Apologies if this is not the correct topic for this question.
I am looking for a general approach/potential references/terms to search for regarding the following situation or similar situations as it seems like there should be references on it, but I haven't been able to find them.
Let's say that we're observing the real-time last traded prices for some asset, denoted as $s_0, s_1, \dots$
First, we observe $s_0.$
Then, we use some predictive model, $P$, to generate a prediction (point, probabilistic, or otherwise) for $s_1$: $$p_1 = P(s_0)$$
Then, we use some decision model, $D$, to decide on an action to take based on the previously generated prediction: $$ d_1 = D(p_1)$$
Then, for the sake of addressing the question, assume that $d_1=-1$ or $d_1=1$, representing the action of either shortselling or buying one unit of the asset at market price, respectively.
This is where my intuition for how to address this situation breaks down.
By acting on $p_1$ through either shortselling or buying a unit of the asset (and assuming we have instant execution of our order), we affect the value of $s_1$ as we decided how to enter the trade through $d_1$ based on our prediction $p_1$.
I feel like this overlap would create potential problems with regards to evaluating the performance of the predictive and decision models.
Is there any framework for evaluating predictions and decisions when the decisions can affect future observations (that we're comparing the predictions against) utilizing only the observations recieved over time? or is this just one of the disadvantages of using something like transaction data versus modeling something with more detail regarding the impact of a trade such as the limit order book?