0
$\begingroup$

I have been struggling to fit the following problem into a linear state space model for a Kalman Filter (KF). I'm having a hard time seeing what I'm doing wrong. I suspect I'm violating some law of KF state space modeling. Perhaps something is non-gaussian, non-linear, non-kosher?

I have data from players who score points in a game each day. The game involves both skill and luck. With a KF in mind I am thinking about skill as the unobserved state and luck as part of the observation noise. Here is some (contrived) data:

        Points Scored
Player  Day 1   Day 2   Day 3
1       10      118     50
2       15      100     49
3       20      130     51

I am also assuming players are endowed with a skill level (relative to the other players) which does not change over time. Further, points scored are exogenously and randomly mean shifted/scaled each day. As you can see in the example data above, Player 3 is more skilled because he scores more points than the other players. Also, the distribution of points moves and scales -- that is to say means and variances of points scored change from day to day.

I'll index players $i=1,\dots,N$ and time $t=1,\dots,T$. Let $y_{it}$ be the points scored by player $i$ at time $t$.

My first approach was to estimate skill using something like a z-score (or a t-stat). For each player at each time step I compute: $z_{it} = \frac{y_{it} - \bar{y_t}}{s_{y_t}}$

Then I average across time to estimate skill: $\widehat{skill_i} = \bar{z_i}$

As far as I can tell, so far so good. Now I want to Kalmanize this thing so I can update estimates of skill as new data arrive. I start to think about the state and its relationship with the observable points scored. But this is where I get stuck.

Suppose the true, unobservable skill level for a player $i$ is $\theta_i$ and this comes from a gaussian distribution with mean and variance $(\mu_\theta, \sigma_\theta^2)$. Luck is $\epsilon_{it}$ with mean and variance $(0,\sigma_{\epsilon}^2)$ Further suppose the common daily mean shift and scaling factor are $\alpha_t$ and $\gamma_t$, respectively, where both come from gaussian distributions with means and variances $(\mu_{\alpha}, \sigma_{\alpha}^2)$, and $(\mu_{\gamma}, \sigma_{\gamma}^2)$.

I model the observed points scored by player $i$ at time $t$ as: $y_{it} = \alpha_t + \gamma_t (\theta_i + \epsilon_{it})$

The state transition equations are easy: $\theta_{it} = \theta_i$. Each player's skill is assumed to be constant across time.

The player's luck is the observation noise.

Although it seems like I have everything defined right, I can't quite seem to figure out how to make $\alpha_t$ and $\gamma_t$ fit into a KF model - should they be part of the state? If so, how can I get a linear KF to multiply states? If they aren't part of the state, but are time-varying coefficients, how can I make that work?

And of course I should also be asking: Is the Kalman Filter even an appropriate estimation strategy for skill?

Help! So many thanks in advance!

$\endgroup$
0
$\begingroup$

If $\alpha_t$ and $\gamma_t$ are known you have a linear state-space model. Even if $\alpha_t$ is unknown and you estimate it online (i.e. add another state) the problem can be solved using KF. However, if $\gamma_t$ is unknown the problem is no longer linear in the states, your measurement model is then non-linear.

Edit: Just to clarify a bit, time-varying constants are not a problem in the KF. The part that makes it "non-linear" is $\gamma_t(\theta_i+\epsilon_{it})$.

$\endgroup$

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.