Different worlds.
Similar shapes.
Discover how two measures move together—even across countries, subjects and units.
To explore weekly inventories or compare daily ports within their own collection, choose “All compatible series” under Discovery scope.
Loading the shape study…
Unexpected pairs to investigate
Selected for varied subjects and useful overlap. Compare the shape and movement scores before forming a story. These are examples, not causal findings or statistically validated connections.
Strongest shape matches
Ranked by absolute Pearson correlation. Negative scores identify opposite movements.
Does the connection hold up?
These sensitivity checks describe the selected window. They are exploratory diagnostics, not significance tests, proof of causation or a prediction.
Prepare a story card
Cards include both scores, overlap dates, source credits and the exploratory caveat. The reporting notes include full attribution and licences.
Sources and reporting context
Aligned observations used in this score
| A period | B period | A input | B input |
|---|
How to read this study
A matching shape is a reporting lead, not evidence of causation. With billions of comparisons, impressive coincidences are expected. These scores are exploratory; they are not significance tests or predictions.
Shape, not size
We calculate Pearson’s r on paired observations. The display converts each input to deviations from its own mean, measured in standard deviations. Units and absolute scale do not influence the score. A positive score near 1 means matching movements; a negative score near −1 means opposite movements.
Two distinct questions
Graph shape uses the observed levels. Movement mode uses first differences between consecutive source periods, then compares their shapes. Differences in percentage rates are percentage points. No percentages of growth are calculated. Shared long-term trends can create a strong level correlation even when the movement score is weak.
Exact overlap and time lags
Each pair needs at least 10 paired observations. We preserve missing values and do not interpolate. Differences never bridge a missing period. Positive lags mean B follows A. Annual, monthly, weekly, daily and marketing-year series are separate. A lag means weeks for weekly histories, days for daily histories, months for monthly histories and years for annual or marketing histories. Daily freight stays at its original daily grain; no averaging is performed.
Discovery screening
The default discovery ranking requires at least 25 paired observations. For each selected method, each complete input history and its actual paired overlap must contain at least five distinct values, fewer than 80% zeros and more than three value transitions. Pairs with |r| ≥ 0.999999 are screened as near-identical standardized shapes, which may reflect shared data or a rescaled copy. These are transparent reporting heuristics, not evidence that the remaining pairs are real connections. Choose the original ranking or a specific pair to inspect screened results; direct inspection retains the ten-observation minimum.
An exploratory search
The full scan retains separate original and discovery-screened top-300 rankings for each method, lag and scope. Calendar, direction and overlap filters narrow the retained overall rankings; they do not rerun the complete batch search. Selecting a series searches its full compatible collection with these filters. Choose a specific second series to inspect any eligible pair directly. Date-window controls change the inspected pair and its exports; discovery rankings use the full overlap. Differences are calculated before filtering the date window, so the first difference may use the preceding observed source period. The half-period checks split the aligned observations chronologically and require 10 observations per half; they describe sensitivity to the chosen window, not out-of-sample prediction. Ordinary p-values would be misleading here because time trends, autocorrelation and searching many pairs affect chance matches. Check source methods, revisions and alternative explanations before reporting a connection.
Study coverage and exclusions
Study 6d196f5ddf73f24ae9e3 · catalogue release wdi-2026-10-08-0a6eb6ab25b7+release-aeeb6ddf-6cce-4970-a73b-0d9cd459923d · prepared 2026-10-15.
84,553 annual series · 4,808 monthly series · 1,092 marketing series · 2,090 daily series · 4 weekly series
- 693,055 records/series: Fewer than 10 present observations
- 2,033 records/series: Constant series (correlation undefined)
- 3 records/series: Source period or methodology is unsuitable for direct correlation comparisons
- 903 records/series: Duplicate source series represented in another collection
- 117 records/series: Latest-only trade, reserves and food balances; ownership records have no comparable numeric time history
All public dataset collections are inspected. Single-year snapshots, ownership relationships and series with too few present observations cannot support a shape study. The minimum excludes the short bilateral-trade histories and sparse survey waves. Source duplicates with the same recognized measure and geography are counted once where identity is available; other collections may contain related or duplicated measures. Close matches can therefore reflect shared underlying data.