Markov Models and Forex | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

New historical Forex models 140 replies

Anyone used ARMA + GARCH models in forex trading? 14 replies

Ecn/stp/dma models and slippage 0 replies

Risk Models 0 replies

MT4 VWB Models 1-4 8 replies

Attachments: Markov Models and Forex
Exit Attachments

Markov Models and Forex

  • #1
  • First Post: Nov 14, 2007 5:38am Nov 14, 2007 5:38am
Some time ago I came across Markov Chains and Hidden Markov Models. It's quite an advanced mathematical model and it wouldn't be possible to describe it briefly here.
From the very start I thought about financial markets application. Now, when I'd like to start some thorough research, I'd like to know if anyone has ever tried or heard about such attempts and their results. I know that some big financial institutions tried to apply it but don't know the results.

Looking forward to seeing your opinions.

Pr1mer
  • #2
  • Sep 2, 2025 1:04am Sep 2, 2025 1:04am
Wow, this thread was started 18 years ago.

I haven't seen a simple, straightforward Markov model used in forex yet.
A high timeframe, 3 by 3 matrix for transition probability seems to work well for GBPUSD and EURUSD.
Pure price action being used as inputs for now, but seems promising. Interestingly enough, > 55% probability worked well in backtests.

Anyone looked into this or using this in practice?
Attached Image(s) (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 8 KB
Click to Enlarge Name: screenshot.png Size: 20 KB
Click to Enlarge Name: screenshot.png Size: 7 KB
Click to Enlarge Name: screenshot.png Size: 22 KB
  • #3
  • Sep 2, 2025 7:58am Sep 2, 2025 7:58am
Quoting carddard
Wow, this thread was started 18 years ago. I haven't seen a simple, straightforward Markov model used in forex yet. A high timeframe, 3 by 3 matrix for transition probability seems to work well for GBPUSD and EURUSD. Pure price action being used as inputs for now, but seems promising. Interestingly enough, > 55% probability worked well in backtests. Anyone looked into this or using this in practice? {image} {image} {image} {image}
Hey carddard,

thanks a lot for your reply to that old thread. It shows you’re genuinely interested in the topic itself and that you have the courtesy not to leave even older but interesting questions unanswered. Kudos!

What I'd still be curious about, though, is which market situations the panels are supposed to represent.

Yes, I've played around with this a few times in the past. Here's something I put together quite a few years ago. I didn't really get very far with it though - I've always seen it more as a descriptive tool than as a predictive model. But of course, I could be wrong.

Initially, I was interested in classifying market phases after CP's post:
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 35 KB


Sure, with MACDs, but also using MAs themselves or differences between reference points you can get to something similar. Pretty quickly, though, it becomes clear that there's only a certain order in which you can move from one phase to another. But if you think the logic through and actually screen the market with it, you'll quickly notice that the direction of the phase evolution can change.
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 16 KB


That's why I went ahead and visualized the Markov chain - and even there I noticed some interesting patterns once self-references were filtered out. It seems there are predetermined sequences in the market:
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 12 KB


So why did I say above that it can’t really be used predictively? Well, we can always use the Markov chain to estimate how often we'd be right with a forecast, how often we'd be right (or wrong) more than x times in a row with our trades. But does the beginning or middle of an uptrend phase really indicates that we're actually in an uptrend? No! They can also appear within downtrends. To get predictive power, we'd need to somehow combine the patterns we see with an actual uptrend, range, downtrend, or different intensities of these phases in order to determine the real forecasting capability of the model.

By "real" I mean that, of course, we want to predict the overall market status - but we'd also need to know the transitions in real time, without lag, to take advantage of potential trend phase shifts. (I mean, we only need it to evaluate or prediction quality; Knowing the true trend in real time would mean that we don't need to predict anything. But it's fine if we can easily determine the true trend in hindsight to "boost" our Markov prediction machine.) That's the point where I pretty quickly hit a wall, and that's why I eventually stopped looking at Markov models alone.

