Power Predictions Explained

Power Predictions Explained

Following the introduction of F1 power predictions, this post explains how the reader should interpret the information providedĀ on power.f1-predictor.com. As a reminder of the legal disclaimer, the website exists for entertainment purposes only and does not provide sports betting tips. Betting is discouraged.

As mentioned in the introductory post, the focus of the power predictions dashboard is on highlighting certain predictions for the upcoming race before the qualifying and the practice sessions have been completed and showcasing their historical accuracy. For each race and confidence level, up to two predictions will be provided. Currently, all predictions refer to whether a driver will finish in the top 3, 6 or 10 positions. Other types of predictions may be added at a later point.

Upcoming race – power predictions table

The table above shows some example power predictions for a given race. It shows three confidence levels/strategies, each having different predictions and characteristics. The table shows that according to the very high confidence level prediction (strategy 1), Verstappen will end up in the top 10 places with 97.2% probability. Similarly, the high confidence level prediction (strategy 2) shows that Ocon will finish in a point-scoring position with 82.4% chance while strategy 3 suggests that the probability of Kvyat finishing in the top 10 positions is 73.6%. Once that race is completed, the same predictions will be moved to the historical power predictions along with the actual results. The goal is to provide total transparency over the historical performance of the power predictions.

Upcoming race – historical performance per strategy

This section shows how the historical power predictions would have performed assuming certain betting odds. The three confidence levels (i.e. strategies) of power predictions were selected because they have different performance characteristics.

As seen in the hypothetical image above, the very high confidence predictions (strategy 1) would have provided a positive average yield (i.e. average amount of return for each unit invested) which is lower compared to the other two cases. Medium confidence (strategy 3) would have provided the highest yield among the given predictions. Partially, this happens due to suggesting higher odds events, i.e. more improbable events, as seen in the ‘median odds’ chart. The yield above assumes all bets equal a single unit of money and remain constant across races.

While very high confidence predictions (strategy 1) would have offered a lower yield, they had the highest hit rate, i.e. the percentage of races where they were accurate. This means that very high confidence predictions have been right more often but with a lower return. In contrast, medium confidence predictions (strategy 3) have been right less often but with higher returns. The same thing is also shown on the ‘gain deviation’ chart which shows the standard deviation of the net gain or loss across the historical races. This again confirms that the very high confidence predictions would have the smaller deviation (which is better) while high confidence and medium confidence would have a higher deviation.

Historical power predictions

In our effort for maximum transparency over the historical performance of our power predictions, we are providing the full list of historical predictions along with the actual betting odds and the result.

As shown in the screenshot above, the table is filterable. The screenshot shows that the high confidence prediction (strategy 2) had suggested Ricciardo to finish in the top 10 positions in the 2020 Russian GP with 80.9% probability. Ricciardo did finish 5th and this meant that the net gain in that case would be 0.35, assuming a bet of a single unit of money. The bet percentage column shows how this single unit of money was split between the predictions of that row (if more than one).

The plots on the lower part of the dashboard are meant to visualize various aspects of the predictions across the seasons in which power predictions were provided. Certain races may be missing entirely because F1-predictor did not have the historical betting odds available for those races.

Feel free to come back with suggestions, feedback and comments. As always betting is discouraged!

Please share the post:

9 thoughts on “Power Predictions Explained

  1. I am working on a similar project, developing a model to predict race or season winner based on previous data. My case is more based on overtaking data collected from previous seasons so we can estimate the future. Can I know more about the algorithm you used?

    1. Hi Ali, in terms of algorithm I’m using LightGBM and XgBoost models. I’m planning to explore neural networks to see if embedding the drivers, constructors, circuits etc. captures any hidden patterns.

      Interesting approach that you’ve used with overtakes. Have you published it anywhere?

  2. Hi, which odds are you using to benchmark your algorithm against? Do you have the historical ones somewhere?


    1. Hi, I have been collecting historical odds since 2018 for almost all races. I don’t want to share which website I’m scraping them from but it’s a well-known betting site. I don’t want to publish the historical odds online but I could share them with you if you need them. Just ping me on a.stergioudis@gmail.com

  3. Hi,

    As someone with a data science interest and very small knowledge of F1, I am interested to know what effect the rule changes in F1 for 2021 have had on your performance this year? Would retraining the data on this year alone be a worthwhile endeavor?

    Thanks for sharing your work on this, interesting stuff.

    1. Hi, overall I think F1 has been more unpredictable this year. However, we have seen bigger changes – for instance in 2014 when the new engines were introduced. The model needs at least a few races after a big change is introduced till it starts making accurate predictions.

      Personally, I’m using data since 2000 (i.e. last 20 seasons) to train the model. If I was using only this year alone, the amount of data points would be too small to train any decent model. Next season will introduce huge changes in the F1 cars, so let’s see how predictable 2022 season will be.

  4. Hi Stergios this is very interesting. Congrats and all the best. I was wondering what stats/probability do you calculate during your EDA phase?

    1. Hi Shev, thank you. I’m not sure I understand what you say. I have built models that predict whether each driver is going to finish in the top-3, top-6 or top-10 in the next race. Then, I compare them against the odds given by a bookmaker and I pick some interesting ones.

Leave a Reply

Your email address will not be published. Required fields are marked *