Darina Goldin, Bayes Esports: Why is it so hard to create good betting odds?

Esports is data-driven like no other sport. Given a good data source, it is possible to have full information about a match at any given second, yet creating good reliable betting odds for esports remains a challenge in the industry, writes Darina Goldin, Director of Data Science at Bayes Esports.

Some of that can be attributed to the relative novelty of esports – traders who are familiar with the games are still hard to find. Thus esports odds often fall in the hands of either bookmakers who have only traditional sport experience or data scientists who have only a limited understanding of the betting market. 

We need companies like Bayes Esports who ensure that parties work together. Only then does it become possible to create esports odds that are fair, entertaining and ultimately profitable.

The first step in creating odds is creating a model that correctly reflects the gamestate. Usually an esports match is played as a set of maps (comparable to sets in tennis) and we start with the map or match winner market. Given all currently available information, we want to correctly estimate the probabilities for each team to win the map.

From this base probability, derived markets can be calculated: the match win probabilities, as well as correct outcomes and handicap lines. Then we can turn to side markets which are very title-specific. In FPS like Counter-Strike, we will want to offer odds for round wins, bomb plants and defuses, headshot and ace counts, among other things. In MOBA titles we will focus on objectives, like the team to fall the first tower, or kill a significant neutral monster like Roshan in Dota 2.

Here already we encounter the first difficulty in predicting within esports.

Side markets are not always related to the map winner

Some side markets are extremely indicative of the map winner – winning the pistol round in CS:GO easily decides the outcome of the next two rounds and getting the first Baron in League of Legends means that you will very likely win the map. Other objectives, however, are more random or subject to different team strategies.

First blood is one such market. This market is won by whichever team gets the first kill, comparable, perhaps, to the first basket in basketball. When predicting the first basket we will take into account how likely each team is to win the tip off, which depends more on the size of their tallest player than on the skill of the team as a whole. 

So while getting the first basket undoubtedly gives you an advantage to win the match, we can end up with very different odds for the two markets. The same is true for first blood, especially in MOBAs, where the correlation between getting the first kill and winning the map is around 0.22. Sometimes a team will have a better setup to get the first kill, because they are playing an early game strategy. Sometimes the first kill is just random. Compare that to the first Baron killed, which in the current patch has a 0.78 correlation with map winner!

The strategy argument is even more compelling when it comes to the first Dragon in League of Legends. Here, too, the correlation between getting the dragon and winning the match currently sits at roughly 0.21. It’s an advantage, but plenty of teams win without needing or planning to get the first dragon! More so, in Dota 2 you can sometimes predict who will get the first Roshan by simply looking at the team compositions – having Ursa on the team, for instance, is a high indicator of a Roshan play.

Ultimately, whether or not the team attempts to kill it is a tactical question somewhat separated from the overall objective of winning the map. Predicting correctly in these situations requires a large amount of game knowledge.

As we saw with the basketball example, this isn’t something that only ever happens in esports. But how often is the soccer team most likely to win, not also the one predicted to get the most corners? In traditional sports these situations are exceptions to the rule. In esports, it IS the rule.

This is something that is very clear to anyone working at Bayes Esports. But it remains hard to communicate to audiences who do not understand esports well and expect the favourite team to also be the favourite for all side markets. It also makes the modelling harder – instead of looking at straightforward win indicators like current net worth or kill counts, we have to take into account the properties of the selected heroes, the skill and preferences of individual teams and players, and so on.

Find your prematch odds on Twitter

Speaking of skill and preferences – one common issue in esports is keeping track of teams at all! There is no single source of information – in fact, there aren’t even consistent naming conventions for teams, a topic Bayes Esports explored in our last whitepaper

It is common for teams to reform under new organisations, to swap out a significant number of players or simply be referred to differently by two different sources. Despite the games taking place almost entirely online, keeping track of professional teams currently still requires a significant manual effort.

Additionally, no definitive rankings exist for esports teams. You can find some attempts at ranking like HTLV’s top teams or Datdota, but these are unofficial third party resources. If you want to have a good idea of team skills, you have no choice but to keep your own records and track the teams yourself, sourcing twitter for most recent developments.

Draft is everything

One common pain point in creating live odds for MOBAgames in particular is dealing with the draft phase. At the start of each map, the teams will ban and pick heroes from a pool of over 100 different options, until each team ends up with five. All heroes have different strengths, weaknesses, and synergies with each other. Usually each team’s strategy is already visible from the hero selection: are they playing the long game or have they selected heroes that are most effective in the first ten minutes of the match? Are they planning to heavily attack buildings or neutral monsters? Will they be able to successfully counter the opposing team?

