Classifying Trend and Mean-Reversion Regimes with Hidden Markov Models
Summary
The note discusses how to identify market periods that favor trend-following or mean-reversion strategies. One suggested approach is regime modeling with Hidden Markov Models, with two research papers offered as starting points. The answer also cautions that useful trading rules are often proprietary, so published methods may not reveal practitioners’ actual criteria.
A second response argues that favorable periods are generally recognized after the fact rather than reliably forecast in advance. It cites trend-following losses during unsettled Treasury yields and recalls a practitioner’s account of cutting trend exposure when volatility was unusually high. That example offers a possible risk filter, but no precise rule or evaluation. The discussion provides ideas for research, not evidence that any classifier can predict regimes or improve live performance.
Key ideas
- Hidden Markov Models are one proposed way to model market regimes.
- The document describes favorable periods for trend and mean-reversion as difficult to predict in advance.
- High volatility is cited as a possible reason to reduce trend-following exposure.
- The suggested volatility filter is anecdotal and cannot be assessed without a specific rule and testing.
Tags
Full text
# What is the common accepted/ best performed method to classify trends and mean-reversion for fixed peroid? # What is the common accepted/ best performed method to classify trends and mean-reversion for fixed peroid? I have knew some strategies only work on trends peroid, and other only works on mean-reversion peroid. But I didn't find how to classify trends and mean-reversion. I wonder the best performed/verified research/paper on this topic, but searching google get a lot of stuffs, hard to verify which is good in real. So ask for advice from someone with experience. ## Answer by Valter (score 1) https://quant.stackexchange.com/a/77733 Up to interpretation, and specifics about this is generally not disclosed since it provides a real edge in trading. One common approach that use is Regime modelling using Hidden State Markov Models. There are several papers available on the topic, below are some suggestions: - https://arxiv.org/pdf/2307.00459.pdf (good starter) - https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4556048 (a bit advanced) ## Answer by nbbo2 (score 0) https://quant.stackexchange.com/a/77730 It is true that there have been historical periods when trend following (tf) methods worked particularly well and others where mean-reversion (mr) worked well. But to my knowledge there is no way to predict this ahead of time, such periods are only identified in retrospect. For example in RiskNet, Aug 2023 we read: "After a stellar 2022, seesawing US Treasury yields in March [2023] led to sharp losses for many trend-following firms." You would think if the firms had a way to predict if the upcoming period was favorable for trend-following or not they would have avoided these losses. Most likely they did not know. OTOH I remember tf pioneer Larry Hite saying in a lecture that he would reduce or even suspend his trend following trades if volatility was very high, because such periods were unfavorable for tf. But he didn't reveal the exact rule he followed, such trading rules are unlikely to be made public and cannot be evaluated. At best they give a clue to what criteria are used.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.