Smart kills and worthless deaths: eSports analytics for League of Legends

1.1 Description of League of Legends

LoL, often referred to as LoL or simply League, is a five versus five computer video game reminiscent of capture the flag. The two teams start in their respective bases on opposite corners, bottom left and top right, of a largely jungle-like environment with three main lanes connecting the two bases, referred to as top, mid, and bot. Each team’s own nexus is placed near its base, with two turrets protecting it. There are also three other turrets on each half of the map along each lane, as well as one inhibitor per team for each lane. You can’t attack the enemy nexus unless at least one of the enemy inhibitors are down, and you can’t attack an inhibitor unless all of the turrets leading to it have been destroyed. Turrets, once destroyed, never respawn, but inhibitors do. The first team to destroy the enemy’s nexus wins the game. Nothing else ultimately matters: not the number of kills or deaths of each player, or how much gold was earned, or even how many total turrets or inhibitors were destroyed. This makes ends of games especially dramatic, including back-door plays and sneaky attacks and last-second races.

Each of the 10 champions can go anywhere they want on the map, and there are essentially no enforced or special positions such as a goalie for whom different rules apply. Much like basketball and the evolution of point guard, shooting guard, etc., positions, even though each of the five players on a LoL team could in principle do anything, a standard form of specialization has evolved. Typically, one player on each team plays the top lane, one plays the mid lane, and two play the bot lane. In the bot lane, one is called the “carry” or the primary damage dealer, and one is the support. The fifth player is called the jungler and tends to roam all over the map and kill the neutral minions found in the jungle.

Minions are non-player characters in the game, and come in two types: neutral ones found in the jungle that any player on either team could kill, and team ones who are like robotic soldiers for each team. Killing a minion grants gold. If it takes multiple hits to kill a minion, only the champion whose damage actually killed the minion gets the gold. Gold can be exchanged at each team’s base for items that upgrade the champion with stronger or new abilities. Thus, in the bot lane, the carry tends to be the champion who kills the minions while the support often heals or offers shields or tries to damage the enemy. That way, the carry earns more gold and gets better items faster. Some items are primarily offensive in nature in that they increase the champion’s combat abilities. Some are primarily defensive in nature in that they make it more difficult for enemies to kill you. Some are a mixture, or provide utility such as vision or shields for allies. Champions that have built primarily defensive items are called tanks and champions that have built primarily for damage are called squishies, because they tend to be easy to kill, so long as you can get past their ally tanks and other teammates looking to protect or “peel” for their carry.

Champions also gain experience or “XP” and level up at certain benchmarks if they are near things that die that are not on their team, for example enemy champions or minions or neutral minions. Leveling up allows you to upgrade one of your champion’s base abilities. For example, you might have an ability that lets you spin over some distance; when you upgrade it to a higher level, perhaps the cooldown is reduced, or the damage it deals is increased, or both. All champions begin the game on level one and can reach a maximum level of 18. Most champions have an “ultimate” ability that is unlocked only when they reach level six. This ultimate ability or “ult” is usually the strongest ability of the champion, causing, for example, a vast amount of damage.

See Appendix 1 for further details about the game.

As is evident from the above description and the Appendix, League is a complex game. It is also constantly evolving, with new patches changing or modifying rules or champions approximately every few weeks. The standard data available to analyze is woefully insufficient to properly analyze gameplay.

1.2 League of Legends data

Using some techniques outlined in this paper, we can collect more comprehensive and granular data. We track every champion’s location multiple times every second. We also track every ability cast, every attack made, and all damages caused and avoided. We track actual vision granted and denied in the fog of war. We track every player’s health and mana and cooldowns. We track everything, continuously. And we do it invisibly, remotely, and live, all without any impact on the user, or even their knowledge. (There are no privacy concerns here: we only track games that allow public spectating.)

