Assessing Hidden Markov Models for High-Frequency Market Data
Summary
The document explores whether Hidden Markov Models might identify useful market regimes more effectively in relatively high-frequency data than in daily or weekly returns. The author reasons that denser observations could make it easier to distinguish groups of returns with different probability distributions, potentially revealing patterns beyond simple return correlation. Possible examples include conditional relationships in which one return is followed by another with a stronger association.
These are hypotheses posed for discussion, not demonstrated findings. The document supplies no data, fitted model, performance comparison, or out-of-sample results. Higher sampling frequency does not by itself guarantee less noise or more exploitable structure, and any apparent states or relationships would require careful validation against overfitting, changing market conditions, and trading costs. The source raises a modeling question but does not establish that Hidden Markov Models outperform volatility models or longer-horizon approaches.
Key ideas
- Hidden Markov Models represent observations through latent states with distinct distributions.
- The author hypothesizes that dense market data could help separate return regimes.
- State separation might expose conditional return relationships beyond simple correlation.
- The document presents conjectures without empirical validation or model comparisons.
- Higher frequency alone does not establish a reliable trading signal.
Tags
Full text
# Hidden Markov Models for Higher frequency trading # Hidden Markov Models for Higher frequency trading I'm curious if anyone can validate my train of thought here with the utility of Hidden Markov models for modeling things happening on higher frequency trading activity versus lower frequency, and in the case of the former I do not necessarily mean HFT—but relatively high frequency. Hidden Markov models have been used all over quant finance for various things, as an example this paper goes into the use of Hidden Markov models over GARCH (1,1) models for predicting volatility. My intuition however tells me that trying to train Hidden Markov models on raw financial data over larger periods of time is not always going to be the best idea given you're training it on something heavily random, i.e., forcing it to train on Brownian motion. On the other hand, with high(er) frequency trading data does the story change a bit? The data is more "dense" and eliminating the noise factor is easier to do with data on higher frequency time scales than on days / weeks thus making it more exploitable by a Hidden Markov model than on longer time scales. With high frequency trading specifically, I would think the applications of Hidden Markov models could be more beneficial for a couple reasons: - It would be easier for the Hidden Markov model to do its job of sorting returns into groups where all of these respective groups have some corresponding probability distribution, and these would be the states for the Hidden Markov model. - Given these states fit together in a nicer way, wouldn't that mean it would be easier for the Hidden Markov model to find certain things of interest? They are able to statistically separate out groups with different price movements—perhaps find things “deeper” (e.g., say it finds that over a period of time the returns are not correlated, it may be able to find another relation that is of more subtle nature, a certain return on $X$ is followed by $Y$ with x% stronger correlation, etc…) Is there any truth to these thoughts?
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.