Market Physics: Why We Model Behaviour Before We Model Price

There are two ways to approach market data.

The first is statistical: find features that correlate with future price, build a model on those correlations, deploy it. The correlation is the evidence, and correlation is all you need. You do not need to understand why the feature works. You need it to work in the next period, and the model takes care of the rest.

The second is physical: understand the mechanical behaviour of the participants generating the data, reconstruct that behaviour from the observed signals, and trade based on understanding what is happening rather than predicting what will happen. The explanation is the evidence.

The market microstructure research programme chose the second approach. Not because the statistical approach is wrong, but because it has a specific failure mode: a correlation that the analyst cannot explain is also a correlation the analyst cannot defend when market conditions shift. Physics-first research produces findings that are either true or demonstrably false. Statistics-first research produces findings that are either statistically significant or statistically insignificant — categories that are useful but that do not tell you whether the underlying mechanism is real.

Institutional Execution Programmes

Large market participants do not execute their position changes through a single transaction. They cannot. The impact on price would be prohibitive, and counterparties would immediately identify and trade against them.

Instead, they execute in programmes: repeated, systematic purchase or sale of small quantities across an extended window, targeting a volume-weighted average price rather than a point price. The programme may run for thirty seconds or three minutes. It involves alternating aggressive takes — market orders hitting the opposing book — with passive resting — limit orders placed at or near the current price to absorb selling pressure. The programme builds, pushes price, and ends. The price partially reverts as the programme’s influence unwinds.

This is not speculative. It is the documented execution behaviour of institutions that have published their own VWAP methodology. The question was whether the programme’s presence in the market could be detected in real-time from publicly available trade tape and order book data.

Reconstructing the Programme

The reconstruction approach begins by clustering raw trade data: individual transactions that occur within a narrow temporal window are grouped into bursts. Within each burst, the directionality, size, and repetition of aggressive trades are measured. A programme requires a minimum number of repeated same-direction pushes within a short interval, minimum position size, and minimum capital deployed to count as a detected episode.

The detected episode is not a scalar or a binary flag. It is a structured record of approximately thirty-five attributes: the build phase — capital spent establishing resting orders before the push begins — the push phase — the aggressive take sequence and its associated cost — the ratio of capital spent absorbing opposing pressure to capital deployed pushing price, whether order book liquidity refills after large order cancellations or remains pulled, and the measured price movement at sixty seconds, five minutes, and ten minutes after the episode completes.

This is a physics measurement, not a prediction. The episode is described completely before any directional claim is made. The question of what price does after a given episode type is a downstream question, answered by aggregating the measured outcomes across hundreds of episodes.

What the Measurement Found

Approximately one hundred to one hundred and thirty-five detectable institutional execution episodes occur per day in the instrument studied. This finding held across multiple data eras and was replicated in two independent implementations built without cross-referencing each other. The structural pattern was consistent whether the tool was built from one direction or the other.

The outcome measurements — post-episode price movement at multiple horizons — revealed regularities. The programmes that ended with visible exhaustion signatures showed one pattern; the programmes that ended cleanly showed another. The build-phase characteristics predicted the push-phase outcome better than the push-phase signals themselves. The fade ratio — how much capital was spent absorbing selling rather than buying into free space — predicted reversion.

These were not hypotheses tested against data. They emerged from the data by measuring the phenomenon completely.

The Taxonomy of Named States

The accumulated record of detected episodes, categorised by their measured characteristics, produced a taxonomy. Named states emerged from the clustering of episodes with similar structural profiles: specific patterns of wall behaviour, specific build-phase and push-phase ratios, specific market context conditions.

The named states are not invented categories. They are observed recurring configurations in a physical process. In the same way that weather patterns are named — a cold front, a stationary system, a ridge of high pressure — because they recur with sufficient consistency and predictability to deserve names, the execution pattern states were named because they recurred across data eras with consistent structural signatures.

A routing system built on this taxonomy can respond to a detected episode not by running a prediction model but by recognising the episode type and referencing the historical outcome distribution for that type. The prediction is implicit in the classification.

The Laws That Emerged

Three regularities survived across datasets, implementations, and market periods.

The exhaustion law: visible institutional execution programmes, when large enough to detect clearly, mark the end of directional moves rather than their continuation. The programme has absorbed the available supply or demand. Price has been pushed. Continuation requires a new programme, and new programmes require a new accumulation phase.

Leg continuation: the internal structure of a directional move — specifically the ratio of push phases to build phases and the profile of fade within each push — predicts the probability that the current leg extends before a directional reversal occurs.

Sensors have habitats: certain signals are only valid within specific market conditions. A flow signal that is robustly predictive during a detected build phase is not the same signal during an ambiguous period. The context is not a modifier on the signal; it is a prerequisite for the signal existing at all.

These are not statistical regularities. They are mechanical consequences of how institutional participants must operate when they trade large size in continuous order-driven markets.

The Connection to TVS

The physics-first approach — measure the mechanism completely before asking about outcomes — is identical to the approach TVS applies to search ranking intelligence.

Ranking research that begins by asking which features correlate with position one misses a prior question: what is Google mechanically trying to achieve when it orders results? The mechanism — relevance, authority, user experience signals, geographic proximity in local search — is not hidden. It is documented. The correlation between a feature and ranking position is meaningful only when you understand the mechanism that produces the correlation.

Reconstructing the institutional programme before modelling price is the same discipline as understanding the ranking mechanism before measuring ranking signals. The search changes when you know what you are looking for and why.

---

The Signal 1 retraction that preceded this structural work is at The Most Valuable Thing We Built Wasn’t a Signal. The specialist committee that processes signals from this reconstruction pipeline is at The Specialist Committee. The TVS market intelligence dataset that applies this methodology to search ranking is at What 249 KZN Businesses Reveal.