A market is not a physics experiment
The data a price-prediction model trains on was generated by people reacting to other people, who were in turn reacting to rules, news and each other. The generating process is not stationary. It changes when the rules change, when the participants change, and when the instruments change. A model that treats every observation as if it came from the same distribution is making an assumption the data itself contradicts.
The break is the point
A structural break — a shift in monetary regime, a listing reform, a clearing-house rule, a sudden change in liquidity providers — is, by construction, something the pre-break data does not contain. No amount of additional historical data will surface it, because it is precisely the moment at which the historical pattern stops applying. Models that announce they have “learned the market” are usually announcing the moment before they meet one.
How we read claims about regime change
We ask whether the people making the claim have named the breaks their model lived through, and whether they tested the model on the far side of each break. A claim that the model “handled” a break is much stronger if the model was evaluated only on data after the break, not on data that included it.
The limit, plainly
Prediction across a regime change is not a harder version of prediction within a regime; it is a different problem. Within a regime, a model can be a useful summary of recent structure. Across a regime change, the model is being asked to extrapolate into a state it has never observed. The honest framing is that the model is reliable up to the next break, and silent about when that break arrives.