Any thoughts on that?
  • #4
  • Sep 2, 2025 12:43pm Sep 2, 2025 12:43pm
Quoting AntiCre
{quote} Hey carddard, thanks a lot for your reply to that old thread. It shows you’re genuinely interested in the topic itself and that you have the courtesy not to leave even older but interesting questions unanswered. Kudos! What I'd still be curious about, though, is which market situations the panels are supposed to represent. Yes, I've played around with this a few times in the past. Here's something I put together quite a few years ago. I didn't really get very far with it though - I've always seen it more as a descriptive tool than as a predictive...
Hi AntiCre would you please tell me about your Mac indicator first I put mac indicator then from navigator I add another Mac (apply to first indicator) then another one. in your picture you have 1 zero line. but I have 3 zero line.
  • #5
  • Sep 2, 2025 4:00pm Sep 2, 2025 4:00pm
Quoting k.k
{quote} Hi AntiCre would you please tell me about your Mac indicator first I put mac indicator then from navigator I add another Mac (apply to first indicator) then another one. in your picture you have 1 zero line. but I have 3 zero line.
Right-click on every indicator manually, select the indicator's properties and set the same fixed minimum and maximum. Hope it helps!
  • #6
  • Sep 6, 2025 2:06am Sep 6, 2025 2:06am
Quoting AntiCre
{quote} Hey carddard, thanks a lot for your reply to that old thread. It shows you’re genuinely interested in the topic itself and that you have the courtesy not to leave even older but interesting questions unanswered. Kudos! What I'd still be curious about, though, is which market situations the panels are supposed to represent. Yes, I've played around with this a few times in the past. Here's something I put together quite a few years ago. I didn't really get very far with it though - I've always seen it more as a descriptive tool than as a predictive...
Hi AC, the panels are basic representations of an uptrend / downtrend. It's my first time looking into Markov chains, so I started with that as my first experiment. What you share is interesting! I don't have the full context, but based on my limited understanding of Markov chains, here's what I would do:

1) You are right that the Markov chain describes the probability of the next state based on the last few. It tells you the immediate, short-term odds. But it may fail to predict the larger, more meaningful market direction (Let's call this the 'true trend'). What I shared in my previous post does exactly that. It tells me the probability of the next x minutes, but it's completely blind to whether a chart of another timeframe is in a massive uptrend or downtrend. But it may not matter if my intention is to just trade the next candle (at least in theory).

2) The Markov model is context-blind. You are absolute right that any short-term prediction could be a pullback / part of a trend in the opposite direction against you.

3) I hypothesize that to get predictive power, we'd need to combine the patterns. Some ideas could be using a trend filter (to group probabilities by regime), or use a volatility filter to achieve the same.

This are just thoughts at this point, will need testing!
  • #7
  • Sep 8, 2025 4:38am Sep 8, 2025 4:38am
Quoting carddard
3) I hypothesize that to get predictive power, we'd need to combine the patterns. Some ideas could be using a trend filter (to group probabilities by regime), or use a volatility filter to achieve the same. This are just thoughts at this point, will need testing!
Dear carddard,

thank you very much for your reply! What you wrote under point 3 is something that keeps being suggested, along the lines of “trade long in an uptrend when the swing also indicates an upward movement.” I think I’ve tried something similar in the past, but it doesn’t seem to be that simple.

To consolidate resources, today I went back to using CP’s market classification: we take the longest EMA, an EMA with half the calculation period, and a short-term EMA to track price movement. This gives us six market conditions. However, to identify the same six phases on a smaller scale, I wanted a swing phase to always end when the trend phase changes. I achieved this by using another intermediate EMA, which in turn has half the calculation period of the long EMA. I then used the long, intermediate, and short-term EMAs, which gives only three possible phases for the swing within a given trend phase. In total, that results in 18 combinations.

Let’s take a look at the Markov model:
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 69 KB

It’s remarkable how clearly we can see certain predefined paths along which the market moves. Of course, there are deviations from these paths, but they are less likely (see the thickness of the edges or their weights, which indicate these probabilities). That sounds promising at first. But what does it actually give us?

