A natural place to start, depending upon your mathematical background might be the modern cardinal generalizations of the voting literature. Consider the following toy model, and feel free to adapt any parts which don't seem appropriate to the problem at hand.
Data: Say we have a fixed, finite population of players $N$ who we observe play. A single observation consists of the result of a single match between two players $\{i \succ j\}$ for some $i,j \in N$. Let's the space of datasets $\mathcal{D}$ consists, for fixed $N$, of all finite collections of such observations.
Solution Concept: Our goal here is twofold: mathematically we're going to try to get a vector in $\mathbb{R}^N$ whose $i^\textrm{th}$ component is the 'measure of player $i$'s caliber relative to all the others.' A player $i$ is better than player $j$ by our rule if the $i^\textrm{th}$ component of this score vector is higher than the $j^\textrm{th}$. We might also like the magnitude of the differences in these components to carry be increasing in some measure of pairwise skill difference (the score vector is cardinal, to use economics lingo).
Edit: This score vector should not be seen as a measure of luck versus skill. But what we will be able to do is use this to 'clean' our data of the component of it generated by differences in skill, leaving behind a residual which we may then interpret as 'the component of the results attributable to luck.' In particular, this 'luck component' lives in a normed space, and hence comes with a natural means of measuring its magnitude, which seems to be what you are after.
Our approach is going to use a cardinal generalization of something known as the Borda count from voting theory.
Aggregation: Our first step is to aggregate our dataset. Given a dataset, consider the following $N\times N$ matrix 'encoding' it. For all $i,j \in N$ define:
$$
D_{ij} = \frac{\textrm{Number of times $i$ has won over $j$}- \textrm{Number of times $j$ has won over $i$}}{\textrm{Number of times $i$ and $j$ have played}}
$$
if the denominator is non-zero, and $0$ else. The ${ij}^\textrm{th}$ element of this matrix encodes the relative frequency with which $i$ has beaten $j$. Moreover, $D_{ij} = -D_{ij}$. Thus we have identified this dataset with a skew-symmetric, real-valued $N\times N$ matrix.
An equivalent way of viewing this data is as a flow on a graph whose vertices correspond to the $N$ players (every flow on a graph has a representation as a skew-symmetric matrix and vice-versa). In this language, a natural candidate for a score vector is a potential function: a function from the vertices to $\mathbb{R}$ (i.e. a vector in $\mathbb{R}^N$) such that the value of the flow on any edge is given by the its gradient. In other words, we ask if there exists some vector $s \in \mathbb{R}^N$ such that, for all $i,j \in N$:
$$
D_{ij} = s_j - s_i.
$$
This would very naturally be able to be perceived as a metric of 'talent' given the dataset. If such a vector existed, it would denote that differences in skill could 'explain away all variation in the data.' Generally, however, for a given aggregate data matrix, such a potential function generally does not exist (as we would hope, in line with our interpretation).
Edit: It should be noted that the way we are aggregating the data (counting relative wins) will generally preclude such a function from existing, even when the game is 'totally determined by skill.' In such cases our $D$ matrix will take values exclusively in $\{-1, 0 ,1\}$. Following the approach outlined below, one will get out a score function which rationalizes this ordering but the residual will not necessarily be zero (but will generally take a specific form, of a $N$-cycle conservative flow that goes through each vertex). If one were to, say, make use of scores of games, the aggregated data would have a cardinal interpretation.
Construction of a score: The good news is that the mathematical tools exist to construct a scoring function that is, in a rigorous sense, the 'best fit for the data,' even if it is not a perfect match (think about how a data point cloud rarely falls perfectly on a line, but we nonetheless can find it instructive to find the line that is the best fit for it). Since this has been tagged a soft question, I'll just give a sketch of how and why such a score can be constructed. The space of all such aggregate data matrices actually admits a decomposition into the linear subspace of flows that admit a potential function, and its orthogonal complement. Formally, this is a combinatorial version of the Hodge decomposition from de Rham cohomology (but if those words mean nothing, don't worry about looking them up). Then, loosely speaking, in the spirit of classical regression, we can solve a least-squares minimization problem to orthogonally project our aggregate data matrix onto the linear subspace of flows that admit a potential function:
$$
\min_{s \in \mathbb{R}^N} \|D - (-\textrm{grad}(s))\|^2
$$
where $-\textrm{grad}(s)$ is an $N\times N$ matrix whose $ij^\textrm{th}$ element is $s_i - s_j$.
If you're interested in seeing this approach used to construct a ranking for college football teams, see:
http://www.ams.org/samplings/feature-column/fc-2012-12
If you'd like to read a bit more about the machinery underlying this including some brief reading about connections to the mathematical voting literature, see:
https://www.stat.uchicago.edu/~lekheng/meetings/mathofranking/ref/jiang-lim-yao-ye.pdf