From this comprehensive data, we are also able to define much better stats and analytics than what is possible from the API. A myriad of things critical to understanding and evaluating the game cannot be determined from the API data. When you died, how many allies were around you? How many enemies? Why did you die? Was it for lack of vision? A deficiency in health or gold spent? Outnumbered? Ganked? Or simply mechanically outplayed? Was your death a worthless one, or did your team come out ahead overall? If you killed someone, was it a smart kill, meaning one that increased your team’s likelihood of winning, or was it merely a kill that padded your stats but did not help your team? How much time did you spend doing productive activity versus how much time was wasted? In team fights, were your formations correct, tanks in the front, squishies in the back? Did you peel? Did you engage and disengage correctly? How effectively did you use your summoner spells, and did you use them offensively or defensively?

Without this new data, there is no systematic way to answer any of these crucial questions, and many others. Yet these are exactly the kinds of questions that LoL players, casual and professional, ask themselves when they review game film.

1.3 Our contribution

We began with LoL because it has by far the largest player base (Paul 2017) but our general methodology is applicable to virtually all eSports. Other MOBA sports either do provide more comprehensive data directly from the game publishers, such as Dota2, or have other tools that have extracted data, such as Matz (2017) for Overwatch, so that particular portion of our methodology would not even be necessary there. We hypothesize as a direction for future research that our findings about individual contributions to team performance would, however, carry over. Furthermore, we argue that insights from eSports analytics can in principle inform analytics in traditional sports as well, because of the vastly larger data that eSports can provide.

We analyzed millions of games of LoL to extract the high-frequency and high-fidelity eSports data. Next, we developed and calibrated a live, in-game win probability model based on the conditions of the game at each moment (see Figure 1). Finally, while we tracked all deaths and kills, we improved the standard metrics to focus on worthless deaths and smart kills. Worthless deaths are ones that do not increase your team’s win probability. Smart kills are ones that do increase your team’s win probability. With these two new metrics, suddenly the relation between individual performance and team performance is virtually a straight line. Figure 2 shows the difference between the upward sloping but relatively flat relationship of the standard stats on the left-hand side, and the practically straight one-to-one relationship of our new advanced stats on the right.

Figure 1:

Win probability.

Figure 2:

Regular deaths and worthless deaths versus team win probability.

Not only do these results suggest specific new ways forward for traditional sports, they also provide a possible novel method for generating brand-new insights: try things on eSports, where the data is far larger and cleaner, and carry those lessons back to the non-e-sports. Finally, with eSports itself poised to take over traditional sports in the future—LoL tournament viewership, for example, exceeds FIFA World Cup viewership (Keiper et al. 2017)—these insights will be valuable for their own sake as well, in addition to improving all other sports analytics.

3.1 Data methodology

We have developed three independent methods for extracting detailed gameplay data. This section explains the basic principles behind our approaches in a high-level overview. The full details and all source code for all methods, both front-end and back-end, are open sourced and available online.

When 10 humans (referred to as “summoners” to distinguish them from the specific “champions” or avatars they control) play a round of LoL, and allow public spectating of the game, anybody can choose to watch the game live as it unfolds. The normal use cases for this are friends watching each other play, professionals streaming for their coaching staff, and commentators discussing a live professional game. When you spectate someone else’s game, you do it through the same game client as if you were playing a game yourself. Rather than being a static video, you have various controls, including camera controls, time controls, and pop up information. You can see both teams if you choose. To avoid cheating, Riot broadcasts spectated games on a 3 min delay. That way, you can’t have a friend spectate your game and tell you where the enemy is.

The network architecture is such that a player’s actions are sent to Riot’s servers, which respond with updates of the game. This happens quickly so that players experience neither a lag between their issuance of a command and the ensuing movement, nor erratic enemy activity. This bidirectional network traffic is highly encoded, difficult to decipher, and subject to change with every patch. Patches happen regularly in LoL, approximately every few weeks.

For our methods of data collection, we rely on the fact that the LoL game client is running and decoding the network traffic appropriately. The results are displayed on the monitor exactly as they would if you were playing the game. Figure 3 shows a snapshot of a spectated game.

