Detecting Market Regimes in Bitcoin

Abstract

Detecting a regime and trading a regime are different tasks. A model can classify historical observations convincingly without making useful forward decisions. This note focuses on the distinction between explaining the full sample and producing an estimate at the actual decision time.

1. A state estimate, not a market fact

Regime-switching models describe a process through a set of possible states and transitions. Hamilton’s 1989 work is a foundational reference for this approach. It supplies a modeling framework—not a universal number of states or a guaranteed forecasting advantage. Hamilton, 1989 · publication record ↗

A label such as “low risk” should therefore be interpreted as an output of a particular model, conditional on its inputs and assumptions. A more careful label would identify the quantity being estimated, such as volatility, instead of implying safety in every dimension.

2. Volatility is one possible modeling target

The authors’ MSGARCH documentation describes tools for fitting, filtering, forecasting and simulating Markov-switching GARCH models, including risk measures. It also lists the Bitcoin application by Ardia, Bluteau and Rüede. These references establish the method’s scope and the existence of that application; they are not a replication of its empirical findings here. MSGARCH documentation ↗

3. The information boundary matters

Consider a hypothetical decision made on Monday. An estimate using observations through Monday is available for a later execution time. An estimate revised using Tuesday or Wednesday’s data is not the same input. Using the revision to score Monday’s trading decision would give the backtest information that the trader did not have.

This is why the data boundary, signal timestamp and execution timestamp belong in the research record. The issue is not whether a model is complex or simple; both can be evaluated with information that arrives too late.

4. Estimation is not allocation

Even an accurate risk estimate leaves an allocation problem. The process must decide how much exposure to take, how rapidly to change it and what to do when an order cannot be implemented as intended. Small differences in these rules may change turnover and the distribution of returns.

A conceptual flow diagram can explain those steps. It cannot substitute for the executable implementation or prove that the steps were applied consistently throughout a simulation.

5. What to ask of a validation study

A defensible evaluation should explain how model choices were selected, which observations were withheld, whether the final settings were frozen before evaluation, and how unsuccessful candidates were treated. It should preserve enough information to distinguish a genuinely untouched test from a favorable retrospective window.

Reporting a positive “out-of-sample” period is not sufficient on its own. The chronology of selection and freezing must support that label. Where that evidence is not public, a reader should treat the report as a source claim rather than independent proof.

Conclusion

A regime estimate is one input to a decision process. Its usefulness depends on timing, implementation and evidence beyond an attractive historical explanation. Read the model claims together with the scope of the supplied evidence, not as a guarantee that the next market transition will be identified correctly.