All these questions affect the odds of the teams in the opening minutes of the game. But how do you assess the draft in an automated fashion? The sheer number of combinations you can have makes it impossible to statistically value each possible hero combination. 

Add to this that some players are a lot more proficient with certain hero choices than with others, and you can have an idea of the difficulties this problem poses to modeling. Hence companies like Bayes Esports need to rely on modelling tricks or expert opinion in order to accurately assess the team strength after the picking phase.

Both approaches are notoriously unreliable. This is especially true for side markets, which are very vulnerable to meta-game changes. It is hard to leave the algorithm to do the work by itself – if you want to be safe, you need a trader and the trader needs to understand what is going on. Very few traders will be able to understand the game to the desired level, however – and it only takes one clever punter to create significant losses.

A similar, though not as hard, problem arises with the map selection in CS:GO. Some teams are better at certain maps, or even at one side of certain maps. This is something we need to keep track of if we want to offer the best odds we can. Since a complete, official database of Counter-Strike results does not exist, a lot of manual effort goes into obtaining this information.

Patches, patches, patches

Now here is something that you will hardly experience in traditional sports: patching. Publishers keep their games interesting and balanced by patching them in regular intervals. Patches can bring minor changes like item price or cooldown duration, or they can introduce new heroes, add or remove buildings, and any number of other tweaks. More often than not, patches will significantly change the game rules. Teams will need to find the best way to play under these new rules and whoever is quicker at adapting to them will get a competitive edge. This often leads to changes in the so-called meta-game – the set of strategies played at the top level of the sport. 

We see patches happen most often in new titles like Valorant and Overwatch. These games are still being developed as an esport and are bound to keep changing dramatically. But even the long-running Counter-Strike is not exempt, as we were two years ago when Valve changed the economy dynamics in the game. There is no saying when they will decide to change it again.

In League of Legend and Dota 2 patches are simply a part of life – you can expect one at least once per month and adapting well is part of the competitive skillset. This is the publishers’ way to keep their games fresh and interesting despite offering only one map and game mode.

If you want to create data-driven odds patches are the bane of your existence. Just look at the draft phase! We have just explained how important it is to value it correctly. Now if a patch changes certain heroes, then their function in teams will change accordingly and your entire research needs to be redone. 

Oftentimes it’s not just old research, but old data that is rendered unusable by a patch. For example Dota 2 that first introduced, then moved, and then removed shrines. This did not just invalidate the data set, but also required Bayes Esports data scientists to first add and then subsequently remove the “first shrine” side market.

After a large update even the pro players cannot fully estimate how the meta-game will develop and who will come out on top in the first few weeks. So how can a bookmaker know? In these situations we have no choice but to pause our algorithms and trade manually until a new standard is achieved.

In the worst case, the patch will destroy not just your odds models, but even your parsing technology. This happened for example when Dota 2 moved from Source to Source 2 engine, which affected all parsers. The patching problem is so inherent and unique to esports that it requires its own solutions. Whoever first finds a robust way of dealing with the ever shifting metagame will have a huge competitive advantage in the world of esports odds.

The need for more talent

We’ve said it before and we’ll say it again: There simply aren’t enough traders who understand esports well enough to trade them. In fact, it’s even wrong to speak of “traders who know esports” – that’s like saying “traders who know sports”. We don’t expect the same person to be able to trade soccer, boxing and cricket equally well, and the same is true for someone trading Dota 2, League of Legends, and Counter-Strike. Each esport title is a very different game with its own nuances. Each requires very different knowledge.

Luckily, the raw talent is already there. We now have an entire generation who grew up watching and playing esports. But only few of these people are aware that esports betting exists and there are career opportunities in this market. The industry needs to get better at reaching these people. The only way forward is to raise and train a new generation of dedicated esports traders. It’s not enough to just post a job on your website – you need to advertise these jobs during large events, and you need to make them entry level.

Check Also

Triggy: focus will continue to be on relevance and simplicity

Christer Larsson, Chief Business Development Officer at Triggy, outlines the main ways in which sportsbooks …

SOFTSWISS clients save €2.8m thanks to highest level of anti-fraud support

The SOFTSWISS Anti-Fraud team helped its clients to save €2.8m in Q1 2022, according to …

Bayes Esports facilitates new growth opportunities with latest funding round

Bayes Esports has closed its latest financing round of €6m, a cash injection that CEO …