Figure 3:

A single-frame image snapshot of a league of legends game in spectator mode.

Note the plethora of information. While the center of the screen shows the action, and is the primary focus of a live gamer, the top shows the overall progress of each team in terms of gold, turrets, and kills, the bottom shows gold earned, items purchased, and the basic kills/deaths/assists stats, and the left and right champion bubbles show health, mana, level, and important cooldowns, but the most important information is in the bottom right minimap. The minimap shows the locations of all 10 champions, turrets, inhibitors, wards, traps, minions, and monsters.

The approach our eSports optical data tracking software takes is to visually process that minimap and infer all the locations it implicitly shows. We also track the other statistics and information from the other areas, but it is the minimap that is the most challenging and most rewarding. Technically, this approach relies on OpenCV, the industry standard open source computer vision library. Much of the difficulty in processing is due to champion overlaps and various other visually confusing information.

Here are some of the technical details for using computer vision. First, notice the white rectangle in the middle right of the minimap at the bottom right of Figure 3. Detecting this white rectangle is relatively straightforward, with a few edge cases: for example, it could be missing one of the sides. The white rectangle indicates what portion of the minimap is shown in the full view. The center of this rectangle indicates the position of the champion being tracked.

On the bottom left, note the various icons and numbers for the particular champion selected, in this case the blonde female mage called Lux. Each of those icons represents the current state of one of her abilities or summoner spells, including cooldown or additional effects. Her health, attack damage, ability power, and other stats are displayed in a standard font. In the bottom middle, a summary table of information about each champion lists the items they have purchased so far and their associated cooldowns if any, their kills, deaths, and assists, expressed as k/d/a, and the total number of minions killed. That screen can be toggled to also display each champion’s gold earned in a standard font and text.

Two of our approaches to data collection rely on such computer vision live tracking using the spectator mode of the game client, one written in Python and one written in Go: the Python version was quicker to develop but the Go version was more robust. However, both approaches had a major downside: to get all 10 champion positions throughout the game, the entire game had to be spectated 10 times, so that the appropriate white rectangle could be detected. Therefore, we implemented a third approach, using direct hooks into the game client to determine actions, in a manner similar to scripting software such as Quipko (2017) that had often been used by players to attempt to cheat at the game (though note that our approach to data collection made any form of cheating impossible as the hooks were into the observer client, not any player’s client).

In each case, cloud servers running Windows and game clients are automatically managed, starting and stopping in response to the volume of games requiring analysis. Each new instance launches and prepares the game client, activates spectator mode for the next game in the queue, and begins the optical tracking, shutting down when complete. The cloud instance needs an expensive graphics processing unit (GPU) for the game client to function; hence the motivation for the third approach to reduce the cost by 90%.

Table 1 lists some the data we are now able to extract from games for each champion, each second.

One analytical advantage of eSports over regular sports is the number of games played. With our data methodology in hand from the Section 2, we can let it loose on an ever-expanding set of players, following each of their matches, and adding all that data to our universe. We start with a list of the top LoL players. Players in LoL fall into one of these ranks: Unranked, Bronze 5 through Bronze 1, Silver 5 through Silver 1, Gold 5 through Gold 1, Platinum 5 through Platinum 1, Diamond 5 through Diamond 1, Master, Challenger. Ranks are determined essentially via an ELO process calculated from your performance, much like a chess rating. By definition, Challenger players are the best 200 players in the world.

Starting with this list of 200, we can spectate all their games. Sometimes their games will include nine other challengers, in which case we simply analyze the game and move on. Sometimes their games will include at least one other non-challenger, so we add that person to our list. And when that person plays games, he likely also plays either with or against other people not already in our corpus, so it continues to grow. Only future played games are spectated, so a player’s history is not imported. By this iterative search process, we scraped several million games.

