How are you constructing your markov chains? What inputs are you using?
I have the raw input data redimensioned into 'bins' of similar amounts of BTC volume. You might think of these bins kind of like candles, with a high and low price, a volume weighted average, etc. The markov is constructed in part from discrete samples the percent change of the VWA from bin to bin going back in the history by a certain number of samples. But I selectively invert these samples depending on if they are a member of either an uptrending or downtrending price regime as identified by what I have been calling a "reversal indicator" which is really troublesome to describe.
Here'e why I am doing this. The reversal indicator shows pretty clearly when we change between uptrending and downtrending price regimes, but only after the fact. It is pretty easy to look at an indicator after the fact to see what it is showing, that's no big whoop. So what I am trying to do is to see if I can project this indicator to see if I can tell when it is about to turn. That's what this is all about.
I see so are you trying to detect a VWA candle in real-time by sampling the chain to detect if its a reversal? How much data do you train it with? I would assume all of the daily candles (market is not that old yet to train fully?) Maybe do a weekly or monthly to create a multi-time frame analysis to improve lower time-frame confidence.
Why can't you use yesterdays daily candle to determine if it was a reversal or not? That is not after the fact?
What about defining the signature of a reversal? Sometimes a sequence of candles provide a higher likely hood that it is a reversal situation. Would this model be able to sample based on a set of related inputs?
How are you defining the up/down trend price regimes in order to invert some samples?
I actually base the tuning on a Bayesian posterior analysis. The calibrations in question are a minimum sample size for the first step in the chain, and the sample size grows at a certain calibrated rate with each subsequent step. I don't care to share exactly what those calibrations are. But I really want to emphasize that time is not used at all as an independent variable, I am only considering the passage of trade volume. Many tests have demonstrated to me that this is key to reducing variance.
As far as the up/down goes, that is based on the reversal indicator. What this does basically is to compare the difference between how BTCOBV and $OBV change in relation to price. The response of this indicator is much sharper at transitions of trend regimes than can be seen just by looking at price. My working theory is that it shows when either BTC or $ at play during the last regime become nearly exhasted, and the trend reverses.
If you're talking about using volume over time which is fair you've really opened a can of worms lol Volume based analysis is a study on its own... THis is why I asked about the signature of a reversal because this can be either:
1) High volume Squat or Doji with stopping volume
2) No Demand bars (end up uptrend) with low volume
3) Accumulation bars with high volume (end of downtrend)
4) Upthrust (end of downtrend or continuation of uptrend by pushing through supply)
I would suggest reading Mastering the markets which focuses on using Volume to determine price action:
http://vsa.pipbuilders.com/mtmv3.pdfJag