Generative Market Hypothesis: Price as a Residue
Efficient-market theory treats price as a summary of information. The generative view treats price as the residue of interaction — a difference that decides what to study, what a tool should look like, and when to admit the old model has expired.
There is a scene that repeats across three very different episodes — the summer of 2015, the spring of 2020, and the technology re-rating of 2023. The model had produced a reasonable range beforehand. After the event, it could not explain why. No parameter was wrong, no out-of-sample test had failed. The problem was that the thing being tested was no longer the same thing.
I call it "the model was right, the world had changed". Those six words are where my research begins, and why I no longer treat forecasting as the core deliverable.
1. Three hidden assumptions in forecasting
Every forecasting model quietly accepts three premises. They are rarely written down, and almost never asked about in a pitch.
First, that structure is stable. The model assumes the relationship between variables holds through the forecast window, or drifts slowly enough for a rolling window to absorb it. When structure switches tracks — a regulator rewrites the rules, a technology converges, the composition of participants changes — the window is filled with samples from the previous world.
Second, that information is exogenous. In efficient-market theory information falls on the market like rain, and price aggregates it. In a real market, participants infer information from price, act on that inference, and the action moves the price. Information is not rain falling on the system from outside. It is rain the system makes.
Third, that error is additive. The model treats deviation as noise and assumes forecast plus error equals reality. Once participants start learning the model itself — the central bank reading market expectations, the market reading the reaction function — error becomes endogenous. It gets arbitraged, and it gets amplified.
These three assumptions are almost harmless in calm periods. Their problem is not that they are wrong, but that their domain of validity gets quietly extended to the turning point.
2. Price is the residue of interaction
The generative hypothesis changes the ontology: a market is not a structure made of pre-existing entities, but a process that keeps taking temporary shape through interaction. Price is the residue that process leaves behind, and at the same time a prompt for the next round.
This is easily misread as "markets are inefficient". It does not claim that. It claims that efficiency is itself generated, and therefore has a history of being generated and of failing. The pricing efficiency of a market depends on who is participating, which information channels exist, and what constraints bind at that moment. Change the structure and the distribution of efficiency changes with it.
Forecasting asks what will happen next. Generation asks what is being produced, and what conditions it needs in order to continue.
3. The research task changes
If price is residue, the center of research shifts from estimating true value to identifying the generative mechanism. In practice I split it into three workable questions.
Who generates? Not "the market" in the abstract, but a set of agents with different objectives, information sets and constraints: market makers, passive funds, industrial capital, regulators, algorithms. Their interaction rules decide what shape the structure takes.
Through what? Liquidity, attention, standards, or regulatory permission? Different markets generate through completely different channels. Pick the wrong mechanism and even a precise model is only fitting noise.
When does it switch? Structural change is rarely gradual; it completes abruptly at certain thresholds. Before the threshold, the old model performs best; after it, the old model is the most expensive thing in the room. What is worth investing in is the early signal near the threshold.
4. What can be tested
A hypothesis that explains everything says nothing. The generative view at least offers three implications that data can reject.
- Price series should be non-stationary and path-dependent, rather than stationary fluctuation around an equilibrium.
- Rising dispersion of participant behavior should lead structural repricing, not move with it or lag it.
- Changes in the connectivity of the market network should predict non-linear contractions in liquidity.
All three can be given a first test with public data, and all three can be falsified. I would rather someone ran them strictly and reached a negative conclusion — that is worth more than another essay agreeing with me.
5. What is unresolved
First, I still have no good way to compress "generativity" into one stable statistic. It may only be identifiable through a combination of indicators, which means it is hard to communicate quickly in practice.
Second, if structure keeps changing, where is the validity boundary of backtesting? For now I approximate with rolling windows and regime segmentation. That is an engineering compromise, not a theoretical answer.
Third, in a market like China's, where policy intervenes heavily, how does generation couple with institutional constraint? Policy is both an exogenous shock and an endogenous participant. That double role makes model design genuinely awkward.
The point of this essay is not to announce a new paradigm. Paradigms are not announced; they are what remains after being used, questioned and revised. If your work contains a "the model was right, the world had changed" moment, write and tell me the detail. That is the material I need.