Very recently, we ran a fun exercise with some friends and clients to predict the score of each Euro 2020 football match. In order to spice up the challenge, we added 3 simple, heuristic robots (we call them our ArtiBots) into the team. You can see more background on this in our 1^st blog related to this topic on LinkedIn from Stefan Ravizza.

You’re wondering about the results?

Now – from the 3 ArtiBots, 2 did outperform all players by far. The bots were analyzing the winning odds from a betting office and then made either 1:0 or 2:1 predictions for the better ranked national team. The final scores have been the result from our analysis of the past Euro and World Cup results as the most “regular” results. So e.g. for #SUIFRA our first ArtiBot has predicted 0:1 for France in the round of 16 (what we know would be a total fail of the prediction – 0 points).

In our prediction community the bots were performing in absolute terms more than 10% better than the human tippers across the entire Euro 2020. His means e.g. while the best human predictor scored 141 points, the bot scored 157 points.

OK, now you could say, we / our community is full of specialists in Data, AI, Analytics and DataScience, but certainly not in predicting football score nor in football itself. And you might be right. So we thought we let our ArtiBots also play against a bigger prediction community. We selected the SRF – Swiss Radio & Television community from the Euro prediction game (see here https://emtippspiel.srf.ch/). From the user statistics from SRF we learned that there were more than 200’000 registered people of which we assume about 50-60% were active and submitting predictions throughout the entire Euro 2020 tournament. This clearly is a wider (and possibly better) comparison community as there are also known football experts (like Andy Egli and Kathrin Lehman or Diego Benaglio and Baschi) submitting their tips.

Applying the scoring mechanism outlined under the SRF game rules (URL: https://emtippspiel.srf.ch/app/hilfe_regeln.jsp), our best performing ArtiBot (the one putting its predictions on 1:0 for the better national team) would have scored 336 Points. This score is achieved without any bonus questions (of which a total 50 bonus points were possible).

So, what does this mean?

With 336 Points our 1^st bot would be ranked 173 out of more than 100’000 active players! This is almost too good to be true!

So, we also checked our 2^nd best ArtiBot (which predicts 2:1 for the better country) – and similar result here – our robot would have scored 327 points and reached rank 538 out of approximately 100’000 players.

In other words – our simple & heuristic ArtiBots are delivering very high performance and are getting into the top 1% of all players – in fact, 99,8% of the players were performing worse than the bots!

Also, we must remember, the score is without the 50 bonus points for special questions like “Until where does the Swiss team manage to get?” or “How many goals will the Swiss score?”…). Assuming our best ArtiBot would have scored 50% of the bonus points – it would have scored 362 and as such would have achieved the 3^rd place overall! The overall winner had 368 points – including all 50 bonus points.

Wow – I need to swallow yet again and still almost can’t believe this performance of our ArtiBots. This does truly fascinate me, how such a simple bot can outperform humans predicting score.

What might be the reasons for this?

Obviously, we are wondering what the reasons for these extraordinary results might be. Maybe there are some answers we could find in the game theory and with the help of pure (but simple) math & statistics and finally psychology.

• Math, Stats & history: Well, this is simple 1&1 on data – and the analysis of the past results will result in the most likely final scores which is then used. Not rocket science but still some basic math and early-grade statistics that must be applied. The most difficult part is truly to collect the data – but luckily there are too many sports portals which make it not too hard to get the past results. J
• Game Theory: In game theory it’s a simple approach to calculate based on odds – and then stick to the winning strategy, no matter what happens. In this case, our ArtiBots have no choice but to put their bets on 1:0 and 2:1, respectively for the better classified national squad. This is based on the data the dominant strategy.
• Psychology: In addition to the above points, this might be the most impactful factor of why machines win over humans. This is part is widely known as the bias theory – but humans make for different reasons (biological, evolutional, mathematical, …) mistakes in their reasoning. One of the biggest mistakes is probably our thinking we can perform better than the statistics. We believe to have a better understanding, a better knowing of the facts or a better analysis of certain circumstances that might be the reasons for one team to win over another. We do not!

Of course, there are always human over-performers that get better scores predicted than the basic math & stats – but they are rare and in fact are mathematical outliers. One could argue that these persons might had better information, a better feeling or luck. We might never be able to fully explain these top performances.

While it’s still unbelievable that the ArtiBots have performed so well – it’s almost sad we have not registered our ArtiBots for the SRF prediction game – who knows how far we could have gotten. Luckily there is another prediction game soon with the World Cup 2022 in Qatar. 😀

Over to you…

What is your take from this? Do you see other reasons why the Bots were better than most of the people participating in the prediction game or do you have any other hypothesis, we could validate in our data? Let’s engage in a open discussion! I bet on you 😀