You Can't Backtest Your Way to a Reason.

The mechanics produce the data, not the other way around.

Every price print, every volume spike, every pattern a model trains on sits downstream of something structural. Order book mechanics. Matching rules. Participant incentives. Information asymmetry. The data is just the footprint those forces leave behind.

So why is quantitative finance built almost entirely on the footprints?

Data-first reasoning won out, and it's not hard to see why. Compute got cheap, datasets got huge, and once you start looking, patterns are everywhere. Then something broke. When the datasets got genuinely massive and ML let people search high-dimensional spaces in a way they never could before, the methodology didn't adjust to its own success. It doubled down. Models that used to do simple pattern-matching turned into correlation engines running across hundreds of variables with no prior theory about why any of them ought to matter. The patterns multiplied. The failure modes multiplied right alongside them.

The problem wasn’t that the tools were too powerful. It was that people used them as the discovery mechanism instead of as instruments for testing a hypothesis that came from the layer producing the data. ML is very good at measuring complex structural effects, pulling parameters out of noisy data, picking out conditions a human analyst would take weeks to find. What it can't do on its own is tell you whether the correlation you found reflects a real market friction or just noise that happened to look like one. That distinction is what the structural framework gives you.

Philosophy of science settled the broader version of this a long time ago. The idea that observation alone gets you to truth, with no prior framework, mostly fell apart. Every observation runs through some set of assumptions. The person who insists they're "just letting the data speak" is still working from a framework, they just never bothered to write it down.

Structure-first starts further down than most people bother to look. Not market narratives, not broad participant behavior, but the actual mechanics: why a tick happens, how price moves through liquidity, what happens when resting orders get eaten at a level. The matching rules don't change when the VIX spikes. What changes is behavior. Liquidity providers widen their quotes, pull resting orders, adverse selection costs climb. A structural model expects this, because the inventory and adverse-selection pieces of the spread are exactly what the mechanics say should respond to that kind of regime shift. VPIN rises. Market makers respond rationally. Structure-first keys on conditions that hold regardless of how the market got there.

From there you build a theory about what those mechanics should produce under particular conditions. If some alignment of order flow, liquidity, and positioning reliably creates directional pressure, that's your hypothesis, and it has to be stated before you open the dataset. That requirement is the whole thing.

If you build a model, watch it work, then reverse-engineer a structural story to explain why, you haven't done structure-first analysis. You've done data-first with a narrative bolted on after. The story feels like an understanding. It isn't. That's where data-first almost always lands: a post-hoc explanation fitted to the result instead of a prediction tested against it. Structure-first wants the commitment up front, to a claim that can fail. That's the only thing separating real inference from a story that happens to fit.

The model comes after. Its job isn't to find meaning in noise. It's to measure whether the friction the structure implies actually shows up as directional pressure in the data. The model ends up being an expression of an understanding that already existed before you ran it.

I'm not claiming you can deduce a market's full behavior from matching rules the way a physicist works from axioms. Markets are complex adaptive systems; you can't. The claim is narrower, and more useful: understanding the mechanics constrains the hypothesis space. It draws the physical boundaries of what the market can actually do, so that when you finally look at the data, you're testing something instead of fishing for anything.

Most models get discovered inside datasets. Structure-first models get derived before the dataset is ever opened.