To find out, I calculated the average up and down movements over 10 weeks in GBPUSD (tick data) from 2018. The most striking result comes when we plot the average up and down price changes within the phases against each other for the 18 market states (BTW - I've used a similar terminology (differentiation between UP & D(OW)N, "new phase" = "start", "start mid phase" = "mid") here as CP used in this post (thus you may wonder whether this really adds to his success):
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 23 KB
What can I say? No phase really has an advantage, as the magnitude of the up and down movements within each phase is almost identical.
  • #8
  • Sep 8, 2025 7:17am Sep 8, 2025 7:17am
Quoting AntiCre
{quote} Dear carddard, thank you very much for your reply! What you wrote under point 3 is something that keeps being suggested, along the lines of “trade long in an uptrend when the swing also indicates an upward movement.” I think I’ve tried something similar in the past, but it doesn’t seem to be that simple. To consolidate resources, today I went back to using CP’s market classification: we take the longest EMA, an EMA with half the calculation period, and a short-term EMA to track price movement. This gives us six market conditions. However,...
Hi AntiCre,

Have followed some of your post over the years & just read this one & wanted to ask would Displaced MA's work for what you're describing with the EMA's. Displace MA's Example = EMA's 199 & -99, 99 & -49, 49 & -24. Just a thought. Your view would view would be most appreciated.
  • #9
  • Edited 4:23am Sep 9, 2025 4:05am | Edited 4:23am
Quoting acetrader
{quote} Hi AntiCre, Have followed some of your post over the years & just read this one & wanted to ask would Displaced MA's work for what you're describing with the EMA's. Displace MA's Example = EMA's 199 & -99, 99 & -49, 49 & -24. Just a thought. Your view would view would be most appreciated.
Thank you so much for your excellent question, acetrader! I actually read your message yesterday and had some time to reflect on it. Still, I went ahead and ran the test today, shifting each EMA back in time by half of its calculation period.

We know that all MAs follow price. It still seems (at least to me) that the FX market largely behaves like a random walk. That means any signal we derive directly from price will only ever have about a 50/50 chance. Unfortunately, I’m seeing the same behavior as with the original EMAs: in every combination of CP’s market phases, we get an identical number of upward and downward price movements:
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 12 KB

n the end, every prediction is bound to fail. Following the market, as CP preached, doesn’t seem to be captured by his market-phase classification through EMAs. That said, in a prolonged trend this classification does work well to gauge the potential of a market move - but it still doesn’t help in determining what you can actually expect to happen.
  • #10
  • Sep 10, 2025 8:44am Sep 10, 2025 8:44am
So the question remains: what’s the difference between ‘using price prediction’ and ‘following the market’ (which is essentially what CP & Co. are doing as well with their EMA-based indicators)? Yes, a philosophical question, isn't it?!
  • #11
  • Edited 1:58pm Sep 12, 2025 1:14pm | Edited 1:58pm
Any trading algo/idea/backtest/... can be framed as {trigger, expected_outcome}. E.g. we saw price closed > ema, we expect price to rise further.
The output will be one of the ordered pairs of truth values { {T, F}, {T, T}, {F, F} {F, T} }.


trigger - any imaginable Boolean expression of any length (close[1] > open[1], friday && RSI > 60, ...) - condition_1
expected_outcome - we check whether condition_2 evaluates T or F within some window.

So, I have some condition -> I wait x_time,
here is results for GU (futures 6B hourly since 2013):
Attached Image (click to enlarge)
Click to Enlarge Name: Cursor_ku09hX2HmE.png Size: 6 KB
{T, T} - true positive bullish signals
{F, T} - ... bearish
{T, F} and {F, F} are failed signals correspondingly
Do calculations within categories and we get 93.34% true bullish and 94.9% true bearish.
Too good to be true? I revised 4600+ plots (and that's for GU only) with my eyes to be sure.

Absorbing Markov chain is the best match for this logic, we can check step-wise probs.

