Introduction

This document was built by filtering the data based on the following criteria:

Please note that for the “Win rate by Elo” plot, all matches that have the lowest Elo player in the match having an Elo greater 800 have been included.

Descriptives

Distribution of Matches Played by Patch

Distribution of Matches Played by Map

Distribution of Matches Played by Game Length

Distribution of Matches Played by Mean Team Elo

Play Rate by Civilisation

Naive Win Rates

Naive Win Rates by Civilisation

Naive Win Rates vs Play Rate

Averaged Win Rates

Averaged Win Rates by Civilisation

Averaged Win Rates vs Play Rate

Miscellaneous Plots

Naive Win Rates by Elo

Naive Win Rates by Game Length

Civilisation v Civilisation Win Rates



Experimental

This section contains outputs that are still being tweaked or reviewed and maybe removed or changed in the future. If you have any feedback or suggestions about them please let me know

Distribution of Players Highest Picked Civilisation’s Play Rate

For this plot we calculate the play rate of each players most played civ (i.e. if I used Franks for 60% of my games, Mayans for 30% and Britains for 10% I would get a value of 60). We then categories the counts into brackets of 10 (i.e. 0-10, 10-20, etc) and count how many are in each bracket. The idea is this should give some indication of how many people are playing random vs how many are 1-civ pickers.

Hierarchical Clustering Dendrogram

This output attempts to highlight civilisations that are “similar” based upon their win rates vs other civilisations. The algorithm works by recursively grouping civilisations (or groups of civilisations) that are the most similar to each other until there is only 1 group. The lower down on the y-axis that civilisations are grouped indicates a higher degree of similarity whilst civilisations that are grouped higher up on the y-axis indicates a lower degree of similarity. That is to say that if two civilisations are linked together low down on the y-axis it means that they tend to win and lose against the same civilisations.

Estimating how Overrated or Underrated each Civilisation is

The following plots attempt to answer the question of “can we tell how overrated / underrated each civilisation is” (credit to SOTL for proposing this idea).

In order to try and quantify this, play rates are normalised using a box-cox transformation and then a robust linear regression is fitted and used to estimate what the expected win rate should be for each civilisation.

The idea is that play rates could be used as a surrogate to indicate what people’s expectations of each civilisation are; that is, civilisations with high play rates are those that people think will do well whilst civilisations with low play rates are those that people think will do badly.

The following is then a plot of the difference between the expected win rate and the observed win rate:

For context the following plot shows both the observed and expected win rates

I must stress that these plots should be interpreted with a massive grain of salt. The fundamental assumption that play rates predict win rates is a strong one (for example I main Vietnamese despite knowing they perform below average). Likewise pick rates heavily bias win rates due to interactions with the Elo system as well as peoples skill/familiarity with the civ. These are biases that the model can’t account for and to which the results are most susceptible to as it is intrinsically the extremes that we are most likely to be interested in. This is all to say that these results should be regarded more as an informal conversation starter rather than as an exact science / analysis.