Note that our corpus of millions of games spanned several LoL patches, each of which can significantly affect gameplay. Hence, we focused on the ∼150,000 games from the latest patch for the purposes of this section, which is still the equivalent of more than a full century of NBA seasons. Of those, about half were Bronze, a quarter Silver, 10% Gold, 7% Platinum, 5% Diamond, and less than 1% in each of Master and Challenger. Because all but about a thousand players had only a handful of games in our database, the breakdown of distinct players by tier and role shown in Table 2 follows approximately the same distribution.

3.2 Analytics methodology

In-Game Win Probability. Our first important analytics contribution is in calculating in-game win probability using a machine learning model.

While there are hundreds of potential metrics each second that could contribute to the win probability (including all the metrics in Table 1), we decided to limit it to the few most salient features so that the win probability could be estimated quickly by players whenever they have a brief few seconds of slow action in-game using a simple interface on a mobile app (see Figure 1).

These features are: the red and the blue team’s kills, towers, and large monsters killed.

We also chose to look at each incremental minute rather than every second.

Our training set thus includes rows like this:

where the Win/Loss indicator output variable is based on the outcome of the complete game, and each of the values on the left is a whole number indicating the number of occurrences.

One immediate issue is that consecutive rows from the same game are highly correlated. To counteract this issue, and because we had the luxury of millions of games, we chose a random minute from each game, leaving us with millions of independent rows. Games typically last between 15 and 60 min, so this is sufficiently granular.

The machine learning model we used was logistic regression:

Following the standard definition and implementation of Wolfram (2017), this models the log probabilities of the indicator variable Yi={1  win0  loss with a linear combination of numerical features xi={Minutes  Elapsedi,  …,  Blue Monstersi} weighted by the coefficients β(i)={β0,…,β7}. The estimation of the parameters by stochastic gradient descent minimizes the loss function:

Notably absent from our model is the total gold for each team. Both casual observers and professional commentators often use the net gold difference between teams as an indicator of win probability. However, this is an inherently flawed approach, especially as the game progresses, as the team with less gold can often win a team fight, ace the enemies, and march down to destroy the nexus. Furthermore, total enemy gold is not revealed to players in game and must be estimated.

Full-Game Win Probability. How much does an individual’s total performance within a game contribute to his overall team’s win probability? A version of this question appears across all sports (c.f. Cervone et al. 2014; Duch, Waitzman, and Amaral 2010).

Consider a dataset with rows of the following type, for each individual in each game:

where Tier refers to the major portion of the rank described in Section 3, i.e., Bronze, Silver, Gold, Platinum, Diamond, Master, and Challenger, ignoring the 5, 4, 3, 2, or 1 division within a tier, Kills or Deaths or other individual metric is a whole number indicating the number of such occurrences within the game, and Win/Loss is a boolean value indicating whether the team the individual belonged to eventually won the game.

Within each tier, we can compute the percentiles for the metric in question. For specificity, consider a player’s raw number of kills. This kills number will be replaced by its percentile, meaning, he did better than what portion of all players in his tier across all games in our dataset. Thus we now have:

Now we accumulate average win/loss percentages within each integer percentile (e.g. 3rd percentile, 27th percentile, and so on) to create a graph of the relationship between individual performance of a single full game metric measured as a percentile πˆ relative to that individual’s tier, and the team’s win/loss probability. This results in graphs such as the left-hand side of Figure 2.

Worthless Deaths and Smart Kills. Given our in-game win probability model, we can refine the standard metrics of kills and deaths depending on if the team’s estimated probability of winning πˆ  increased or decreased following the event.

Specifically, let us call a worthless death one that did not increase your team’s estimated win probability and a smart kill one that did. If you died but took out an enemy inhibitor, that is virtually always “worth.” (“Worth,” as opposed to “worth it,” is a standard term in LoL. Players often type it either in ally chat or all chat to indicate that the value of the objectives achieved outweighed the death itself.) But if you chased a weak enemy for 2 min to kill them, that is almost surely not a smart kill, because it was a waste of time and you likely missed out on an objective.

With these new metrics, we can repeat the process of Section 3.2 to generate graphs such as the right-hand side of Figure 2.