Transient states - states that the system will eventually leave and never return to. - everything from the initial condition and within waiting window.
'U_3' (bullish, 3 steps remaining) are transient:
- Either a break occurs → go to absorbing state ;B'
- or time passes → go to 'U_2' (fewer steps remaining)

Absorbing states - states that once entered the system never leaves - expected_outcome
- 'B' - expected_outcome
- 'U_0', 'D_0' - no more time steps remaining
Attached Image (click to enlarge)
Click to Enlarge Name: Cursor_i599kVLenO.png Size: 6 KB

Attached Image (click to enlarge)
Click to Enlarge Name: Cursor_G8jkYqiXjZ.png Size: 6 KB
(0.008 points = 80 pips)

I then tested the bullish signals against of buying after a random bullish hourly candle:
below the median they're quite the same. But - S1 has MUCH less false signals.
Attached Image (click to enlarge)
Click to Enlarge Name: Cursor_uzHMU5g5qO.png Size: 4 KB

I encourage everyone to test their findings against random buy/sell... Because, you know, maybe it's not a finding.

Probabilities don't change over time. Of course it can fluctuate in the short run just like head/tail ratio when flipping a coin - but when analyzing in chunks, we have ~same underlying probability
Air force
  • #12
  • Sep 14, 2025 9:04am Sep 14, 2025 9:04am
Quoting W0lfram
I revised 4600+ plots (and that's for GU only) with my eyes to be sure.
Forgot to attach plots, random sample (bullish signal)
Attached Image (click to enlarge)
Click to Enlarge Name: evo.png Size: 53 KB


Querying a dataset with logical operations and exploring its future is a simple and effective practical method

I'll test some links from google later.
https://medium.com/@deepml1818/pytho...e-c54d2710f71b
https://www.quantstart.com/articles/...s-in-qstrader/
https://blog.quantinsti.com/regime-a...rading-python/

https://github.com/search?q=markov%2...e=repositories
Air force
  • #13
  • Sep 16, 2025 3:25am Sep 16, 2025 3:25am
Quoting W0lfram
Any trading algo/idea/backtest/... can be framed as {trigger, expected_outcome}. E.g. we saw price closed > ema, we expect price to rise further. The output will be one of the ordered pairs of truth values { {T, F}, {T, T}, {F, F} {F, T} }. trigger - any imaginable Boolean expression of any length (close[1] > open[1], friday && RSI > 60, ...) - condition_1 expected_outcome - we check whether condition_2 evaluates T or F within some window. So, I have some condition -> I wait x_time, here is results for GU (futures 6B hourly since 2013): {image}{T,...
What a shame I haven't seen your posts earlier! However, I need to digest this first. The idea of testing how the “expectation” changes over different time intervals is a way to see how long one should trust the signal. Unfortunately, running such code on tick data in R is always quite slow. Still, it’s interesting to observe how quickly the expectation horizon decays (think this is what one of your images show, too). Of course (imo), this isn’t really a result of the indicators themselves (they can only shorten it slightly), but rather of the price behavior itself.

I’m definitely very curious to see whether the HMMs will actually prove useful in allowing a quick enough switch to the most appropriate market model (though I do have some inner doubts about that). In any case, please do share your future findings here! Many thanks - great stuff as always!!!
  • #14
  • Sep 18, 2025 4:07am Sep 18, 2025 4:07am
dīmēnsiō (genitive dīmēnsiōnis) → meaning "a measuring, measuring out, quantity."

