sample.
03 · Output

What a result looks like

Every finding is reported in five layers, in a fixed order. The statement is the same whether it is read by a statistician or a non-specialist, because the plain-language layer is a restatement of the same numbers rather than a separate claim.

Output specimenillustrative values
↘Decreasing
Validated
Effect estimate
Rate of change
−3.214 units/step
95% interval
−4.502 … −2.106
Usable sample
128 / 131
Inference
Test
Mann-Kendall (τ)
τ
−0.412
p
0.003
α
0.05
Decision
Reject H₀ — trend detected
Provenance
Stationarity verdict
trend stationary
Detrending
STL
Cycle length
12
Outliers removed
3 (2.3%)
In plain language

This series is decreasing overall, by about 3.21 units per step on average (95% likely between 2.11 and 4.50 units per step). The trend is very likely real, not random noise.

  1. 01

    Decision

    The direction of the series and whether it is distinguishable from chance at the stated alpha.

  2. 02

    Effect estimate

    The rate of change, its 95% interval and the unit it is measured in. An interval, not only a point estimate.

  3. 03

    Inference

    The test that was used, the rank correlation, the p-value, alpha and the resulting verdict, each number carrying its verification status.

  4. 04

    Plain language

    One sentence restating the same finding for a reader who does not read test statistics, with no new claim added.

  5. 05

    Trace

    Every automatic choice in order - diagnostic, rule, threshold, method - with the verification row behind it.

04 · Evidence

Four independent lines of evidence

Agreement with another implementation only proves that two programs compute the same thing. It does not prove that the thing is right. The workspace therefore carries four separate kinds of evidence, and each of them can fail on its own.

  1. 01

    Agreement with an independent implementation

    Every user-visible statistic is compared against statsmodels, SciPy, pymannkendall or ruptures, with the tolerance fixed before the comparison runs. A failed row blocks the build rather than producing a warning.

  2. 02

    Recovery of a known truth

    Synthetic series are generated with a planted slope, a planted seasonal amplitude, a planted change point and a planted level shift. The estimators must recover the parameters that were put in, within a tolerance that is stated in advance.

  3. 03

    Metamorphic invariance

    Negating, scaling, shifting or monotonically transforming the series must transform the output in exactly the predicted way, and reversing time must mirror the slope and the change points. These relations hold for every correct implementation.

  4. 04

    Calibration under the null

    The realised type-I error of the trend test is measured on data with no trend and must match the nominal level. The change-point penalty is calibrated so that noise produces no boundary while a one-standard-deviation shift is still found.

05 · Verification

Correctness is a reproducible comparison, not a promise

The public matrix lists every statistic with its oracle, its tolerance, the rationale for that tolerance and the fixture it was checked against. A failed row blocks the build. A statistic that appears in the interface without a verification row is treated as a gate failure, and a method with no available oracle is permanently badged as unvalidated rather than quietly trusted.

Inspect the verification matrix92 validated · 6 unvalidated · 98 rows
06 · Privacy

Your file never crosses the network

There is no upload endpoint, no server action and no API route that accepts file bytes. Parsing, diagnostics, STL, trend tests and charts run on this device in Web Workers. Closing the tab discards the data. The only stored preferences are your display depth and whether you have seen a chart introduction.

Read the privacy statementzero upload endpoints

Thirty seconds from workbook to a defensible answer

Drop a file and the workspace selects the sheet with the most numeric columns, picks the first numeric series, runs the stationarity diagnostics and states a finding with its interval. Nothing has to be configured to get a correct first answer, and nothing has to be taken on trust to trust it.