Basic and Advanced Stats. Beyond worthless deaths and smart kills, we can define dozens of new advanced stats. Table 3 lists the basic API stats and the new stats. For space considerations, we discuss only the most important. The bottom panel of Table 3 lists convenient bundles, chosen to be as orthogonal to each other as possible and to correspond to important higher level concepts.

Some of the stats measure activity. This includes the number of attacks, the number of each of the four abilities used, and the damage per minute caused from combinations of attacks and abilities. Others measure your team fights. A favorable team fight is one where you and your allies outnumber the enemies. How often did such fights happen, and how many kills and deaths resulted? Similarly for unfavorable team fights where you were outnumbered, and balanced fights.

Smart kills and worthless deaths (both always computed on a per-minute basis) are further broken down depending on why the death occurred. Was it due to being in an unfavorable team fight? Or from a lack of vision? Or a large difference in health or gold spent between the champions in the fight? And so on.

Determining something as fundamental as team fights is a non-trivial problem. First, of course, it requires our new optical tracking data to know the locations of all champions every second. Second, it further requires some judgment and game knowledge to disambiguate team fights. We do this by starting with engagements: an engagement is a continuous sequence of champion-on-champion damage until 5 s elapse with no damage.

Five seconds is consistent with the concept of “in-combat” used within the game by certain powerups: for example, Boots of Mobility add a burst of speed after 5 s out of combat.

During engagements, we track the damaged champion and the one who dealt the damage, and include in the team fight any champions within experience range of any champion already deemed to be in the engagement. Thus, engagements get concatenated together as champions join or leave a team fight, either by dying or retreating. Recall that champions earn XP or experience points whenever they are within a small range of an enemy’s death, even if they did not participate in the death. Thus, an otherwise uninvolved ally is deemed part of the engagement if and only if they are within XP range.

The distributions of the basic and advanced stats typically exhibit positive skewness. Almost all players fall along a bell curve but the few that had great results had particularly great results.

4.1 Examples of advanced analytics

Some of the immediately useful information is simply visual presentation of the raw data, while others also require additional computation, such as the in-game win probability. One immediate benefit of our approach is better detection of which champion played which role, a categorization that the official Riot API often fails at. Human inspection of a game rarely leaves any doubt about roles, but the Riot API uses a heuristic based on early game positioning, so, for example, a top that commits early with the bot lane to attempt to score an early kill might be erroneously deemed to be a support. The Appendix shows pseudocode for our improved detection.

Ward Maps and Heat Maps. As a visualization example, Figure 4 shows a ward and a heat map, neither of which are possible with API data. The ward map shows where wards were placed to provide vision, coded to indicate the type of ward and its fate: whether it expired, was replaced or destroyed, or persisted until the end of the game. The heat map shows the main locations a given player occupied on the map. In the case shown, it was clearly a player in the mid lane, who roamed bottom a few times, and to the baron pit, before finally pushing through to the nexus.

Figure 4:

Ward map showing placement and state of wards (left) and heat map showing position (right).

Vision. Beyond simply counting or plotting wards, we can also compute the union of all the regions of vision granted by still-living wards using their specific radius at each moment in time for each team. Figure 5 shows this visualization, with the chosen team’s vision as a percentage of the total map on top in blue, the enemy team’s vision on the bottom in red, the chosen team’s net vision as the yellow line, and the chosen individual’s vision contribution to the chosen team in the middle in green. In this specific game, the chosen team slowly but surely built a vision lead, and the chosen individual’s contribution was crucial.

Figure 5:

Map vision: team (blue), opponent team (red), net team (yellow), and individual (green).

Post-Game Win Probability with Automatic Annotations. It is common in many sports to see graphs of win probability after a game. The large fluctuations especially at ends of games can be dramatic. We further extend this standard type of graph with automatic annotations indicating key states of the game at local minima and maxima.