https://www.forexfactory.com/thread/...83#post6584683

How do you design a system that remains self similar - IID random - within each boundary, identifies transitions and at the same time recalculates when the market empirically shifts to non-IID, fat tail behaviour?
  • #15
  • Sep 21, 2025 10:29pm Sep 21, 2025 10:29pm
Here is a good direction for you to play with it. Instead of using HMM to generate entry signals, it's more useful as a filter. It goes the same for other machine learning algorithms. However, you need a good strategy at first.
  • #16
  • Sep 22, 2025 1:53pm Sep 22, 2025 1:53pm
Quoting zkchyo
Here is a good direction for you to play with it. Instead of using HMM to generate entry signals, it's more useful as a filter. It goes the same for other machine learning algorithms. However, you need a good strategy at first.
Will it be a quality improver or sample size reducer kind of filter?
Any examples?
Air force
  • #17
  • Sep 22, 2025 6:53pm Sep 22, 2025 6:53pm
Quoting W0lfram
{quote} Will it be a quality improver or sample size reducer kind of filter? Any examples?
Quality improver. Let's say you have a good strategy. With HMM, you could find when to bet bigger and when to bet smaller. It could affect equity curve well. Never try to use ML to generate signals. It's always overfitting due to the unstable nature of financial data series. It's never stationary. From 2020 to now, I've seen enough "promising" trading models fail in reality.
  • #18
  • Sep 22, 2025 7:13pm Sep 22, 2025 7:13pm
Quoting zkchyo
Here is a good direction for you to play with it. Instead of using HMM to generate entry signals, it's more useful as a filter. It goes the same for other machine learning algorithms. However, you need a good strategy at first.
they can’t replace a well-grounded strategy. The edge still comes from the underlying logic, not the algorithm.
  • #19
  • Sep 24, 2025 12:23pm Sep 24, 2025 12:23pm
Quoting zkchyo
{quote} Quality improver. Let's say you have a good strategy. With HMM, you could find when to bet bigger and when to bet smaller. It could affect equity curve well. Never try to use ML to generate signals. It's always overfitting due to the unstable nature of financial data series. It's never stationary. From 2020 to now, I've seen enough "promising" trading models fail in reality.
In any case we'll get sample reducer. Quality is debatable.
Like in this Hypothetical distribution:
Attached Image (click to enlarge)
Click to Enlarge Name: WolframNB_r74rpSwen4.png Size: 14 KB

Quote
It's always overfitting due to the unstable nature of financial data series. It's never stationary.
Only the price itself is not stationary, i.e. modelling "close price EOD" etc. is doomed to failure. Bc the market is highly efficient = random walk behavior.
Stationarity is easily achievable if needed.
Even the count of ticks inside 1m candle has its own distribution and stays the same over the years.


Quote
From 2020 to now, I've seen enough "promising" trading models fail in reality.
Have you ever seen unfounded claims like that: "Let's say you have a good strategy. With HMM, you could find when to bet bigger and when to bet smaller". - failing in reality?
Maybe one could, maybe couldn't. Any examples?
Air force
  • #20
  • Edited 2:50pm Sep 24, 2025 1:40pm | Edited 2:50pm
Quoting AntiCre
{quote} What a shame I haven't seen your posts earlier! However, I need to digest this first. The idea of testing how the “expectation” changes over different time intervals is a way to see how long one should trust the signal. Unfortunately, running such code on tick data in R is always quite slow. Still, it’s interesting to observe how quickly the expectation horizon decays (think this is what one of your images show, too). Of course (imo), this isn’t really a result of the indicators themselves (they can only shorten it slightly), but rather...

Well, the links from google weren't very useful. The code in one of them didn't work at all

Anyway I was researching and trying to implement HMM filtering.
Some brief results:
run_up - best unrealized profit during the signal
S1 & S2 are from the previous post.
And HMM filtering applied to S1
We have slightly shifted mode and higher variance. But no difference between S1 and HMM. - it only reduced the S1 sample by a factor of 2.
Attached Image (click to enlarge)
Click to Enlarge Name: WolframNB_8DKGaMaXqo.png Size: 10 KB
Attached Image (click to enlarge)
Click to Enlarge Name: screenshot.png Size: 5 KB




Quote
It still seems (at least to me) that the FX market largely behaves like a random walk. That means any signal we derive directly from price will only ever have about a 50/50 chance.
Can't agree more.


Suppose there are 2 approaches:
- random entry, wait and hope for n_hours (fundamental stats of price excursion) - then check for run_up and drawdown
- we trade any indicator with exits at opposite signals - for profit we'll have a normal distribution, for run_up it's the same quantiles assuming we test everything on a hourly dataset

Attached Image (click to enlarge)
Click to Enlarge Name: Cursor_JaQ6I3kF2i.png Size: 14 KB
Air force
0 traders viewing now