Sparse sensing · Real financial records
A few filings.
A fuller picture.
Historical reconstruction · six quarters · actual reported values
Patterns learned across company histories give a few new observations a much larger context. Here, four selected financial readings estimate hundreds of other line items. Explore what the model has learned—and how those readings bring the wider picture into focus.
4 financial readings · 500+ line items inferred · 140+ companies · one quarter
Reads four financial values.
The six tiles of a block are revenue, capital spending, deferred revenue, inventory, remaining performance obligations and purchase commitments. The method selects four line-item readings that help distinguish the learned patterns.
Infers the rest.
Four readings determine how strongly the four learned patterns enter the estimate. Those patterns reconstruct the other values, above or below their historical means. Only four current-quarter values enter the reconstruction.
Checks against what was filed.
The historical reconstruction is scored tile by tile against reported values withheld from the inference.
Knowledge comes first.
Before reading the quarter, the model learns patterns across earlier company histories. Each pattern has a footprint—which financial measures rise and fall together—and a history of how strongly it appeared over time.
The model estimates a new quarter as a combination of these patterns. Four well-chosen readings determine four amounts: how strongly each pattern enters that estimate. The learning supplies the relationships; the new readings locate the quarter within them.
Learn patterns of shared financial variation from earlier-quarter histories, and measure how much of that training variation each carries.
Select readings that distinguish those patterns, so each observation contributes information to the reconstruction.
Use the four readings to solve for the pattern amounts. Reconstruct the wider picture and check it against the withheld reported values.
Mix the patterns yourself.
Each slider is the amount of one pattern, in units of that pattern's usual quarterly size: +1 is a typical quarter's worth, 0 is none. Drag them and watch the whole economy recolour coherently: sectors move together because the patterns say so. Press Kenshin's mix for the amounts inferred from four readings, or Fit from all values for the pattern combination estimated using the full quarter.
Your mixKenshin's mix
What was filed
Why the choice of four matters.
Same quarter, same budget of four readings: the method’s selected readings and a random selection. Tiles the inference got wrong by more than 1.5 units are lit magenta; the brighter, the further off. The random example is the draw nearest median error among 1,000 selections.
Kenshin's four tiles wrong
A random four tiles wrong
Against 1,000 random selections.
Each dot represents one percentage point of the 1,000 random selections. Lime shows the share beaten by the selected four.
How much closer.
Mean distance between inferred and filed tiles, in standardized units after transforming the financial values; shorter is closer.
Why the readings matter.
Where the method looks.
Across the six quarters shown, the method made 24 choices of a reading. These are the financial measures it selected and the line items it returned to more than once. They serve as the panel’s strain gauges: observations chosen to help distinguish the learned patterns.
Read through time.
A reported value jumps. How much reflects persistent activity, and how much a temporary effect? This illustration models the two components with different responses over time. A single reading permits many explanations; a second can distinguish them when the timing response is known.
Insect strain sensors also convey information through the timing of their signals. Move readings seen from one to two to watch the possible explanations narrow. Here the components are nonnegative and the readings are noise-free; the fading control shows how separation becomes fragile when their responses look alike.
The Kenshin connection
A signal has meaning
within a larger system.
An insect can respond to a gust without forecasting its arrival. For Kenshin, the related research question is which economic signals to observe, how to interpret their changes, and how that understanding can inform a response.
The reconstruction above illustrates part of that idea: learning relationships first, then using a few selected readings to estimate the wider picture.
You don’t need certainty about the future to make a better-informed decision today.
Retrospective research demonstration using SEC financial values, with 136 to 152 companies and 478 to 542 evaluated line items across the six quarters shown. The fixed display includes 155 company blocks, organized by sector; its layout is not a discovered economic topology. Values are transformed with asinh(USD/1 million), then standardized against each line item’s historical mean and variation. The four patterns use earlier-quarter histories, including interpolated missing entries. The calculation does not enforce filing-date availability and uses full-panel information for eligibility; it is not an ex-ante backtest. The random comparison uses 1,000 selections per quarter and includes the four observed values in its error score. A magenta tile marks an error above the stated 1.5-unit threshold; other tiles can still have error. Training variance explained is not out-of-sample accuracy. The spreading animation illustrates one reconstruction calculation. The two-component time example assumes its timing response is known. These examples do not establish investment performance or a portfolio-protection capability. Research inspiration: data-driven sparse sensor placement and wing sensing and neural encoding.