Figure 6 shows an example. The chosen team got a few early kills (69%), but then had seven net deaths and a lost dragon (35%), a brief mid game bump followed by a long slump of lost towers and lives (17%), before finally rallying with a sequence of good kills and likely an ace. The terminal win probability shown is below 100% because the last partial minute of game play is not included.

Figure 6:

Enhanced post-game win probability analysis showing key turning points and game state.

Tilt Analysis. A common psychological issue in LoL is the concept of tilt. Borrowing from poker, tilt in LoL refers to situations where a player plays suboptimally because of emotional turmoil. The most extreme form of tilt occurs when a player dies in lane then returns immediately and plays too aggressively and dies again. We can proxy for tilt by calculating the time spent dead by the player over the preceding 3 min and dividing it by the total time spent dead by anyone on his team over the same period. Three minutes was chosen because death timers range up to around 60 s, and it takes approximately 30 s to walk back to the middle of the map, so “inting” or intentionally dying to the enemy immediately on respawn will quickly show up on this metric. This measure can be calculated from basic Riot API data, and produces the blue chart shown below in Figure 7, starting by convention at 0.2 instead of 0.0 to indicate balance because there were zero individual deaths and zero team deaths. However, this calculation by itself provides little context or explanation. If we further highlight all the player’s deaths in red, and note the kinds of team fights they were involved in (which requires the new data), then the context helps to understand and possibly alleviate future examples of tilt. In the example below, notice the player first died in a favorable and then a balanced team fight, before taking three unfavorable fights in a row.

Figure 7:

Tilt analysis (portion of recent team deaths responsible) and engagements.

Time Management. Given the location and damage target information, we can partition each player’s activity at any given moment into one of several types. An example of a post-game time management analysis is shown in Figure 8. As is common, before minions spawn, most time is essentially wasted, meaning, not engaged in productive activity. Reducing the amount of wasted time during the rest of the game, however, is an important metric. Transition time ought to be minimized. Laning time should also decrease over time, with the possible exception of split pushers.

Figure 8:

Time management, indicating activities and portion of game phase spent.

Individual Performance and Team Win Probability. How does an individual’s performance within a game correlate to his team’s probability of winning? We previewed earlier in Figure 2 that the relationship between standard kills or deaths with winning is positive but imperfect, similar to the relationship one might expect between, say, an individual’s scoring and his team’s winning in a traditional sport such as basketball. Filtering individual production for activities that benefit the team, however, has a strong effect.

Figure 9 shows the win probability relative to the percentile for various non-overlapping bundles of advanced stats, bundled as shown in the bottom of Table 3, and separated by tier. The OVERALL figure is for the average of all the bundles, as it is a proxy for an individual’s overall in-game performance. Averaged across games, it can also represent a player’s overall individual rating.

Figure 9:

Individual performance in a bundle of stats versus team win probability.

The OVERALL metric does well. It satisfies the main criteria of being a good metric, namely, as the percentile performance on the metric in a game increases, the team’s overall win probability also increases, for each tier. However, it is not optimal because it includes several bundles that do very poorly at explaining a team’s win probability: laning, carry focus, vision denial, and vision. An individual’s performance in these four bundles seems to be irrelevant to a team’s win probability. An OVERALL metric ignoring these four bundles would do even better.

Some of the bundles that do well are intuitively reasonable—deaths and favorable fights line up perfectly and make sense: don’t die, and don’t take bad fights. Gold also is increasing but not as much. As we know, gold is good but it is not the whole story. The same is true for abilities and survivability.

Time management is an intriguing one, as it measures an individual’s impact across space and time, through map usage and portion of time spent in useful activities. To the extent you had one of the best games in terms of spacetime presence for your rank, you are virtually guaranteed to win.

These graphs, however, only show the isolated effects of each bundle separately. We can further consider the marginal contributions in a complete model with a logistic regression.

Logistic Regression. We fit a standard logit model on the 10 stat bundles percentiles and a constant term for each individual in each game. The estimates βiˆ and their corresponding z-statistics in the logistic model yˆ=1/(1+e−(β0+β1f1+β2f2+…+β10f10)) are shown in Table 4 below, for the corresponding fi that map onto the 10 bundles listed.

