1
$\begingroup$

Suppose I have a very simple asset whose price takes only three possible values: $X_t\in \{-1,0,1\}$. I also got some discrete time series $X = (X_t)_{t\geq 0}$ and I would like to come up with a trading rule based on these observations. Let's focus on the following naive approach: given the current level of the asset, I would like to estimate of what the next change would be. Thus, I am fitting this time series into Markov Chain where I disregard transition that do not change the state. For example, based on the following sample from the time series: $$ \dots0,0,0,0,0,0,0,1,1,1,0,0,0,-1,-1\dots \tag{1} $$ and suppose that there are no appearances of $0$ anymore. I can conclude that out of $10$ appearances of $0$, $7$ are followed by $1$ and $3$ are followed by $-1$, hence a naive algorithm would say that transition probabilities are $p(1|0) = 0.7$ and $p(-1|0) = 0.3$.

Now, I wanted to make it faster, so I preprocessed the data to get rid of repetitions since I thought that it shall not affect the end result. For example, the sample $(1)$ transforms into $$ \dots0,1,0,-1\dots \tag{2} $$ but now $p(1|0) = 0.5$ and $p(-1|0) = 0.5$ which is quite different from the previous estimation. Of course, that's just a simple example, but it gives a general impression: the lumping procedure $(1)\to(2)$ changes the estimates of transition probabilities. It surprised me first, but now it seems very natural: in $(2)$ when sampling I give equal weight to each interval of consecutive $0$'s whereas in $(1)$ more weight goes to a longer interval of consecutive $0$'s.

The question is: given my purpose, what would be the correct method to estimate probabilities? Note that I am trying to predict the price move without taking into account for how long have I stayed at the current price of the asset, only the price level itself. From that perspective, I guess the second approach is more appropriate: if I observe new prices and I see the price is $0$, I'd say I'd rather rely on that the next price change happens with probabilities $(0.5,0.5)$ rather than $(0.7, 0.3)$. At the same time, here I do not feel confident since my background in statistics is rather weak, so any feedback on this topic is highly appreciated. Namely: what is formally more correct, to use probabilities from $(1)$ or from $(2)$, or if both are correct depending on how they are used, what is a proper way of using them. Practical comment on that approach is also welcome.

$\endgroup$
1
$\begingroup$

It sounds to me like you have a Markov model that is not "lumped", it's just that certain transitions don't provide you with any payout. I would model the true transition probabilities.

Now, let's ask what the probability of getting a one is, assuming that we won't stay at zero, i.e. $P(X_1=1 | X_0=0, X_1 \not = 0)$. We recall that $$P(A|B)=P(A,B) / P(B)$$ and therefore get

$$ \begin{align*} P(X_1=1 | X_0=0, X_1 \not = 0) &=\frac{P(X_1=1, X_1 \not = 0 | X_0=0)}{P(X_1 \not = 0| X_0=0)}\\ & =\frac{P(X_1=1 | X_0=0)}{P(X_1 = 1| X_0=0) + P(X_1 = -1| X_0=0)} \\ & =\frac{P(1|0)}{P(1|0) + P(-1|0)} \end{align*} $$

Which is just your second definition where you found it to be $(.5,.5)$ as indeed half the times it transitioned away from zero it transitioned to one.

Importantly, this is not $P(1|0)$! I think this is where you were getting confused before.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.