All the estimates are highly significant at the p < 0.0001 level except for Abilities, which has a p-Value of 0.03. As noted in footnote 2, our sample size was approximately 150,000 games. The McFadden (1974) likelihood ratio index or pseudo-R-squared for the regression is 0.42, which represents an excellent fit, as it is even higher than the 0.20 to 0.40 range that McFadden (1978) refers to as “an excellent fit.”

The Effect Size column computes the difference in the probability of winning if the respective factor is 55 versus 45% while all the other input factors are fixed at 50%: only Favorable Fights, Deaths, and Time Management are highly sensitive and have a high impact; the others require larger variance to have a strong effect on the win probability.

Surprisingly, several of the bundles have negative marginal contributions, including survivability. Conditional on all the other bundles, it is actually better for you to have less survivability. Similarly with vision, vision denial, carry focus, and laning. There are two legitimate possible explanations for these negative marginal stat bundles. One is that those metrics tend to correspond to a measure of passivity rather than aggressiveness, and aggressiveness tends to win games even if on the margin it makes those stat bundles lower. The other possibility is of multicollinearity among the features.

Table 5 shows the pairwise correlations of each of the features. Very few exhibit substantial multicollinearity. The biggest positive correlation is 0.51 between carry focus and favorable fights. This makes sense because someone who is able in team fights to focus the carry (defined as the enemy’s largest damage dealer) is likely also able to avoid fighting in unfavorable conditions. The biggest negative correlation is −0.75 between vision denial and gold. This also makes sense because support players tend to sweep for enemy wards, and they usually earn relatively little gold.

With roles in mind, we can redo the logistic regressions for each of the five standard roles: top lane, jungle, mid lane, attack damage carry (“adc”), and support. Table 6 lists the z-Statistic for the relevant coefficient in a logistic regression of player-games restricted to the given role.

Certain bundles are highly significant across all roles: favorable fights, deaths, vision, time management, and survivability. Carry focus is significant for all roles except the top lane; the top laner often split pushes and does not join team fights, so he ends up focusing whoever comes to stop him rather than the enemy carry. Laning is important for laners (top, mid, and adc), and not important for junglers and supports, while abilities are the opposite.

For completeness, we can see the separate correlations of each bundle with win probability, by role. Table 7 shows that the top factors are as expected and relatively constant across roles. The implied advice to be a better player is: fight good fights, don’t die, and don’t waste time.

Fast and Frugal Trees. We compute a fast and frugal tree to see the best way to predict your team’s performance given only the most important details about an individual’s performance. Fast and frugal trees (described in general as well as the implementation used here in Phillips 2016) are decision rules that are both easy-to-use and make good predictions. We train on all the metrics of Table 3, with 25% of the data, about 35,000 observations, reserved for cross-validation. Figure 10 shows the best resulting tree. It makes the correct prediction about 80% of the time despite only using three metrics.

Figure 10:

Fast and frugal tree analysis for individual percentile performance to predict team victory.

The final decision rule is simple indeed: as a rule of thumb, if your bad death percentile was below average (meaning you had more worthless deaths than players at your rank typically do in games), then you likely lost. If, on the other hand, you didn’t die that much, then, if you had more smart kills than your peers, you likely won. Finally, if you had both above average worthless deaths and below average smart kills, then whether your team won or not depends on how much influence you had over the map. If you had more map coverage then average, you probably won; otherwise, you probably lost. Thus, the fast and frugal tree (FFT) clarifies our story: don’t die for no reason; if you do, try for some smart kills; if you can’t do that either, at least roam and have decent map coverage.

Note also that the FFT was free to choose any decision tree using any of the metrics in Table 3, including both the basic and advanced stats, and it ultimately chose to use only the advanced stats, and exactly the ones we had identified in the previous section as having a consistent impact on win probability across all roles. Note also that the overall FFT accuracy is 80% with a false positive rate of only 13%.